Kimi K3 vs Claude Fable 5 on DeepSWE: Cost and Coding

A model that solves more tasks after several attempts can still be less consistent on repeat runs. Together AI illustrates this in its DeepSWE comparison, which tested 113 software tasks four times each. Moonshot AI's Kimi K3 passed at least once on 89.4% of tasks, slightly above 88.5% for Anthropic's Claude Fable 5.

Fable nevertheless passed all four trials on 58 tasks, compared with Kimi's 45. That separates coverage, reaching a passing solution at least once, from repeatability. The vendor's results suggest different strengths depending on whether a workflow values multiple chances or consistent success. The benchmark's hidden tests identify successful attempts; a production workflow would also need a way to recognize which output is usable.

“Kimi K3 landed in DeepSWE on July 16, 2026, with 452 graded rollouts at max effort: 113 real, long-horizon feature requests from live open-source repos, four trials each, graded pass/fail by a hidden test suite.”

Together AI

“Kimi reaches 89.4% of the benchmark - higher than any peak, with only 12 tasks it never cracks (Fable: 13). But it's less reliable on 4/4 tries: 76.6% reliability and only 45 tasks solved four-for-four, against Fable’s 79.0% and 58.”

Together AI

“Per solved task, Kimi delivers 14.7 solves per \$100 versus Fable’s 5.3 - 2.8x the work per dollar.”

Together AI

“Per-task correlation between Kimi K3 and Fable is 0.72 - the highest cross-vendor similarity in the entire benchmark. In fact the top four cross-vendor similarities in the export are all Kimi-K3-versus-Anthropic pairs.”

Together AI

“In practice, Kimi K3 and Claude Fable 5 succeed and fail on nearly the same tasks, so pairing them buys you almost no diversity: their union covers 105 of 113 tasks, barely above Kimi alone at 101.”

Together AI