Kimi K3 achieves 68.5% on DeepSWE, costs 2.8x less per task than Claude Fable 5

Together AI released a DeepSWE benchmark analysis on July 24 showing Moonshot AI's open-weight Kimi K3 matching Anthropic's Claude Fable 5 on software-engineering coding tasks at a sharp cost discount. On pass@1, Kimi K3 scored 68.5% versus Fable's 69.9% — a 1.4-point gap. But when benchmarks allow multiple attempts (pass@2 and pass@4), Kimi pulls ahead: 82.0% vs 80.2% at pass@2, and 89.4% vs 88.5% at pass@4.

Cost is the headline: Kimi K3 costs $4.65 per task rollout versus $13.41 for Claude Fable 5 at its xhigh setting. Measured per solved task, Kimi delivered 14.7 solves per $100 against Fable's 5.3 — 2.8x the output per dollar. The performance gap reflects a strategic difference: Fable is more reliable on single attempts (higher pass@1, more four-for-four solves); Kimi casts a wider net across retries and achieves better pass@2/4 coverage. Both models fail on nearly identical task types (0.72 correlation).

Kimi K3 is open-weight, deployed through Together AI's inference platform, while Fable 5 is closed and available only through Anthropic. This makes Kimi suitable for teams building agentic workflows where multiple attempts and self-hosting are acceptable — voice agents, code generation loops, search-and-iterate tasks. Fable remains preferred for single-shot production paths requiring determinism.

For architects choosing between open and closed LLMs for reasoning and coding tasks, this signals a milestone: open-weight models now achieve flagship-tier reasoning capability at production cost/token ratios closer to small models. DeepSWE is a narrow benchmark, but it measures real software engineering. The cost advantage ($4.65 vs $13.41) may drive adoption among teams building CI/CD agents and autonomous programming systems, especially where task retry is tolerable.