Availability update — 2026-09-25: This is a May 2026 snapshot. Of the models compared, GPT-5.5, Claude Sonnet 4.6, Claude Opus 4.7 (as its thinking variant) and Gemini 3.1 Pro Preview were listed on 2026-09-25, alongside DeepSeek V4 Flash and the newer DeepSeek V4.1 Flash, while DeepSeek V4 Pro, Doubao and Qwen were not; as of the 2026-10-13 catalog, Gemini 3.1 Pro Preview is no longer listed — the live catalog is the source of truth, and earlier notes are in the availability history at the end.
For Chinese developers in May 2026, the choice is no longer "use GPT or use Claude." The frontier shifted dramatically in April: Claude Opus 4.7 dropped April 16, GPT-5.5 on April 23, DeepSeek V4 on April 24. Domestic models — Qwen 3.5 / 3.6 from Alibaba, Doubao 2.0 from ByteDance, DeepSeek V4 — have closed enough of the gap that for many tasks they are first-pick, not fallback. Add the network reliability advantage (domestic models are hosted in Chinese data centers and reachable in 30 ms instead of 300 ms), and the calculus shifts further.
This guide is the head-to-head: Qwen 3.5, Doubao 2.0, and DeepSeek V4 against Claude Opus 4.7 / Sonnet 4.6 and GPT-5.5, on the dimensions that decide whether to use a Chinese or international model for a given task. Coding, reasoning, multilingual quality, latency, cost, and where each model breaks.
The May 2026 Landscape
A short orientation before the comparisons; each entry describes the field as it stood in May 2026.
Qwen 3.5 / 3.6 is Alibaba's flagship line. Qwen 3.5 (released February 16, 2026) ships a 397B-A17B Mixture-of-Experts model and is the data-rich tier with public benchmarks; Qwen 3.6 Max (April 20, 2026) is the newer, even-more-capable preview. Strong all-rounder; particularly good at Chinese-language tasks and coding.
Doubao 2.0 is ByteDance's flagship (February 14, 2026), optimized for low latency and high throughput. Doubao 2.0 Pro is the API tier; Doubao Vision adds multimodal. ByteDance positions it at parity with GPT-5.2 / Gemini 3 Pro on math, coding, and logical reasoning.
DeepSeek V4 (released April 24, 2026) ships in two sizes: V4-Pro at 1.6T total / 49B active (MoE), and V4-Flash at 284B. Open-weight option; the hosted API is also available. Notably strong at coding and math; meaningfully cheaper than competitors.
Claude Opus 4.7 (April 16, 2026) — Anthropic's top tier at the time. 1M context, best agentic loop in production, deepest reasoning. Claude Sonnet 4.6 is the everyday tier; Claude Haiku 4.5 the cost tier.
GPT-5.5 (April 23, 2026) — OpenAI's flagship at the time. 1.05M context. Strong instruction following; broad coverage. GPT-5.5 pro is the high-accuracy variant and sits in a much higher cost band than the standard tier.
Gemini 3.1 Pro (February 19, 2026) — Google's flagship, 1M context, with a higher rate band for prompts over 200K tokens. Long-context champion.
For deeper coding-specific benchmarks see DeepSeek V3 vs Claude 4 vs GPT-4.1 for coding; for the August 2026 catalog generation — Claude Opus 5, the GPT-5.6 family, Grok 4.6 — see the August 2026 new-model guide (its DeepSeek V4 Flash section is historical; V4 left the catalog on 2026-09-05, and deepseek-v4-flash was listed again on 2026-09-10 next to the new deepseek-v4.1-flash).
Coding
Public benchmark scores as of May 2026. HumanEval has saturated at the frontier (top models cluster 93-99%), so we omit it; SWE-bench Verified and LiveCodeBench are the discriminating tests in 2026:
| Model | SWE-bench Verified | LiveCodeBench |
|---|---|---|
| GPT-5.5 | 88.7% | 92% |
| Claude Opus 4.7 | 87.6% | 89% |
| DeepSeek V4-Pro | 80.6% | 93.5% |
| Gemini 3.1 Pro | 80.6% | 84% |
| DeepSeek V4-Flash | 79.0% | 91.6% |
| Claude Sonnet 4.6 | 77% | 82% |
| Qwen 3.5 (397B) | 76.4% | 83.6% |
| Doubao 2.0 Pro | ~72% | ~78% |
The story changed in late April. GPT-5.5 took the SWE-bench Verified #1 slot (88.7%) just edging out Opus 4.7 (87.6%); both are meaningfully ahead on cross-file repo tasks. DeepSeek V4-Pro leads on LiveCodeBench (93.5%) — competitive programming is its specialty — and is essentially tied with Gemini 3.1 Pro on SWE-bench at 80.6%. Qwen 3.5 is solid in the second tier; Doubao 2.0 trails on hard coding but ships the lowest latency.
