Router One Changelog & Updates

This changelog covers the major features, improvements, and milestones shipped to the Router One platform. We update this page as new releases go live.

October 2026

Gemini 4 Argon Listed

As of the 2026-10-14 catalog, Router One lists Gemini 4 Argon (google/gemini-4-argon); at 01:55 UTC on 2026-10-14 the catalog held 68 ids — 60 text and 8 image. The catalog entry lists text and image input, text output and the chat, streaming, tool-calling and vision flags. The gateway also accepts Google's own id, gemini-4-argon, for the same catalog entry. As a Gemini id it is served on POST /v1/chat/completions only. As of the 2026-10-14 plan response, Gemini 4 Argon is in the Premium models tier of the Pro, Max and Ultra plans, beside GPT-6.1 Sol, Claude Opus 5 and the other models that tier lists; on a plan, its calls count against the Premium models allowance, and calls beyond the allowance bill the wallet at the posted rates. The Gemini 4 Argon API guide covers the catalog id, the endpoint and a first request; the October 2026 new-model guide, the Gemini API in China page and the pricing facts are updated.

Gemini 3.1 Pro Preview No Longer Listed

As of the 2026-10-13 catalog, Router One no longer lists Gemini 3.1 Pro Preview (google/gemini-3.1-pro-preview); the other Gemini chat ids stay listed, and the Gemini API in China page lists them. Choose a Gemini id the live catalog lists; each model page shows its endpoints and rates. The October 2026 new-model guide, the September 2026 new-model guide and the pricing facts are updated.

Cancelled Streams: Output Already Streamed Is Billed on an Estimate When No Final Usage Arrives

As of 2026-10-13, when a client cancels a stream on POST /v1/chat/completions or POST /v1/responses after output has started and no final usage arrives within the gateway's five-second window, the request is no longer settled at only the usage already observed: the part the upstream did not report is estimated — output tokens from the text already streamed, input tokens from the request — and billed. A stream cancelled before any output is still not billed if no usage arrives, a cancelled request is still never retried or failed over, and Dashboard → Logs shows the settled cost while its token counts keep what the upstream reported, so such a row can show a cost next to zero tokens. Details are in the streaming guide and the pricing facts.

Claude Haiku 5.5 Listed; Gemini 3.5 Flash Lite in the Standard Models Tier

As of the 2026-10-10 catalog, Router One lists Claude Haiku 5.5 (anthropic/claude-haiku-5.5); the catalog holds 68 ids — 60 text and 8 image; no id left the catalog. The catalog entry lists a 1,048,576-token context window, text and image input, the chat, streaming, tool-calling and vision flags, and two rate lines split at 100,000 prompt tokens; no aws/, vertex/ or azure/ twin is listed. The gateway also accepts Anthropic's own id claude-haiku-5-5. As a Claude-family id it is served on POST /v1/messages and POST /v1/chat/completions, not on POST /v1/responses. No plan tier in the 2026-10-10 plan response lists it, so its calls bill per token from the wallet balance; Claude Haiku 4.5 stays in the Standard models tier of all three plans. Per Claude Code's changelog and model configuration docs (Claude Code 2.1.296, checked 2026-10-10), 2.1.293 and later resolve the haiku alias to claude-haiku-5-5, so /model haiku and subagents defined with model: haiku request Claude Haiku 5.5 and bill per token from the wallet; the optional ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-4-5 line in the Claude Code setup guide keeps them on Claude Haiku 4.5. The Claude Haiku 5.5 API guide covers the id, both endpoints and a first call, and model comparisons gained Claude Haiku 5.5 vs Claude Haiku 4.5, Claude Haiku 5.5 vs Gemini 3.5 Flash Lite and Claude Haiku 5.5 vs Claude Sonnet 5.5.

As of the 2026-10-10 plan response, Gemini 3.5 Flash Lite (google/gemini-3.5-flash-lite) is in the Standard models tier of the Pro, Max and Ultra plans, beside Gemini 3.5 Flash, Claude Haiku 4.5 and the other models that tier lists; the 2026-10-09 plan response listed it in no tier. On a plan, its calls count against the Standard models allowance, and calls beyond the allowance bill the wallet at the posted rates. Gemini Nano Banana 2.1 is still in no plan tier. The October 2026 new-model guide, the Gemini API in China page and the pricing facts are updated.

Gemini 3.5 Flash Lite and Gemini Nano Banana 2.1 Listed; Claude Fable 5.1 No Longer Listed

Update (2026-10-10): as of the 2026-10-10 catalog Router One lists Claude Haiku 5.5 (claude-haiku-5-5), in no plan tier as of the 2026-10-10 plan response, and as of the 2026-10-10 plan response Gemini 3.5 Flash Lite is in the Standard models tier of the Pro, Max and Ultra plans — see the Claude Haiku 5.5 and Gemini 3.5 Flash Lite entry in October 2026 above; Gemini Nano Banana 2.1 is still in no plan tier.

As of the 2026-10-09 catalog, Claude Fable 5.1 (anthropic/claude-fable-5.1) is no longer listed, and two ids are new: google/gemini-3.5-flash-lite and gemini-nano-banana-2.1; the catalog holds 67 ids — 59 text and 8 image. Claude Fable 5 (anthropic/claude-fable-5) stays listed. Claude Code's fable alias resolves to Claude Fable 5.1 (claude-fable-5-1) per its model-config docs (Claude Code 2.1.295, checked 2026-10-09), so the Claude Code setup guide pins it to Claude Fable 5 again with ANTHROPIC_DEFAULT_FABLE_MODEL=claude-fable-5; in the 2026-10-09 plan response Claude Fable 5 is in no plan tier. The guide also notes that Claude Code 2.1.293 and later resolve the haiku alias to claude-haiku-5-5, which the 2026-10-09 Router One catalog does not include. Gemini 3.5 Flash Lite, which the Gemini API release notes (checked 2026-10-09) list as generally available, has a 1,048,576-token context window, text and image input and the chat, streaming, tool-calling and vision flags in the catalog; as a Gemini id it is served on POST /v1/chat/completions. It is a different model from Gemini 3.5 Flash. Gemini Nano Banana 2.1, which the same release notes describe as an update to Nano Banana 2, is an image model billed per generated image: it is served at POST /v1/images/generations and takes a reference image on POST /v1/images/edits; the image generation API page and the Nano Banana 2 API guide cover both calls. No plan tier in the 2026-10-09 plan response lists Gemini 3.5 Flash Lite or Gemini Nano Banana 2.1, so their calls bill from the wallet balance at their posted rates; Gemini 3.5 Flash stays in the Standard models tier. Model comparisons gained Gemini 3.5 Flash Lite vs Gemini 3.1 Flash Lite and Gemini 3.5 Flash Lite vs Gemini 3.5 Flash. The October 2026 new-model guide collects the month's catalog and plan changes.

Long Non-Streaming /v1/messages Calls Are Held Open; Disconnected Non-Streaming Calls Settle at Actual Usage

Update (2026-10-13): a later change altered how cancelled streams are billed on POST /v1/chat/completions and POST /v1/responses — see the cancelled-stream entry in October 2026 above.

