DeepSeek V4 Pro beats Fable 5 on two key code benchmarks and costs a third per solved task

The DeepSeek V4 Pro model version 0813, now running on Fireworks, outperforms Fable 5 on two of four core software-development benchmarks—SWE-Bench Verified and LiveCodeBench v6—while costing roughly a third per solved task. Unlike certain closed models, it does not refuse legitimate security-review requests, making it a more practical partner for teams that need automated code scans without artificial refusals.

Fireworks provides the model via serverless endpoints, dedicated deployments and fine-tuning support, including SFT, DPO and RFT. The context window spans one million tokens, with native tool calling and preserved reasoning history, targeting long-horizon agentic workloads that require the agent to retain context over many steps.

The evaluation suite comprised four families: SWE-Bench Verified, LiveCodeBench v6, Aider Polyglot v1 and Terminal-Bench 2.1. V4 Pro leads as a single model on two of them. On SWE-Bench the cost per solved task is $0.309 versus $0.808 for Fable 5. On LiveCodeBench the gap widens to $0.040 (about 15 agorot) versus $0.225 (about 82 agorot). Terminal-Bench shows two notable cost and performance gaps for Fable 5, but the multilingual results are more striking: in Aider Polyglot the model scores 90.0% on Rust and 80.8% on C++, yet drops to 48.9% on Java, the largest suite with 47 tasks, compared with 74.5% for Fable 5. The Java performance accounts for almost the entire overall gap.

The team also tested Oracle routing, a method that runs each task on both models and retrospectively selects the cheaper correct answer. Across 409 combined tasks, Oracle routing achieves 92.4% accuracy versus 86.6% for Fable 5 alone, a 5.9-point improvement, at $0.279 per task versus $4.510. That is 16 times cheaper, with 86.3% of tasks routed to V4 Pro. The conclusion is that the more expensive model is the exception rather than the default. It is worth noting that Oracle routing is a theoretical upper bound computed retrospectively; real routers achieve only a fraction of that, but the gap size indicates substantial room for improvement.

Any pairing of V4 Pro with Fable 5 outperforms the comparable Kimi K3 pairing: LiveCodeBench reaches 94.0% versus 92.0%, Aider 93.3% versus 91.6%, and the combined cohort 92.5% versus 91.1%. This holds even though K3 scores higher on CyberGym, because V4 Pro’s wins fall precisely on the tasks where Fable 5 fails, making it a more natural complement.

Users who wish to build a domain-specific model can fine-tune V4 Pro directly in Fireworks on a frontier-grade infrastructure without artificial platform limits. After training, the customized checkpoint can be deployed to production with a single click. Because training and serving run on the same co-optimized backend, production performance matches the training benchmarks each time. DeepSeek’s official benchmarks run on DeepSeek Harness (dsh), an original agentic runtime that handles the entire chain from model adapter to execution core.