DeepSeek V4 Pro Tops Fireworks Coding Tests

DeepSeek V4 Pro Tops Fireworks Coding Tests
Monday, September 21, 2026
  • •DeepSeek V4 Pro 0813 scored 95.2% on SWE-Bench Verified and 92.0% on LiveCodeBench v6
  • •Fireworks measured $0.309 per solved SWE-Bench task for DeepSeek V4 Pro versus $0.808 for Fable 5
  • •Oracle routing with DeepSeek V4 Pro and Fable 5 hit 92.4% at $0.279 per task

Fireworks said on 8/26/2026 that DeepSeek V4 Pro 0813 led its SWE-Bench Verified and LiveCodeBench v6 runs while costing less per solved task than Fable 5. Fireworks enabled DeepSeek on its platform the previous week with serverless endpoints, dedicated deployments, and training support. The model offers a 1M-token context window, native tool calling, and preserved reasoning history for long-horizon agentic workloads.

Fireworks evaluated DeepSeek V4 Pro 0813 against Kimi K3 and Fable 5 across SWE-Bench Verified, LiveCodeBench v6, Aider Polyglot v1, and Terminal-Bench 2.1, using the same eval suite it used for Kimi K3 in July. DeepSeek V4 Pro scored 95.2% on SWE-Bench Verified (500), ahead of Kimi K3 at 92.6% and Fable 5 at 85.4%. It scored 92.0% on LiveCodeBench v6 (100), ahead of Kimi K3 at 88.0% and Fable 5 at 91.0%. Fable 5 led Aider Polyglot (225) with 88.0%, compared with Kimi K3 at 76.4% and DeepSeek V4 Pro at 73.8%. Terminal-Bench 2.1 (89) put Kimi K3 at 80.9%, while DeepSeek V4 Pro and Fable 5 both scored 76.4%.

DeepSeek V4 Pro delivered lower cost per solved task on standard code tasks in Fireworks' tests. On SWE-Bench, it ran at $0.309 per solved task versus $0.808 for Fable 5. On LiveCodeBench, it ran at $0.040 versus $0.225 for Fable 5. On Terminal-Bench, Fable 5 matched DeepSeek V4 Pro's score but required multiple attempts, raising task costs to $16.85; on a first-attempt baseline, Fable 5 scored 74.2% and cost per task rose to $75.40.

Fireworks said DeepSeek V4 Pro was weaker on multi-language breadth. On Aider Polyglot, the model reached 90.0% in Rust and 80.8% in C++, but fell to 48.9% on Java, the largest slice at 47 tasks, while Fable 5 reached 74.5% on Java. Fireworks said Java performance accounted for nearly the entire Aider Polyglot gap and argued that Java-heavy codebases may need multiple models rather than a single-model setup.

Fireworks tested oracle routing (after-the-fact best model selection) between DeepSeek V4 Pro and Fable 5 by picking the better model per task and breaking ties toward lower cost. On a 409-task combined cohort, routing reached 92.4% versus 86.6% for Fable 5 alone, a 5.9-point gain, at $0.279 per task versus $4.510 for Fable 5. Fireworks said the routed setup was 16 times cheaper and sent 86.3% of tasks to DeepSeek V4 Pro. The company cautioned that oracle routing is an upper bound computed after the fact, not a deployable policy today.

Fireworks said every DeepSeek V4 Pro pairing beat the equivalent Kimi K3 pairing. LiveCodeBench reached 94.0% versus 92.0%, Aider reached 93.3% versus 91.6%, and the combined cohort reached 92.5% versus 91.1%. Fireworks said DeepSeek V4 Pro was a better routing partner than Kimi K3 even though Kimi K3 scored higher on CyberGym, because DeepSeek V4 Pro won on tasks that Fable 5 missed.

Fireworks also said DeepSeek V4 Pro training is live with support for SFT, DPO, and RFT, and that fine-tuned checkpoints can be deployed to production after training. The company said DeepSeek's own benchmark runs use DeepSeek Harness, a native agent runtime (software loop for agent actions) that handles model adapters and execution. Fireworks said a minimal-first setup with bash and editor tools produced more consistent tool calls than starting with the full standard catalog. Across 840 traced adversarial security runs, Fireworks said DeepSeek V4 Pro recorded zero refusals and zero output-length truncations.

Subscribe to the AIB Newsletter

The fastest, most reliable AI news, straight to your inbox

Companies related to this article