As of 2026-10-09, a stream: false call on POST /v1/messages is read from the model as a stream and returned as one Anthropic Message body, the way long non-streaming Chat Completions and Responses calls already were. Once such a request has been running for about 25 seconds, the gateway commits HTTP 200 and writes JSON leading whitespace as a keepalive, so a failure after that point arrives as HTTP 200 with the Anthropic error envelope, {"type": "error", "error": {...}} — check type before reading content. A generation that outlasts the gateway's window for non-streaming requests ends with the message generation exceeded the non-stream response window; use stream=true for long generations. On all three endpoints, closing the connection does not stop a non-streaming call: if the model finishes, the request is billed for the usage it reports and recorded in Dashboard → Logs as HTTP 499 client_cancelled with those tokens; if it fails, nothing is billed. Streams the client closes are settled as before. Details are on the error codes page, in the API compatibility facts and in the streaming guide.

GPT-6 Sol and GPT-6.1 Sol in the Premium Models Tier

As of the 2026-10-06 plan response, GPT-6.1 Sol (openai/gpt-6.1-sol) and GPT-6 Sol (openai/gpt-6-sol) are in the Premium models tier of the Pro, Max and Ultra plans, beside GPT-5.6 Sol, GPT-5.5, Claude Opus 5 and the other models that tier lists; the 2026-10-03 plan response listed neither. On a plan, their calls count against the Premium models allowance — 1 request per call, 2 when the total input is strictly above 272,000 tokens — and calls beyond the allowance bill the wallet at the posted rates. When config.toml names no model, Codex 0.159.1 and later use their bundled default, gpt-6.1-sol (per the 0.159.1 release notes and the rust-v0.160.1 source, checked 2026-10-06), so on a plan such sessions draw the Premium models allowance too. GPT-6 Astra stays the only model in the Flagship models tier, which only Max and Ultra carry (2026-10-06 plan response). The pricing page lists each tier's models and allowances; the GPT-6.1 Sol API guide, the GPT-6 Sol API guide and Codex CLI in China are updated.

October 3GPT-6 Astra API Guide and a Claude Opus 5.5 vs GPT-6 Astra Comparison

A new GPT-6 Astra API guide covers openai/gpt-6-astra: the catalog id, the endpoints that serve it, what OpenAI's docs say about its tool calls and reasoning effort, Codex CLI setup, billing and first calls. In the 2026-10-03 plan response, GPT-6 Astra is the only model in the Flagship models tier, which only Router One's Max and Ultra plans carry; on Pro, and once a Max or Ultra cycle's Flagship allowance is used, its calls bill per token from the wallet balance. Model comparisons gained Claude Opus 5.5 vs GPT-6 Astra, and the existing GPT-6 Astra pairs now link to the guide.

CORS Preflights on /v1 Allow Any Request Header; Chat Completions Accepts max_completion_tokens (October 1–2)

Since October 2, the API endpoints under https://api.router.one/v1 allow every request header that a CORS preflight lists, including SDK headers such as x-stainless-* and anthropic-dangerous-direct-browser-access that the earlier fixed list did not include; preflight responses carry Access-Control-Allow-Origin: * as before. Web pages and WebView-based clients built on the OpenAI or Anthropic SDKs can now call them directly. No cookies or other credentials are involved: the API key travels in the request header as usual. Anyone who can load a page can read a key shipped in it, so never put a key in a public web page; give a browser or desktop client its own key with a maxSpend cap. The API compatibility fact sheet and the API connection troubleshooting guide cover the details and what to check when a browser still reports a CORS error, and the Copilot for Obsidian guide now has you keep Copilot's Enable CORS setting off.

Since October 1, the API reference documents that POST /v1/chat/completions accepts max_completion_tokens, OpenAI's current name for the output cap, alongside max_tokens: send either one; when both are sent, max_completion_tokens wins. On POST /v1/responses the output cap is max_output_tokens.

September 2026

September 30Claude Fable 5.1 Listed Again; GPT-6.1 Sol Listed

Update (2026-10-06): in the 2026-10-06 plan response GPT-6.1 Sol is in the Premium models tier of the Pro, Max and Ultra plans — see the GPT-6 Sol and GPT-6.1 Sol entry in October 2026 above; Claude Fable 5.1 is still in no plan tier.

Update (2026-10-09): Claude Fable 5.1 is not in the 2026-10-09 catalog — see the Gemini 3.5 Flash Lite, Gemini Nano Banana 2.1 and Claude Fable 5.1 entry in October 2026 above; the Claude Code setup guide pins the fable alias to Claude Fable 5 again.

The 2026-09-30 catalog re-read at 06:48 UTC contains 66 ids — 59 text and 7 image; none left. anthropic/claude-fable-5.1, delisted on 2026-09-11 (see the September 11 entry), is listed again: a 1,048,576-token context window, text and image input, and the chat, streaming, tool-calling and vision flags. Per Claude Code's model-config docs (checked 2026-09-30), its fable alias resolves to Claude Fable 5.1 (claude-fable-5-1, which the gateway accepts for this model), so the Fable pin in the Claude Code setup is no longer needed; the Claude Code setup guide has the updated pins (plan holders pin ANTHROPIC_DEFAULT_SONNET_MODEL=claude-sonnet-5; ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-4-5 is optional). As a Claude-family id it is served on POST /v1/messages and POST /v1/chat/completions, not on POST /v1/responses. openai/gpt-6.1-sol (catalog description "GPT-6.1 Sol") is listed: a 1,050,000-token context window, text and image input, and the same four flags; as a GPT-family id it is served on POST /v1/chat/completions and natively on POST /v1/responses. No plan tier in the 2026-09-30 plan response lists either model, so their calls bill per token from the wallet balance, as Claude Fable 5's do. Dated details are in the September 2026 new-model guide.

September 29Claude Sonnet 5.5 Listed

The 2026-09-29 catalog re-read at 18:20 UTC contains 64 ids — 57 text and 7 image; none left. anthropic/claude-sonnet-5.5 is listed (catalog release date 2026-09-29): a 1,048,576-token context window, text and image input, the chat, streaming, tool-calling and vision flags, and one rate line across the whole window (no long-context tier). The gateway also accepts Anthropic's own id claude-sonnet-5-5 for the same model. As a Claude-family id it is served on POST /v1/messages and POST /v1/chat/completions, not on POST /v1/responses; no aws/, vertex/ or azure/ twin is listed. No plan tier in the 2026-09-29 plan response lists it, so calls bill per token from the wallet balance; Claude Sonnet 5 stays in the Mid-tier models tier of all three plans. Claude Code 2.1.284 and later send claude-sonnet-5-5 for the sonnet alias, so subscribers who want Sonnet requests on plan quota set ANTHROPIC_DEFAULT_SONNET_MODEL=claude-sonnet-5 — see the Claude Code setup guide and Claude Code in China. Dated details are in the September 2026 new-model guide.

