Fireworks reports Kimi K3 within a few points of Fable at a fraction of the cost

Sources: Fireworks blog, HN discussion
Summary: Inference provider Fireworks published a comparison of the open-weight Kimi K3 against Anthropic's Fable across about 1,030 tasks in five categories (repository bug-fixes, long agentic terminal operations, algorithmic problems, multi-language implementation, and legal tasks). It reports K3 at 92.4 percent versus Fable at 92.6 percent on the software-engineering set, with the two staying within a few points across categories, and states K3 was up to roughly 50 times more cost effective than Fable alone on long agentic loops. An oracle router selecting between the two reached 93 percent. Fireworks serves K3 on its own platform.
Comments: HN commenters flagged the post's promotional framing (close results described as a tie when Fable leads but a win when K3 leads) and noted it omits GPT-5.6 Sol from the comparison.
Why it matters: A serving provider's own head-to-head is not independent, but a near-parity, far-cheaper open-weight coding model continues the pressure that closed US labs faced all week.
Follow-up: Watch for the K3 weight release due 2026-07-27, the technical report, and independent agent-harness reproduction against both Fable and GPT-5.6 Sol.