Kimi K3 gets 2.8x more DeepSWE solves per dollar than Fable 5

The coding benchmark turns into a cost benchmark

For teams running coding agents at scale, the useful metric is not always the top single-shot score. Together AI posted on X on August 3, 2026 that Kimi K3 Max delivered “2.8× more solved tasks per dollar” in its DeepSWE analysis. The tweet says Kimi K3 Max came close to Fable 5 xhigh on Pass@1 while costing about one-third as much per rollout.

Together’s linked analysis describes DeepSWE as a software engineering benchmark built from 113 real, long-horizon feature requests from live open-source repositories, with four trials each and hidden pass/fail tests. On the headline metric, Fable 5 xhigh leads Kimi K3 max by 69.9% to 68.5% pass@1, a 1.4-point gap. But the ranking changes when multiple attempts are allowed: Kimi reaches 82.0% pass@2 versus Fable’s 80.2%, and 89.4% pass@4 versus 88.5%.

The cost profile is the sharper signal. Together reports $4.65 per Kimi K3 rollout and $13.41 for Fable 5 xhigh. Across the 452-rollout sweep, that means $2,103 for Kimi and $6,010 for Fable. Per solved task, Kimi returns 14.7 solves per $100, while Fable returns 5.3. In workflows where tests can reject bad patches and trigger another attempt, that can matter more than a small pass@1 lead.

The caveat is reliability. Fable 5 remains steadier on first attempts and solves more tasks four-for-four, while Kimi casts a wider net across retries. Watch whether Kimi’s upcoming open-weight availability and inference optimization reduce latency, and whether the DeepSWE pattern holds inside private repositories with real CI and review constraints. The source tweet is available on X.