September 27Logs Show Time to First Token; Dashboard Overview Redesigned

Dashboard → Logs shows each request's total time; a streamed request that produced output also shows its time to first token (TTFT). Both are timed by the gateway from the same start — once it has accepted the request and begins executing it — and TTFT ends when the first output (text, reasoning or a tool call) is written back to your client. TTFT therefore includes any failed attempt before failover, and neither timing includes the network between your client and the gateway. Non-streamed requests, and streams that failed before any output, have no TTFT. The logs CSV export carries a TTFT (ms) column next to Latency (ms); a timing that was not recorded exports as an empty cell, never 0. How to read the two timings is on the LLM observability and LLM streaming pages. The dashboard overview now estimates how many days your wallet balance lasts at your average pay-as-you-go spend over the last 7 days (plan-covered requests don't count) and flags it under 3 days, ranks the last 7 days' spend by model, and shows both base URLs, a round-trip check from your browser to api.router.one and platform announcements.

September 26Long-Context Plan Quota Bands Now Cover Claude, GPT-5.5 and Gemini Pro

Update (2026-10-13): Gemini 3.1 Pro Preview is not in the 2026-10-13 catalog — see the Gemini 3.1 Pro Preview entry in October 2026 above.

On Pro, Max and Ultra, a request whose total input is strictly above a model's long-context threshold draws more than 1 request. The catalog now lists bands for Claude Opus 5, Opus 4.8, Opus 4.7 (thinking), Opus 4.6, Sonnet 5 and Sonnet 4.6, Gemini 3.1 Pro Preview and Gemini 3 Flash Preview (2 requests above 256,000 input tokens, 4 above 512,000) and GPT-5.5 (2 above 262,000, 4 above 512,000); GPT-5.6 Sol, GPT-5.6 Terra and GPT-6 Astra gain a second band (4 above 512,000). Cached input counts once, output does not count, and wallet-billed calls keep the posted token rates. The live table sits under the plans on pricing, and the pricing facts list every band.

September 25Console: Per-Key Usage, Logs Filtered by Key, CSV Exports and Account Notifications

Update (2026-10-06): as of 2026-10-06, the in-app low-balance notice goes only to accounts that have topped up their wallet.

Dashboard → Usage ranks your API keys by spend, tokens and requests for the chosen range, and each key links to its own logs. Dashboard → Logs filters by model, API key, status and date range; the API key column (and the request detail) names the key that sent each request — web-console calls are labelled as such. Logs (current filters) and the wallet ledger (chosen range) export to CSV, and the ledger opens on a money-movements view that leaves out per-request API charges; switch to Everything to see each call's charge. A new account settings page shows the account ID to quote to support and sets or changes the sign-in password with an emailed code — accounts created with Google or GitHub can add a password there too. The console also sends in-app notifications when a top-up is credited, when a referral commission is credited, when a key reaches 80% of its maxSpend and when it hits the cap, when the balance runs low, and when a subscription renews or expires; key-cap notices are in-app only, and there are no webhook alerts. For attribution below the key level — per end customer or per run — keep logging your own labels and request IDs; LLM cost attribution per key walks through the workflow.

September 24Routes with a Sustained 502–504 Failure Rate Are Tried Last

When several routes serve the same model, a route whose recent requests show a high share of 502, 503 or 504 failures now moves behind that model's healthy routes. It stays available as a fallback and returns to its usual place once it recovers. A request still gets the model it names — only the order in which same-model routes are tried changes. What counts as a retryable failure is on the routing methodology and LLM fallback pages.

September 24Invoices Are Emailed Automatically

Every payment completed through the hosted checkout (card or Alipay) is invoiced automatically: the invoice is emailed to the address entered at checkout as soon as the payment succeeds, so the email-to-support request step from the August 4 entry is no longer needed. The invoice is issued by the hosted checkout and is not a Chinese VAT invoice (fapiao); pricing, the FAQ, pricing methodology, the payment facts sheets and the card and Alipay payment guide, which had described VAT invoices, now say so. Stablecoin top-ups remain outside invoicing.

September 24Claude Opus 5.5 Listed

The 2026-09-24 catalog re-read contains 63 ids — 56 text and 7 image; none left. anthropic/claude-opus-5.5 is listed: a 1,048,576-token context window, text and image input, the chat, streaming, tool-calling and vision flags, and one rate line across the whole window (no long-context tier). The gateway also accepts Anthropic's own id claude-opus-5-5 for the same model; claude-opus-5.5, anthropic/claude-opus-5-5 and -thinking variants are not accepted. As a Claude-family id it is served on POST /v1/messages and POST /v1/chat/completions, not on POST /v1/responses, so Codex CLI cannot call it; no aws/, vertex/ or azure/ twin is listed. No plan tier in the 2026-09-24 plan response lists it, so calls bill per token from the wallet balance; Claude Opus 5 stays in the Premium models tier of all three plans. Claude Code 2.1.280 and later default to Claude Opus 5.5 and send claude-opus-5-5 through Router One unless a model is pinned, so subscribers who want to keep drawing plan quota set ANTHROPIC_MODEL=claude-opus-5. The Claude Opus 5.5 API guide covers the id, billing and first calls, Claude Code on Opus 5.5 covers model pinning, background tasks, effort and the 1M window, the Opus 5.5 migration guide covers Anthropic's breaking changes and their 400 errors, and Model comparisons gained Claude Opus 5.5 vs Claude Opus 5, Claude Opus 5.5 vs GPT-6 Sol and Claude Opus 5.5 vs Claude Sonnet 5. Dated details are in the September 2026 new-model guide.

September 23GPT-6 Sol and Grok 4.7 Build Fast Listed

Update (2026-10-06): in the 2026-10-06 plan response GPT-6 Sol is in the Premium models tier of the Pro, Max and Ultra plans — see the GPT-6 Sol and GPT-6.1 Sol entry in October 2026 above.

The 2026-09-23 catalog re-read contains 62 ids — 55 text and 7 image; none left. openai/gpt-6-sol is listed (catalog description "GPT-6 Sol": a 1,050,000-token context window, text and image input, the chat, streaming, tool-calling and vision flags, and the same whole-request price tier above 272,000 input tokens as GPT-6 Astra); the bare name gpt-6-sol resolves to the same model. As a GPT-family id it is served on POST /v1/chat/completions and natively on POST /v1/responses, so Codex CLI can point at it; it is not served on POST /v1/messages. No plan tier in the 2026-09-23 plan response lists it, so calls bill per token from the wallet balance. The GPT-6 Sol API guide covers the id, the tier rule and first calls, the coding-tools setup covers Codex CLI, OpenCode, Cline, Cursor, Continue and Goose, and Model comparisons gained GPT-6 Sol vs GPT-6 Astra, GPT-6 Sol vs GPT-5.6 Sol and GPT-6 Sol vs Claude Sonnet 5. grok-4.7-build-fast is also listed (catalog description "Grok 4.7 Build Fast supports text, image understanding and tool calling": a 500,000-token context window, text and image input, the same four flags and a whole-request tier above 200,000 input tokens) as a separate id from grok-4.7 with its own rate lines; as a Grok chat id it is served on Chat Completions and natively on Responses, and the Grok API page lineup lists it. GPT-6 Luna and a channel twin of GPT-6 Sol are not listed. The September new-model guide carries the dated update.