For day-to-day coding through Claude Code, Cursor, or Cline, Claude Sonnet 4.6 is the daily driver and Opus 4.7 is the "hard task" backup. DeepSeek V4-Pro is the credible cost-conscious alternative — competitive on benchmarks from a materially lower cost tier.
Reasoning
Reasoning means: structured thinking, math, logic, "what would happen if" simulations. Public benchmarks as of May 2026 (MMLU-Pro is near-saturated at the frontier, so the spread is narrow):
| Model | GPQA Diamond | AIME 2026 |
|---|---|---|
| Qwen 3.5 (397B) | 88.4% | 91.3% |
| Claude Opus 4.7 | 84% | 90% |
| GPT-5.5 | 83% | 90% |
| DeepSeek V4-Pro | 80% | 89% |
| Gemini 3.1 Pro | 78% | 87% |
| Claude Sonnet 4.6 | 73% | 83% |
| Doubao 2.0 Pro | 65% | 75% |
Two surprises: Qwen 3.5 leads GPQA Diamond at 88.4% — the highest of any tracked model — and is competitive on AIME 2026. DeepSeek V4-Pro continues the V3 tradition of being notably strong at math. The Chinese models are competitive on reasoning benchmarks now, not "second tier with a discount." Check Chinese-language reasoning separately for tasks where that matters.
Chinese-Language Quality
This is where domestic models lead, sometimes meaningfully. We've found:
- Native Chinese conversation: Qwen 3.5 and Doubao 2.0 read more naturally than translated-feeling Claude/GPT output. The difference is most visible in longer-form generation (essays, marketing copy, cultural commentary).
- Idiom and cultural context: Qwen 3.5 handles 成语, regional dialects, and cultural references better than Claude or GPT. GPT-5.5 has improved noticeably here but is still second-tier.
- Technical writing in Chinese: Surprisingly, Claude Sonnet 4.6 holds up well — possibly because of strong code+technical training. Qwen 3.5 is excellent.
- Code with Chinese comments: All models handle this; Qwen 3.5 and DeepSeek V4 are slightly more natural at it.
For consumer-facing Chinese content (chatbots, customer service, content generation), Qwen 3.5 and Doubao 2.0 are usually the right starting point. For technical-internal Chinese content (docs, internal tools), Claude Sonnet 4.6 is competitive.
Latency
Network plus inference. Numbers from a Beijing connection in May 2026:
| Model (provider) | Time-to-first-token | Tokens/sec |
|---|---|---|
| Doubao 2.0 Pro (Volcengine) | 0.3-0.5 s | 80-120 |
| Qwen 3.5 (Alibaba Cloud) | 0.4-0.6 s | 70-100 |
| DeepSeek V4-Flash (DeepSeek API) | 0.4-0.8 s | 60-100 |
| DeepSeek V4-Pro (DeepSeek API) | 0.5-1.0 s | 40-80 |
| Claude Sonnet 4.6 (via Router One) | 0.6-1.2 s | 50-80 |
| Claude Opus 4.7 (via Router One) | 0.7-1.5 s | 40-70 |
| GPT-5.5 (via Router One) | 0.7-1.4 s | 50-90 |
| Gemini 3.1 Pro (via Router One) | 0.6-1.3 s | 50-90 |
Domestic models running on Chinese cloud infrastructure are substantially faster from China than international models, even routed through Router One. For latency-sensitive applications (chatbots, autocomplete, voice), the domestic models win on this dimension regardless of quality.
Cost
We do not reprint provider rate cards here — they move, and a stale table is worse than no table. What is stable enough to plan against is where each model sits relative to the others on price. This column is price only; it says nothing about capability:
| Model | Relative price |
|---|---|
| DeepSeek V4-Flash | Lowest |
| DeepSeek V4-Pro | Lowest |
| Doubao 2.0 Pro | Lower |
| Qwen 3.5 (397B) | Lower |
| Claude Sonnet 4.6 | Middle |
| Gemini 3.1 Pro | Middle, higher for prompts over 200K tokens |
| GPT-5.5 | Higher |
| Claude Opus 4.7 | Higher |
| GPT-5.5 pro | Highest |
DeepSeek V4 redrew the cost frontier on April 24: it landed at the lowest price point while scoring within 8 points of the frontier models on SWE-bench Verified. For high-volume tasks where 90% quality is enough, V4-Pro is the new default. For tasks where quality matters more than cost (one-off agent work, hard reasoning, long-horizon agents), Opus 4.7 or GPT-5.5 pro is worth the premium.
