On September 14, 2026, AI inference platform Fireworks AI released a review of DeepSeek V4.1 Flash. In its DeepSWE code tasks, the model scored a maximum of 74.34%, with an average cost of $0.430 per task; GPT-6 Astra scored 74.12% in the xhigh inference range, costing $6.524. Based on these results, the former's single-task cost is about one-fifteenth that of the latter, but this is only a comparison based on specific platforms and benchmarks, not that all development scenarios maintain the same gap.
How close are the results of the four models?
| Model | DeepSWE scores | Cost per task |
|---|---|---|
| DeepSeek V4.1 Flash | 74.34% | $0.430 |
| GPT-6 Astra xhigh | 74.12% | $6.524 |
| Gemini 3.8 Flash | 73.83% | $2,362 |
| Claude Opus 5 | 73.65% | $11,838 |
Fireworks also emphasizes that DeepSWE fluctuates by about 1.4 to 3.2 percentage points per run, while the highest scores of the four models differ by only 0.69 percentage points. Therefore, this table is not enough to declare who is absolutely superior in code capability; It better illustrates that when quality is in a similar range, inference cost becomes a key variable in routing and procurement decisions.
Where does low cost come from?
DeepSeek V4.1 Flash uses a hybrid expert architecture with 552 billion parameters, activating about 8 billion parameters per token during input and about 16 billion parameters at output. DeepSWE is another typical input-intensive task: each task handles about 36.9 million input tokens and 211,000 output tokens on average, a ratio of about 174:1, with 99.6% of input hits in the cache. Fireworks' billing breakdown shows cache input accounts for about 59.9% of total cost and output about 32.5%. This means model architecture, cache pricing, and workload form together create results, which cannot be explained by total parameters alone.
Cross-benchmark viewing, the conclusions are not the same
In Terminal Bench 2.1, DeepSeek V4.1 Flash scored 86.5%, Astra 87.5%, and Fireworks reported a total cost difference of about 12 times; But in Humanity's Last Exam, the two scores were 34.52% and 50.40%, respectively, showing a clear gap in capability. The report also provides a theoretical upper limit of 54.80% for "Oracle Routing," assuming the ideal result of always selecting the correct model, which may not be possible on real routers.
How should companies use this data?
A safer approach is to conduct small-scale retesting using your codebase, context length, cache hit rate, and failure retry rules, while also recording completion rates, manual rework time, delays, and final billing. Entering highly repetitive code review or warehouse understanding tasks may make it easier to earn caching bonuses; Tasks requiring difficult inference or low cache hits may have completely different cost curves. The value of this evaluation is not to select a unique model for the enterprise, but to show that multi-model routing should consider both quality intervals and true total costs.