September 22Grok 4.7 Listed; Azure GPT, AWS and Vertex Claude Channel Ids

The 2026-09-22 catalog re-read contains 60 ids — 53 text and 7 image. grok-4.7 is listed (catalog description "Grok 4.7 supports text, image understanding and tool calling": a 500,000-token context window, text and image input, the chat, streaming, tool-calling and vision flags, and the same whole-request price tier above 200,000 input tokens as Grok 4.6) and served on POST /v1/chat/completions and natively on POST /v1/responses — the Grok chat ids (grok-4.7, grok-4.6, grok-4.5, grok-4.3, grok-build-0.1) are served natively on the Codex CLI path from this date; Model comparisons gained Grok 4.7 vs Grok 4.6, Grok 4.7 vs GPT-5.5 and Grok 4.7 vs Gemini 3.8 Flash, and the Grok API page lineup leads with it. Twenty-one sales-channel ids are listed beside their default-channel twins: azure/gpt-5.5, azure/gpt-5.6-luna, azure/gpt-5.6-sol, azure/gpt-5.6-terra and azure/gpt-6-astra (GPT-family: Chat Completions and native Responses), nine aws/claude-* ids and seven vertex/claude-* ids (Claude-family: Chat Completions and Messages). A channel id is a separately priced product of the same base model: routing stays inside the channel with no fallback to the default channel, the default id never borrows channel routes, rates and long-context tiers are the channel's own and do not inherit the default id's discount, a plan quota that covers a default id does not cover its channel twin unless the live plan response lists the channel id, and the response model, the Logs row and the settled charge carry the channel id. The catalog shows the channel and the customer price only; bare and vendor-prefixed ids are unchanged. The channel model ids guide has the full rules and client examples, three channel pairs — GPT-5.6 Sol (Azure) vs GPT-5.6 Sol, Claude Sonnet 5 (AWS) vs Claude Sonnet 5 and Claude Opus 5 (Vertex AI) vs Claude Opus 5 — render the price gap live, and the pricing, API compatibility and routing fact sheets carry the same rules.

September 19Higher Rate Limits on Request, and CONTENT_FILTERED Documented

Every API key and every account runs under default request-per-minute and token-per-minute limits, and a gateway 429 names the one you hit in its X-RateLimit-Scope header — api_key, subject (the account) or subject_model (the account's use of one model), with upstream_provider marking the model vendor's limit instead. These limits are not edited in the dashboard, where maxSpend is the per-key setting; key and account limits can be raised on request — email support@router.one with your account ID, the key and your expected peak traffic. The 429 guide and the error codes page were rewritten around this. The error codes page and the API reference also document 400 CONTENT_FILTERED: a request blocked by content policy before any model is called, and not billed — an image or video prompt that fails content moderation, or a request to a Grok model classified as sexual content involving minors.

Correction (2026-09-25): the gateway does not block Grok chat requests with CONTENT_FILTERED; the code covers image and video prompts that fail content moderation — see the error codes page.

September 19GPT-5.3 Codex Spark Delisted, Three Comparison Pairs, Six Integration Guides

The 2026-09-19 catalog re-read contains 38 ids — 31 text and 7 image. openai/gpt-5.3-codex-spark is no longer listed, and its two comparison pairs — GPT-5.3 Codex Spark vs Claude Sonnet 5 and Grok Build 0.1 vs GPT-5.3 Codex Spark — were retired; posts that named the id now point at currently listed GPT-family ids. Model comparisons gained three pairs: Claude Opus 5 vs Claude Sonnet 5, Claude Fable 5 vs GPT-6 Astra and Grok 4.6 vs Claude Sonnet 5, all rendered from the live catalog. Six client guides went live: Mastra, DSPy, Microsoft Agent Framework, Flowise, Langflow and RAGFlow. Each one states which field takes the base URL, which endpoint the tool calls, and that embeddings stay on another provider, because the gateway serves no embeddings endpoint.

September 18GET /v1/balance for Server-Side Balance Monitoring

GET /v1/balance returns the prepaid wallet balance of the account that owns the key, in USD: balance (spendable right now), reserved_balance (held for in-flight requests) and total_balance. Values are live (Cache-Control: no-store), and the endpoint has its own limit of 60 requests per minute per key that does not draw on inference rate limits. It is built for relay operators and production teams that want a low-balance alert before an empty wallet answers every key with HTTP 402. There is no top-up API; top-ups stay in the dashboard. See the API reference and the balance monitoring guide.

September 18Long Non-Streaming Responses Requests Are Held Open

Update (2026-10-09): the same contract now covers stream: false calls on POST /v1/messages, and a non-streaming call whose client disconnects is billed for the model's actual usage when the model finishes — see the non-streaming entry in October 2026 above.

A stream: false call on POST /v1/responses for a natively served model is now read from the model as a stream and returned as one JSON body, the way long non-streaming Chat Completions calls already were. Once such a request has been running for about 25 seconds, the gateway commits HTTP 200 and writes JSON leading whitespace as a keepalive, so a failure after that point arrives as HTTP 200 with a top-level error object — check for error before reading output. Generations that take minutes, such as hosted web search or agentic runs, should use stream: true. Details are on the error codes page and in Codex and the Responses API.

September 17Long-Context Requests Draw More Subscription Quota

On the Pro, Max and Ultra plans a standard request still uses 1 request from its model tier. A request whose total input is strictly above a model's long-context threshold now uses more, by input-length band: above 128,000, 256,000 and 512,000 tokens on the Gemini 3.5 to 3.8 Flash ids (2, 4 and 8 requests), above 272,000 tokens on GPT-5.6 Sol, GPT-5.6 Terra and GPT-6 Astra (2 requests), and above 200,000 tokens on Grok 4.6 (2 requests). Cached input counts once and output does not count. Wallet-billed calls are unchanged and use the posted token rates. The live table sits under the plans on pricing.

September 11DeepSeek V4.1 Flash and V4 Flash Listed, GPT Image 2.5 Flare and Sunburst Listed, Claude Fable 5.1 and GPT-5.4 mini Delisted

Correction (2026-09-29): 2026-09-08 is the release date the catalog records for the two GPT Image 2.5 ids, not the day they were listed — the 2026-09-09 re-read did not include them, and the 2026-09-11 re-read is the first that does.

Update (2026-09-30): anthropic/claude-fable-5.1 is listed again — see the September 30 entry; its two retired comparison pairs were not restored. openai/gpt-5.4-mini is still not listed.

Update (2026-10-09): Claude Fable 5.1 is not in the 2026-10-09 catalog — see the October 2026 entry above; its two comparison pages stay retired.

