Fireworks AI says DeepSeek V4 Pro 0813 scored higher than Fable 5 on SWE-Bench Verified and LiveCodeBench v6 in the company’s coding tests. Fireworks reported costs of $0.309 per solved SWE-Bench task for DeepSeek, versus $0.808 for Fable 5, and $0.040 versus $0.225 per solved LiveCodeBench task.
Fireworks compared DeepSeek with Kimi K3 and Fable 5 across four benchmark families, using the same evaluation suite it had used for Kimi K3. The company says DeepSeek scored 95.2% on SWE-Bench Verified (500 tasks) and 92.0% on LiveCodeBench v6 (100 tasks). Fable 5 scored 85.4% and 91.0%, respectively.
Results vary by language
DeepSeek did not lead every test. On Aider Polyglot v1, Fable 5 scored 88.0% to DeepSeek’s 73.8%; on Terminal-Bench 2.1, Kimi K3 led with 80.9%, while DeepSeek and Fable 5 each scored 76.4%. Fireworks also reported that DeepSeek scored 48.9% on Java tasks in Aider Polyglot, compared with Fable 5’s 74.5%.
The benchmark table is labeled as results “with Oracle Routers.” Fireworks defines oracle routing as testing each task with multiple models and choosing the best result afterward. It is a theoretical upper bound, not a routing system customers can deploy as-is. On a combined 409-task cohort, Fireworks reports that routing between DeepSeek and Fable 5 reached 92.4% accuracy, versus 86.6% for Fable alone, at $0.279 versus $4.510 per task.
What Fireworks offers
Fireworks says DeepSeek V4 Pro is available through its serverless API and dedicated deployments, and that customers can train custom versions using supervised fine-tuning, preference tuning or reinforcement fine-tuning. The source also describes a 1-million-token context window and tool calling. It does not provide independent validation of the benchmark results.
Fireworks notes that its testing setup affected performance: the company found that starting with a minimal set of tools produced more consistent tool calls than beginning with a larger standard catalog. That is a practical caveat for developers trying to reproduce the reported scores.
Why it matters
For teams paying to run coding agents, Fireworks’ reported costs suggest DeepSeek V4 Pro could be a cheaper option for tasks like SWE-Bench and LiveCodeBench. But the results come from Fireworks’ own tests, performance varied by benchmark and language, and the strongest routing result relies on an after-the-fact method that customers cannot directly deploy.