The models above that Router One carries are priced on the model catalog, each listed next to its context window and capability tags, and pricing goes up to 90% off official prices on select models.
Where Each Model Breaks
Knowing failure modes is more useful than scoring averages.
Claude Opus 4.7 / Sonnet 4.6 — Sometimes too cautious; will add unrequested error handling, validation, "what about this edge case" thinking. The Opus 4.7 tokenizer is also new and uses up to 35% more tokens than 4.6 for the same text, which inflates real-world cost. Mitigation: instruct precisely on scope; budget extra for tokenizer overhead.
GPT-5.5 — Can hallucinate API signatures more readily than Claude on niche topics; it also repriced upward from GPT-5, which is noticeable for high-volume use. Mitigation: verify generated code against actual library docs; reach for cheaper models on simple tasks.
DeepSeek V4-Pro — Output formatting drifts under long context; sometimes forgets schema constraints late in long generations. Mitigation: re-send schema in critical sections.
Qwen 3.5 — Tends to over-explain in Chinese; brevity prompts work less reliably than they do for Claude. Mitigation: explicit "no preamble, no postscript" instruction.
Doubao 2.0 — Less precise on hard coding tasks; can produce plausible-but-wrong patterns at edge cases. Mitigation: pair with a stronger coding model for review.
Practical Routing Strategy
Most teams in China end up with a 3-tier strategy:
- High-volume cheap tier for chatbot turns, classification, summarization: DeepSeek V4-Flash, Doubao 2.0 Pro, or Qwen 3.5 32B — at the time compared side by side on the Doubao Seed 2.0 Lite vs DeepSeek V4 Flash and DeepSeek V4 Flash vs GPT-5.4 mini pages, both since retired (GPT-5.4 mini left the catalog by 2026-09-11); the DeepSeek ids listed as of 2026-09-11 are
deepseek-v4.1-flashanddeepseek-v4-flash, and V4.1 Flash is paired on DeepSeek V4.1 Flash vs Claude Haiku 4.5. - Mid-tier for writing, structured reasoning, most coding: DeepSeek V4-Pro, Qwen 3.5 397B, or Claude Sonnet 4.6.
- Top tier for hard tasks (complex agent loops, cross-file refactors, deep reasoning): Claude Opus 4.7 or GPT-5.5 (or GPT-5.5 pro for the very hardest).
This ladder works because cost ratios are roughly 1 : 5-10 : 30+ across the tiers — using top-tier for everything is genuinely wasteful, but reserving it for the hard 5% of tasks gives big quality gains where it matters.
Router One makes this practical: one API key gives you the models above that are in the catalog — check the model catalog for the current line-up — plus the rest of the field. Your application chooses the model for each step, while the gateway selects an eligible provider route and records the model, request ID, tokens, latency, and cost. See AI model routing explained for the routing boundary.
Multimodal
| Model | Image | Long PDF | Video | Audio |
|---|---|---|---|---|
| Claude Sonnet 4.6 | ✅ | ✅ | Limited | ❌ |
| Claude Opus 4.7 | ✅ | ✅ (1M ctx) | Limited | ❌ |
| GPT-5.5 | ✅ | ✅ (1.05M ctx) | ✅ (frames) | ❌ |
| Gemini 3.1 Pro | ✅ | ✅ (1M ctx) | ✅ | ✅ |
| Qwen 3.5 VL | ✅ | ✅ | Limited | ❌ |
| Doubao 2.0 Vision | ✅ | ✅ | Limited | Limited |
| DeepSeek V4 | Limited | Limited | ❌ | ❌ |
For multimodal-heavy workloads the international leaders (Gemini 3.1 Pro, GPT-5.5, Claude) are still ahead. Qwen 3.5 VL and Doubao 2.0 Vision are catching up but lag on edge cases. See the Gemini API China guide for accessing the Gemini models in the Router One catalog from China.