The 2026-09-11 catalog re-read contains 39 ids — 32 text and 7 image. deepseek-v4.1-flash (catalog description "DeepSeek V4.1 Flash": a 1,000,000-token context window, text and image input, and the chat, streaming, tool-calling and vision flags) was listed on 2026-09-10, and the older deepseek-v4-flash (text input only) is listed again alongside it; both are served on POST /v1/chat/completions, on POST /v1/messages and natively on POST /v1/responses, so the endpoint rule reads: Messages = the currently listed Claude-family and DeepSeek ids, Responses native = the currently listed GPT-family and DeepSeek ids. Model comparisons gained three pairs: DeepSeek V4.1 Flash vs Gemini 3.8 Flash, DeepSeek V4.1 Flash vs Claude Haiku 4.5 and DeepSeek V4.1 Flash vs GPT-5.5. Two image ids listed on 2026-09-08 — gpt-image-2.5-flare ("GPT Image 2.5 Flare (fast)") and gpt-image-2.5-sunburst ("GPT Image 2.5 Sunburst (high-fidelity)") — are served at POST /v1/images/generations at a flat per-image unit price shown on each model page, and the image generation API page lists them. anthropic/claude-fable-5.1 and openai/gpt-5.4-mini are no longer listed (Claude Fable 5 and Claude Opus 5 remain; no mini GPT id is currently listed), and their five comparison pairs — Claude Fable 5.1 vs GPT-6 Astra, Claude Fable 5.1 vs Claude Opus 5, Gemini 3.7 Flash vs GPT-5.4 mini, Gemini 3.1 Flash Lite vs GPT-5.4 mini and Claude Haiku 4.5 vs GPT-5.4 mini — were retired. The September 2026 new-model guide carries the dated update.

September 9GPT-5.4 and GPT-5.6 Luna Delisted; Gemini 3.1 Pro Preview Listed Again

Update (2026-10-13): Gemini 3.1 Pro Preview is not in the 2026-10-13 catalog — see the Gemini 3.1 Pro Preview entry in October 2026 above.

The 2026-09-09 catalog re-read contains 37 ids. openai/gpt-5.4 and openai/gpt-5.6-luna are no longer listed (GPT-5.6 Sol and GPT-5.6 Terra remain), and google/gemini-3.1-pro-preview, which had left between the 2026-09-01 and 2026-09-05 snapshots, is listed again: a 1,048,576-token context window, text and image input, and the chat, streaming, tool-calling and vision flags; as a Gemini id it answers on POST /v1/chat/completions. GPT-5.4 mini, a separate id, was still listed that day and had left by 2026-09-11 (see the September 11 entry). The two comparison pairs built on GPT-5.4 — Claude Sonnet 5 vs GPT-5.4 and GPT-5.6 Terra vs GPT-5.4 — were retired with it. The September 2026 new-model guide carries the dated update.

September 7September Model Wave: Eight Ids Listed, Twelve Delisted, Six Comparison Pairs, Five Integration Guides, and a -thinking Default

Update (2026-09-29): two ids in the delisted list below are listed again — google/gemini-3.1-pro-preview since the 2026-09-09 re-read (see the September 9 entry) and deepseek-v4-flash since 2026-09-10 (see the September 11 entry); the other ten were still not listed on 2026-09-29.

Update (2026-10-13): of those two, Gemini 3.1 Pro Preview is not in the 2026-10-13 catalog — see the Gemini 3.1 Pro Preview entry in October 2026 above.

Between the 2026-09-01 and 2026-09-05 catalog snapshots, eight ids were listed — anthropic/claude-fable-5.1, openai/gpt-6-astra, google/gemini-3.8-flash, google/gemini-3.1-flash-lite, google/gemini-3.1-flash-lite-preview, google/gemini-2.5-pro, grok-4.3 and grok-build-0.1 — and twelve left: deepseek-v4-pro, deepseek-v4-flash, glm-5.1, glm-5.2, glm-5.2-fast, MiniMax-M2.7, doubao-seed-2-0-lite-260428, doubao-seedream-5-0-260128, google/gemini-3.1-pro-preview, google/gemini-3-pro-preview-11-2025, openai/o3 and openai/o4-mini. The September 2026 new-model guide covers each new id's catalog entry, endpoint family and whole-request price tier without reprinting rates. Model comparisons gained six pairs: Claude Fable 5.1 vs GPT-6 Astra, Claude Fable 5.1 vs Claude Opus 5, GPT-6 Astra vs Claude Opus 5, GPT-6 Astra vs GPT-5.6 Sol, Gemini 3.1 Flash Lite vs GPT-5.4 mini and Grok Build 0.1 vs GPT-5.3 Codex Spark, all rendered from the live catalog (the two Claude Fable 5.1 pairs and the GPT-5.4 mini pair were retired on 2026-09-11, and the Codex Spark pair on 2026-09-19, when those ids left the catalog). Five client guides went live: GitHub Copilot (its Custom Endpoint provider, so Copilot Chat and agent sessions run on your own key), OpenHands, NextChat, CrewAI and Spring AI. The API compatibility fact sheet now documents the -thinking default: a request to a catalog id ending in -thinking that carries no thinking field is sent with thinking enabled (adaptive where the gateway adds the field), and a request that already carries a thinking value keeps it (disabled is forwarded as sent; the Opus 4.7-generation ids translate a legacy enabled/budget_tokens value to adaptive).

September 1Gemini 3 Flash GA Listed Alongside the Preview

google/gemini-3-flash, the general-availability Gemini 3 Flash, joined the catalog next to google/gemini-3-flash-preview — chat, streaming, tool calling and vision, 1M context — and the Gemini API page lineup lists both ids. Specs and the live rate are on the model catalog; the GA and preview ids are priced separately.

September 1GPT-5.6 Luna Back in the Catalog

Update (2026-09-29): openai/gpt-5.6-luna left the catalog again — the 2026-09-09 re-read no longer lists it (see the September 9 entry), and it was still not listed on 2026-09-29. azure/gpt-5.6-luna, listed 2026-09-22, is a separately priced Azure-channel product, not this id.

openai/gpt-5.6-luna, the GPT-5.6 family's high-volume budget tier, is listed again after leaving the catalog on 2026-08-22 — the same roughly 1M context as Sol and Terra. The model catalog is the source of truth for its availability, and the August 2026 model guide carries the updated status note.

September 1Anthropic code_execution Accepted on /v1/messages

On POST /v1/messages, the Anthropic code_execution server tool — including its bash and text-editor sub-tools and container reuse across turns — is now accepted and metered at the model's standard token rate, joining web_search; web_fetch and mcp_servers are still rejected with a 400 invalid_request_error before any model is called. The endpoint-by-family map is on the API compatibility fact sheet.

