Kimi vs Claude for Coding: Agentic Open Weights Tested

On paper this is the most lopsided matchup in our arena. Kimi K2.7 Code is Moonshot AI’s open-weight coding specialist: published weights, an agentic training recipe, and a roughly $4-per-million output price. Claude Fable 5 is Anthropic’s Mythos-class flagship: closed, positioned above Opus, and at $10 in / $50 out per million tokens the most expensive model on our stand. One lists at 12.5x the other per output token. The question the votes exist to answer is whether the code is 12.5x apart too.

Two different bets on the same job

What Kimi K2.7 Code brings

Kimi K2.7 Code is Moonshot’s answer to a specific question: how much of the agentic coding loop — write, run, read the error, fix — can you get from a model whose weights you can actually download? It is trained as a coding agent rather than a chat model with a code habit, the weights are published so you can self-host it or rent it from whichever provider is cheapest this month, and it is no longer a curiosity: GitHub added it as a Copilot option, which is about as mainstream as an open-weight coder gets.

What Claude Fable 5 charges for

Fable 5 is the opposite bet: the first Mythos-class model, generally available at $10 per million input tokens and $50 per million output — double Opus 4.8’s output price, more than ten times Kimi’s. What that buys, by our own tally, is real: as of this writing Fable 5 holds the second-most community votes in the arena overall, behind only Opus 4.8 at high effort, and in our agentic Godot pipeline it rebuilt a full marble platformer in 22 self-correcting turns. Long-horizon agentic work is exactly where Anthropic says the money goes, and so far the arena mostly agrees.

How the arena settles it

• Same one-shot prompt. Both models get the identical published instruction per challenge — landing pages, arcade games, emulators, physics sandboxes. One shot, no retries, no cherry-picking.
• Real outputs, running live. The generated apps execute in your browser, bugs included — not screenshots.
• Blind community votes. Compare mode hides the model names until after you vote, and every vote lands in the same public tally.
• Honest cost math. Output tokens and estimated cost per task are disclosed next to every entry, so “slightly better” gets weighed against “twelve times the price”.

What the votes say so far

I will not paste vote counts here, because they would be stale within a week — the tally is live and moves with every vote. The durable pattern so far: the Claude tier leads the overall standings with Fable 5 near the top of them, while Kimi takes individual challenges off the closed flagships regularly — on real prompts the open-weight cluster sits much closer than the price sheet implies. Watch the head-to-head yourself: Kimi K2.7 Code vs Claude Fable 5 runs both against the same landing-page brief, live, and the leaderboard has the current totals, cost per task and generation times.

If your shortlist is bigger than these two — GPT, Gemini, GLM, DeepSeek and the rest of the 20-family roster — our guide to the best AI for coding walks the whole field. Or skip the reading: open the arena, put Kimi next to Fable 5 in compare mode, judge blind, and vote. The ranking on this site is not my opinion; it is the running total of people who did exactly that.