Provider pricing (live)
Pulled from the OneInfer pricing catalog; rates can change. Capture the timestamp with any benchmark or production decision.
| Provider | Input $ / 1M tokens | Output $ / 1M tokens |
|---|---|---|
| novita | $0.150 | $0.470 |
| deepinfra | $0.150 | $0.470 |
| oneinfer | $0.150 | $0.470 |
Vendor-reported benchmarks
Scores come from the Qwen/Qwen3.8-27B benchmark record in the OneInfer model catalog and have not been independently reproduced.
| Evaluation | Score |
|---|---|
| General agentic | |
| Claw-Eval Avg | 72.4 |
| Claw-Eval Pass^3 | 60.6 |
| QwenClawBench | 53.4 |
| SkillsBench | 48.2 |
| Agentic coding | |
| SWE-Bench Verified | 77.2 |
| SWE-Bench Pro | 53.5 |
| SWE-Bench Multilingual | 71.3 |
| TerminalBench 2.0 | 59.3 |
| NL2Repo | 36.2 |
| QwenWebBench | 1487.0 |
| Multimodal | |
| MMMU | 82.9 |
| MMMU-Pro | 75.8 |
| MathVista mini | 87.4 |
| DynaMath | 85.6 |
| VlmsAreBlind | 97.0 |
| RealWorldQA | 84.1 |
| MMStar | 81.4 |
| MMBench EN-DEV v1.1 | 92.3 |
| SimpleVQA | 56.1 |
| General capabilities and reasoning | |
| MMLU-Pro | 86.2 |
| MMLU-Redux | 93.5 |
| SuperGPQA | 66.0 |
| C-Eval | 91.4 |
| GPQA Diamond | 87.8 |
| Humanity's Last Exam | 24.0 |
| LiveCodeBench v6 | 83.9 |
| AIME 2026 | 94.1 |
| HMMT Feb 2026 | 84.3 |
| HMMT Nov 2025 | 90.7 |
| IMOAnswerBench | 80.8 |
| Document understanding | |
| CharXiv RQ | 78.4 |
| CC-OCR | 81.2 |
| OCRBench | 89.4 |
| Spatial intelligence | |
| ERQA | 62.5 |
| CountBench | 97.8 |
| RefCOCO Avg | 92.5 |
| EmbSpatialBench | 84.6 |
| RefSpatialBench | 70.0 |
| Video understanding | |
| VideoMME | 87.7 |
| VideoMMMU | 84.4 |
| MLVU | 86.6 |
| MVBench | 75.5 |
| Visual agent | |
| V* | 94.7 |
| AndroidWorld | 70.3 |
Decision snapshot
Both models are flash-tier with 1M+ context windows, released in the same fortnight (Qwen3.8-Flash on 26 Aug 2026, DeepSeek V4 Flash 0731 on 31 Jul 2026). The decision is between price and verified multimodal/agentic capability.
| Decision factor | Qwen3.8-Flash | DeepSeek V4 Flash 0731 |
|---|---|---|
| Input $ / 1M tokens | $0.16 | $0.030 (5.3× cheaper) |
| Output $ / 1M tokens | $0.47 | $0.075 (6.3× cheaper) |
| Context window | 1,000,000 tokens | 1,310,720 tokens |
| Multimodal (image + text input) | Image + text | Not stated |
| Independent agentic benchmarks | Verified via MarkTechPost / The Decoder | Pending independent reproduction |
| Tool-use track record on Claude Code / coding agents | Yes — flash-tier tuned | Yes — same tier |
| License | Hosted API | Hosted API (0731) |
| Best when | Multimodal flash at $0.16 input | Cheapest flash tier with 1.3M context |
Ready to test the workflow?
Pricing anchor (3:1 input:output blend)
With three input tokens for every output token (a 3:1 input:output ratio), Qwen3.8-Flash blends to $0.238 per 1M tokens and DeepSeek V4 Flash 0731 blends to $0.041 per 1M tokens. Change the ratio in any pricing sheet to test your workload.
When neither is the right answer
For the hardest desktop automation tasks (OSWorld 2.0 binary ≈ 19.4) or scripted RPA, both models trail Claude Opus 4.6. Pick a frontier model instead.