August 2026

August 26Grok 4.20 Left the Catalog; Its Comparison Pair Retired

The Grok 4.20 non-reasoning tier is no longer in the catalog, so the Grok API page lineup now lists grok-4.6 and grok-4.5 and the Grok 4.20 vs DeepSeek V4 Flash comparison pair was retired, following the Kimi precedent. Model comparisons also gained an answer capsule under each title, a dated "rates read from the live catalog" stamp beside the spec table, catalog-filtered related pairs, and links to the family landing pages, the cost calculator and the pricing methodology.

August 25Nano Banana 2 API Guide and an API Reference Refresh

Correction (2026-09-29): Codex does not send codex-auto-review for /review. Per the codex-rs source (v0.100.0 through v0.159.0, checked 2026-09-29), /review uses review_model when it is set and otherwise the session model; codex-auto-review is an id Codex's Auto-review may request (per the rust-v0.159.0 source), and it is in the catalog but in no plan tier, so those requests bill the wallet. The Codex / Responses API page carries the current explanation.

A new Nano Banana 2 API guide covers gemini-3.1-flash-image-preview end to end: text-to-image at POST /v1/images/generations, reference-image edits at POST /v1/images/edits, flat per-image pricing, and when to pick it over Nano Banana Pro (gemini-3-pro-image-preview); the image generation API page now carries both names on its Gemini cards. The API reference was brought in line with the gateway: GET /v1/models is documented, error bodies show the UPPER_SNAKE_CASE code values and the request_id the gateway actually returns, the 429 rows say to read Retry-After and name the three quota codes, a new section lists the rate-limit response headers, /v1/messages documents its 400 in the Anthropic error envelope for a model outside the Claude and DeepSeek V4 families, and the image and video examples use catalog model IDs with the fixed clip specs. The Codex / Responses API page explains the codex-auto-review model ID Codex sends for /review.

August 24Video Page Rewritten, Four New Comparison Pairs, and Gemini 3.1 Flash Image in the Image Lineup

The video generation page now leads with the video models the catalog actually carries and the fixed clip each one produces, and states plainly that Veo and Grok Video are not in the catalog. The image generation API page names Gemini 3.1 Flash Image alongside gpt-image-2, Gemini 3 Pro Image, Doubao Seedream and the two Grok Imagine entries, and documents reference-image edits at POST /v1/images/edits. The error codes page opens with a 400 row covering a model called on the wrong endpoint. Model comparisons gained four pairs — DeepSeek V4 Pro vs DeepSeek V4 Flash, Gemini 3.1 Pro vs GPT-5.6 Sol, Grok 4.20 vs DeepSeek V4 Flash and MiniMax M2.7 vs GLM-5.2 (all four have since been retired: Grok 4.20 was delisted on 2026-08-26 and the other three pairs' models left the catalog on 2026-09-05) — while three Kimi pairs were retired, because Kimi is not in the catalog.

August 23Six Video Models, Each with a Fixed Clip Length and Resolution

Correction (2026-09-29): a request's trace in Dashboard → Logs records its cost, not a unit price; the per-clip unit prices were shown on each model page. None of the six video models below is listed today: the catalog listed no video model as of 2026-09-05, and none on 2026-09-29.

viduq3-turbo (5s, 720p), viduq3-turbo-540p (4s, 540p), pixverse-video (5s, 540p), MiniMax-Hailuo-2.3 and MiniMax-Hailuo-02 (6s, 768P), and wan2.6-i2v (5s, 720P, image-to-video — exactly one image_url) are served at POST /v1/videos/generations: submit a task, poll it by ID, download the clip. Each model produces one fixed combination of clip length and resolution, so a duration or size field in the request body changes nothing about the clip you get, and every job costs one flat per-clip unit price recorded in its request trace. Live prices are on the models page under the video category; the lineup and call shape are on the video generation page.

August 23Gemini Image Models: Generation and Reference-Image Edits

The catalog's two Gemini image entries — gemini-3-pro-image-preview and gemini-3.1-flash-image-preview — serve text-to-image at POST /v1/images/generations after a fix to the gateway's image path, and both list image input, so either one takes a reference image at POST /v1/images/edits (multipart/form-data, the file in the image field). Both are unit-priced per generated image like the rest of the image lineup. Details are on the image generation API page.

August 23Claude Model IDs Rejected on /v1/responses with a 400

A Claude-family model ID sent to /v1/responses is now rejected with a 400 invalid_request_error before any model is called — model '<id>' must be called via /v1/messages or /v1/chat/completions — so the mistake costs no tokens and returns a message that names the fix. /v1/responses is served natively for GPT models (and, at the time, DeepSeek V4 — delisted 2026-09-05; DeepSeek returned on 2026-09-10 as deepseek-v4.1-flash and deepseek-v4-flash, both again served natively there); Claude models answer on /v1/messages and on /v1/chat/completions. Each model's detail page lists the endpoints it speaks, and the full endpoint-by-family map is on the API compatibility fact sheet.

August 23An Empty Reply That Reports Usage Is Served as 200

A reply that carries no content, no tool calls and no refusal but reports token usage is a completed response, not a failure: it is returned as HTTP 200 exactly as it came back, with no retry and no failover, and settlement follows the usage the response reports. Don't auto-retry it — read finish_reason, refusal and usage on the reply, since a retry loop pays for the same silence again. A reply that reports no usage at all is still treated as an upstream failure: the gateway tries the next candidate and returns a 502 when none succeeds. The troubleshooting answer is on the error codes page.

August 22Structured Outputs Guide, Responses Endpoint Matrix, Gemini 3.7 Flash Comparison Pairs, an API Compatibility Fact Sheet, and Five More Markdown Mirrors

A new LLM structured outputs guide documents response_format on the OpenAI-compatible endpoint — what the gateway validates before any upstream call, and why schema enforcement is model-side. The Codex / Responses API page gained an endpoint matrix listing which Responses request features are accepted and which are rejected. Model comparisons gained two Gemini 3.7 Flash pairs: Gemini 3.7 Flash vs GPT-5.4 mini (retired 2026-09-11 when GPT-5.4 mini left the catalog) and Gemini 3.7 Flash vs Gemini 3.6 Flash. A machine-readable API compatibility fact sheet states which endpoint families serve which model families. And five trust pages — Security, Data retention, Routing methodology, Pricing methodology, and the China latency benchmark — now ship .md mirrors, so AI crawlers read exactly the same copy as the rendered pages.

August 20Structured Outputs: json_schema Passthrough on Chat Completions

response_format with type: "json_schema" is now validated at the gateway and forwarded to the model unchanged. A request whose json_schema block is missing its name or schema is rejected with a 400 invalid_request_error before any upstream call — response_format.json_schema.name is required / response_format.json_schema.schema is required — so a malformed request costs no tokens. json_object mode was already passed through. Enforcement is model-side: the output is constrained when the model supports it, so parse and validate on the client. When a model route rejects a schema that uses $defs or $ref, the gateway treats the rejection as a candidate-specific 400 and fails over to the next candidate for the same model. Details and request shapes are on the structured outputs guide.