When to Pick a Domestic Model Specifically
A pragmatic checklist:
- ✅ Chinese-language consumer interaction → start with Qwen 3.5 or Doubao 2.0
- ✅ High-volume, cost-sensitive throughput → Doubao 2.0 Pro or DeepSeek V4-Flash
- ✅ Math-heavy / reasoning-heavy with cost ceiling → DeepSeek V4-Pro
- ✅ Latency-sensitive Chinese chatbot → Doubao 2.0 Pro
- ❌ Hard agentic loops touching many files → Claude Opus 4.7 or GPT-5.5 still wins
- ❌ Complex multi-step reasoning over English text → Claude Opus 4.7 or GPT-5.5
- ❌ Strict instruction-following with multimodal input → Gemini 3.1 Pro or GPT-5.5
FAQ
Are domestic models really much cheaper? On the posted token line, yes — they sit a tier or two below the Western frontier. The catch: for a given task they often produce more tokens (longer explanations, more verbose code), which partly closes the gap. Compare on cost per completed task, not on the headline rate, and check the current numbers on the model catalog rather than a table like this one.
Can Qwen 3.5 handle English as well as Chinese? The 397B-A17B flagship is fully bilingual and competitive on English benchmarks (88.4 on GPQA Diamond, the highest tracked). Smaller Qwen variants weaken faster on English than on Chinese.
Is DeepSeek V4-Pro really comparable to Claude Opus 4.7 at coding? On SWE-bench Verified, V4-Pro is at 80.6% vs Opus 4.7's 87.6% — a real gap, ~7 points. On LiveCodeBench V4-Pro actually leads at 93.5%. In production agent loops, Opus 4.7 still has the edge for long-horizon tasks. V4-Pro is excellent for one-shot generation and review at meaningfully lower cost.
What about Qwen Coder / Doubao Coder specifically? Both vendors ship coding-specialized variants. They're competitive with Claude Sonnet 4.6 for code generation in their training distribution but weaker on cross-file repo tasks. Worth trying for greenfield code generation; less compelling for agentic workflows.
How do I access these from outside China?
DeepSeek V4 has open weights (Pro and Flash) and a hosted API accessible globally. Qwen and Doubao have official APIs but signing up from outside China requires Chinese phone verification — workable but extra friction. As of 2026-09-05, DeepSeek V4, Doubao, and Qwen were not in the Router One catalog; on 2026-09-10 deepseek-v4.1-flash was listed with deepseek-v4-flash alongside it, while V4 Pro, Doubao and Qwen remain unlisted. The comparisons in this article are historical; use the providers directly or choose a currently listed model from the live catalog.
Will Western models keep their lead on agentic workflows? Hard to say. The April 2026 frontier shift (Opus 4.7, GPT-5.5, V4-Pro all within 8 days) shows the race is tighter than ever. Anthropic and OpenAI specifically post-train for tool use and long-horizon agents; Chinese labs are catching up fast. Check current benchmarks before betting on a 6-month-old picture.
Can I run DeepSeek V4 locally? V4-Pro (1.6T total / 49B active) needs very serious hardware. V4-Flash (284B) runs on more modest setups in quantized form. For most teams it's cheaper to call the hosted API than to operate the infrastructure.
Conclusion
The May 2026 reality for Chinese developers is that the right choice is "all of the above, routed by the task." Qwen 3.5 and Doubao 2.0 for Chinese-native and cost-sensitive work; DeepSeek V4-Pro for math-heavy and code-heavy at scale; Claude Opus 4.7 for hard agentic loops; GPT-5.5 and Gemini 3.1 Pro for the long tail. A unified gateway can simplify access to the models it currently lists. The historical lineup above is not a statement of current Router One availability; check the live catalog.
For the broader story of model routing across providers see AI model routing explained; for cost-specific levers see 5 ways to reduce LLM API costs.
Availability history
- 2026-08-26: Claude Opus 5 had joined the Router One catalog, Claude Opus 4.7 was listed only as its thinking variant, and Qwen was not in the catalog. The benchmark scores, the Beijing latency numbers and the line-up in this post stay as of May 2026.
- 2026-09-05: DeepSeek V4 and Doubao were no longer listed, and Qwen was not listed either. The vendor comparisons in this post do not imply current gateway availability.
- 2026-09-11:
deepseek-v4.1-flash(DeepSeek V4.1 Flash) was listed on 2026-09-10, withdeepseek-v4-flashlisted again alongside it — both are served on /v1/chat/completions, on /v1/messages and natively on /v1/responses; DeepSeek V4 Pro, Doubao and Qwen remained unlisted.