August 19Responses API: Hosted Tools, File Inputs, Service Tiers, and Server-Side Context Accepted

On /v1/responses, requests that use custom tools, previous_response_id, conversation, prompt, hosted tools (file_search, code_interpreter, computer_use, mcp, web_search), file_id / file_url input parts, and any service_tier value are now accepted on models whose detail page lists POST /v1/responses, billed at the model's standard rate. Function tools were already supported. Still rejected with a 400 invalid_request_error: the image_generation tool and image_generation_call items (use /v1/images/generations instead) and background: true. The full matrix is on the Codex / Responses API page.

August 17Gemini 3.7 Flash Listed in the Catalog

google/gemini-3.7-flash joined the catalog with chat, streaming, tool calling, and vision, and a 1M-token context window. Live rates are on the Gemini 3.7 Flash model page.

August 16New-Model Wave Coverage: Five Comparison Pairs, an August 2026 Model Guide, n8n and Immersive Translate Integration Guides, and an SLA Markdown Mirror

Correction (2026-09-27): Immersive Translate's custom API address no longer has to be the full /chat/completions path. Versions 1.30.2 and later also accept a base URL ending in /v1, as the field's own help text now says, and 1.33.3 completes even a bare host; 1.30.2, the build Firefox Add-ons still serves, turns a bare host into /chat/completions and gets a 404. The full https://api.router.one/v1/chat/completions works in every version and remains the recommended value — see the updated Immersive Translate guide.

The August 2026 new-model guide is the narrative companion to the catalog's August 2026 generation — Claude Opus 5, the GPT-5.6 family, Grok 4.6, Kimi K3 and DeepSeek V4 Flash (Kimi K3 has since left the catalog; DeepSeek V4 Flash left on 2026-09-05 and was listed again on 2026-09-10) — with a route-key A/B pattern for trying them through one key. Two client guides went live: n8n, the workflow automation platform — point its OpenAI credential's Base URL at the gateway and every step's model call lands in one ledger under a per-key maxSpend cap — and Immersive Translate, the bilingual browser translation extension, whose custom OpenAI-compatible endpoint takes the full /chat/completions path. The Grok API page lineup now covers Grok 4.6, and the SLA page ships a markdown mirror for AI crawlers. Model comparisons gained five pairs: Grok 4.6 vs GPT-5.6 Sol, Grok 4.6 vs GPT-5.6 Terra, Grok 4.6 vs Grok 4.5, GPT-5.6 Sol vs GPT-5.6 Terra, and Doubao Seed 2.0 Lite vs DeepSeek V4 Flash — all rendered from the live catalog.

August 15DeepSeek V4 on the Anthropic-Compatible /v1/messages Endpoint

deepseek-v4-pro and deepseek-v4-flash now answer on /v1/messages, the Anthropic wire-format endpoint Claude Code uses — so Claude Code can point at either model with the same base URL and key it already uses for the Claude family. /v1/chat/completions continues to serve every chat model; each model's detail page lists the endpoints it speaks. Both DeepSeek V4 ids left the catalog on 2026-09-05; on 2026-09-10 deepseek-v4-flash was listed again next to the new deepseek-v4.1-flash, and /v1/messages serves the currently listed Claude-family and DeepSeek ids (V4 Pro remains unlisted).

August 12Zed, Continue, and Open WebUI Integration Guides, a Grok API Page, and Three New Comparison Pairs

The client integration catalog grew by three: Zed, Continue, and Open WebUI each got a setup guide with the exact configuration shape the current client version expects, verified against the live client rather than transcribed from old docs. A dedicated Grok API access page covers calling the Grok family through the gateway, including mainland-China reachability. And the model comparison set gained three pairs: Claude Fable 5 vs Claude Opus 5, Claude Fable 5 vs GPT-5.6 Sol, and Grok 4.5 vs GPT-5.5 — all rendered from the live catalog so specs and rates cannot go stale.

August 7Cursor and OpenClaw Integration Guides

Two more setup guides went live: Cursor, covering the custom base URL and key configuration that shares one wallet and one request trace with the CLI tools, and OpenClaw, covering how to point the agent at the OpenAI-compatible endpoint. Both pages ship bilingual.

August 4Invoicing for Hosted-Checkout Top-Ups

Invoicing policy is now uniform: any top-up completed through the hosted checkout (card or Alipay) can be invoiced — ordinary pay-as-you-go wallet top-ups included, not only enterprise or subscription accounts. Stablecoin top-ups are a separate rail and are outside the invoicing scope. At the time, invoices were requested by emailing support@router.one with the account ID; since September 24 they are emailed automatically after payment (see the September 24 entry).

July 2026

July 24One-Click Claude Code and Codex Setup, Plus Quickstart and FAQ Guides

Connecting a CLI no longer means hand-editing config files. A single install command configures Claude Code (writing the gateway base URL and auth token into the env block of ~/.claude/settings.json) or Codex CLI (writing ~/.codex/config.toml with env_key and appending the key export to your shell profile). Scripts ship for macOS, Linux, and Windows PowerShell. The docs gained two companion pages: a quickstart that gets a first request out in five minutes across both protocol surfaces, and an FAQ and troubleshooting guide covering base-URL selection and 401/402/429 debugging.

July 24Alipay on the Hosted Checkout, with a Free Amount Input

Wallet top-ups now run through one hosted checkout that offers a card or Alipay, and the amount field accepts any value instead of only fixed tiers. Top up with a card or Alipay through one hosted checkout, or with USDT/USDC on six chains (Tron, BSC, Ethereum, Polygon, Base, Arbitrum). No US credit card required.

July 18Image and Video Generation Scenario Pages

Three new surfaces cover generation workloads: the image generation API page for text-to-image models behind the OpenAI-compatible endpoint, the Veo video API page for asynchronous video generation, and a bilingual blog guide for calling both from China. Media models are priced per generated image or clip rather than per token, and each model's unit price lives in the model catalog.

July 10Automatic Site Language Detection

The marketing site now negotiates language on a first visit using the browser's Accept-Language header, and remembers a manual switch for a year. English stays at the apex and Chinese under /zh, so every page keeps one canonical URL per locale.

June 2026

Playground Experience Upgrade with Image Generation

The playground received a full experience upgrade. Streaming responses render more smoothly with polished markdown and code highlighting, errors come with actionable next steps, Esc stops a generation in progress, and the model selector shows per-call pricing. Image generation is now a first-class playground mode: pick an image model, write a prompt, optionally attach up to four reference images, and generate directly in the conversation flow.

Four New Landing Pages and Expanded Machine-Readable Facts

Four new bilingual landing pages went live — /smart-routing, /llm-cost-tracking, /helicone-alternative, and /use-cases/ai-coding-tools — completing the gateway content cluster. The machine-readable facts surface also expanded with /facts/routing.md and /facts/pricing.md, plus markdown mirrors for the gateway and comparison pages, so AI crawlers read exactly the same copy as the rendered pages.

CC Switch GUI Setup Guide

The docs now include a guide for CC Switch, a GUI tool that configures Claude Code and OpenAI Codex base URLs and API keys without hand-editing config files. The guide is available in English and Chinese.

May 2026

Subscription Plans Launch

Pro, Max, and Ultra subscription plans went live on May 21 for teams that want predictable monthly pricing. Wallet pay-per-token billing remains available unchanged — subscriptions run alongside it, not in place of it. Full details are in the launch announcement.

April 2026

Google and GitHub OAuth Login

You can now sign in to Router One using your Google or GitHub account. This complements the existing email-based authentication and makes onboarding faster for developers who already use these identity providers. OAuth and email sign-in reach the same Router One account and dashboard; Router One does not currently expose an organization/RBAC management layer.

Google Search Console Integration

Correction (2026-09-29): this integration covers Router One's own website only. Router One does not host customer documentation or landing pages and offers users no SEO tooling, so the last sentence below does not apply.

We have integrated Google Search Console verification and sitemap management into the platform. This is part of our ongoing effort to improve the discoverability of Router One's documentation and public-facing pages. For users building on our platform, this also means better SEO tooling for any documentation or landing pages served through Router One.

Improved Dashboard Performance

Correction (2026-09-29): no measurement or method was published for the "40 percent faster" figure below, and we cannot substantiate it today, so it is withdrawn.

The main dashboard now loads 40 percent faster thanks to server-side rendering optimizations and lazy loading of chart components. Usage graphs and cost breakdowns render incrementally, so you see critical numbers immediately without waiting for the full page to hydrate.

March 2026

Model Marketplace with Detailed Pricing

The new model marketplace gives you a clear view of every LLM available through Router One, with real-time pricing per input token and output token, capability tags, and context window sizes. Compare models side by side and make informed routing decisions without leaving the dashboard.

At launch the marketplace listed models from OpenAI, Anthropic, Google, Mistral, DeepSeek, and Meta. The catalog turns over continuously — the models page is the source of truth for what is callable today.

Full i18n Support: English and Chinese

Correction (2026-09-29): the language is not set in account settings. It follows the URL — English at the root, Chinese under /zh — and the language switcher in the site header changes it; first-visit detection from the browser's Accept-Language header arrived on July 10 (see that entry). We also cannot substantiate the claim below that every translation was reviewed by native speakers, so it is withdrawn.

The entire Router One dashboard and documentation are now available in both English and Chinese. Language selection is automatic based on browser locale and can be overridden in account settings. All UI text, error messages, help content, and API documentation are fully translated — not machine-translated, but reviewed by native speakers for accuracy and clarity.

Real-Time Usage Dashboard

Correction (2026-09-29): three claims in this entry do not match the dashboard. The usage page does not refresh on a timer — reload it to fetch newer figures. Its charts group usage by day; there is no hourly view. And there is no usage anomaly indicator: to bound spend, give each key a maxSpend cap, and the console sends in-app notices when a key reaches 80% of its maxSpend and when it hits the cap (see the September 25 entry).

The usage dashboard now updates in real time with no manual refresh required. Watch token consumption, cost accrual, and request volume as they happen. The dashboard includes:

• Per-model breakdown — see exactly how much each model is costing you
• Per-key breakdown — give each project, tool, or team a separate key and track its usage
• Per-request traces — inspect model, token, cost, latency, status, and key metadata
• Time-series charts — visualize trends over hours, days, or weeks
• Anomaly indicators — automatic highlighting when usage deviates significantly from historical patterns

Landing Page Redesign

The public-facing website received a complete redesign with improved messaging, faster load times, and a clearer explanation of Router One's value proposition: smart routing, automatic fallback, and per-request observability for LLM API traffic from mainland China.

February 2026

Provider Route Selection and Health Signals

Correction (2026-09-25): production routing is exact-model routing with same-model failover. A request is served by the model it names; a route that fails with a 429, a 5xx or a timeout is retried on another route for the same model, and routes with a sustained upstream-error rate are moved to the back of the order until they recover (see the September 24 entry). Routes are not ranked by measured latency or posted cost. The current contract is on the routing methodology page.

The core routing engine is now live. A request names a model and the gateway selects an eligible provider route. Server-side adaptive strategies can use recent EWMA latency, posted cost, and reliability signals when ranking configured candidates.

The public Dashboard does not expose project-level or API-key routing weights. Candidate policy is configured on the gateway, while the request trace records the model and request ID.

Automatic Failover

When a provider returns a retryable supply, rate-limit, service, or timeout failure, Router One can try another eligible candidate. Health state and cooldowns help deprioritize degraded routes and reintroduce them after recovery.

Actual fallback time includes the failed attempt and the next provider response, so there is no universal millisecond guarantee. The Dashboard records the model, request ID, usage, cost, latency, and status; support can investigate issues using the request ID.

WeChat and Alipay Payment Integration

Superseded (2026-07-25): WeChat Pay is not currently supported. RMB top-up now runs through Alipay on the hosted checkout, alongside cards.

Chinese users could add funds to their Router One account using WeChat Pay and Alipay, in addition to international cards. Payments were processed in real time with immediate balance updates. This removed a major friction point for teams operating primarily in China.

Budget Controls and QPS Limits

Each API key can now carry its own maxSpend cap, and every key runs under request (rateLimit) and tokens-per-minute (tokenLimitTpm) limits set at platform defaults, which support can raise on request. Give each application, project, or tool a separate key to isolate its exposure. When maxSpend is exhausted, requests on that key stop even if the account wallet still has balance.

January 2026

Platform Launch

Router One is live. The initial release includes the foundational infrastructure that everything else is built on:

Unified LLM API Endpoint. A single OpenAI-compatible endpoint (https://api.router.one/v1) that accepts Chat Completions requests and routes them to the right model. GPT, Claude, Gemini, Mistral, and DeepSeek model families supported at launch (the catalog has since turned over — see the models page), behind one request and response schema.

Claude Code and Codex Support. Router One works as a drop-in API base URL for both Anthropic's Claude Code CLI and OpenAI's Codex CLI. Configure the base URL, provide your Router One API key, and all requests flow through the platform with full tracking and budget enforcement.

API Key Isolation. Create multiple API keys under one account and assign a separate key to each application, project, tool, or environment. Each key has its own spend cap and is rate-limited on its own at platform defaults; organization roles and RBAC are not part of this public release.

Observability Foundation. Every request is logged with billing and routing metadata: model used, tokens consumed (input and output), cost incurred, latency measured, status, and request ID. This data powers the usage dashboard and per-request traces. Current data-handling clarification (September 5, 2026): Router One does not retain prompt or model-response bodies from direct API calls. It records request metadata for billing, usage tracking, and troubleshooting. Playground saves conversation history so users can revisit and continue their chats.

Prepaid Credit System. Add funds to your account and consume them as you use LLM APIs. You see the wallet balance decrease with usage and can top up when needed; a key-level maxSpend cap limits an individual workload.

What Is Coming Next

We do not pre-announce features before they ship — roadmap items appear in this changelog once they are live. Subscription plans, previously listed here as upcoming, launched on 2026-05-21; see the May 2026 entry above.

Follow this changelog for updates as new releases ship.