Kimi K3 vs Claude Fable 5: two leaderboards, two different winners (and one pricing myth to kill)

Moonshot AI shipped Kimi K3 on 16 July 2026, and within a day it was sitting at #1 on Arena.ai's blind Frontend Code Arena — above Anthropic's flagship Claude Fable 5. Meanwhile, on the general-purpose Artificial Analysis Intelligence Index, Fable 5 still leads and K3 sits third, behind GPT-5.6 Sol. Both facts are true at the same time, and the gap between them is the whole story. We also ran K3 through our own executed coding benchmark on launch day, which turned up the number nobody else is printing: K3 is not a cheap model.

The split verdict

Most coverage of Kimi K3 picks one leaderboard and runs with it. The Chinese-AI-wins takes cite Arena.ai. The nothing-has-changed takes cite Artificial Analysis. Both are reading real data, and both are telling you half the story.

Here is the whole thing, as of July 2026:

• Arena.ai Frontend Code Arena (independent, blind pairwise human vote): Kimi K3 is #1 at 1,679 points, ahead of Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618. That is off 1,757 valid votes, and it is a 17-place jump from K2.6, which sat at #18. K3 took first in 6 of 7 frontend domains; the one it lost was Gaming, to Fable 5.
• Artificial Analysis Intelligence Index (independent, composite benchmark): Claude Fable 5 leads at 59.9. GPT-5.6 Sol is at 58.9. Kimi K3 is third at 57.1, ahead of Claude Opus 4.8 at 55.7.

Neither result cancels the other, because they measure different things. The Arena number is human preference for generated interfaces — taste, polish, does-this-look-right. The Index is a composite of correctness-style evaluations. A model can genuinely produce the frontend people prefer while sitting behind on general reasoning. K3 is that model.

Two caveats worth carrying. First, the top of the Intelligence Index is tightly clustered: the top four span 59.9 down to 55.7, just over four points. Second, even a single source can give you two different rank numbers — Artificial Analysis's own launch article puts K3 at #3, while its live K3 model page currently shows #4 of 187 against a ranked leaderboard of 165 models. Quote the score, name the source and date it, because this ordering will move again.

What each model actually is

Kimi K3 (Moonshot AI, released 16 July 2026) is a 2.8-trillion-parameter mixture-of-experts model, the largest open-weight model released to date. The architecture is the interesting part: Kimi Delta Attention, a hybrid linear attention scheme that uses recurrent state instead of a conventional full KV cache everywhere, which Moonshot says delivers up to 6.3x faster decoding at million-token context. Moonshot's earlier Kimi Linear paper reported up to 75% lower KV-cache use on the same family of techniques. There is also a second mechanism, Attention Residuals. Around 16 of 896 experts are active per token. Context runs to 1M tokens, and vision is native. Full breakdown in our Kimi K3 review.

The open part carries an asterisk. Weights are promised for 27 July 2026 under a Modified MIT license — promised, not shipped. If you are reading this before that date, K3 is an API-only model in practice, and any architecture that assumes self-hosting is a bet, not a plan.

Claude Fable 5 is Anthropic's current top-tier flagship: closed weights, no self-host path, and priced accordingly. It is the model to beat on general intelligence, and on the Index it is still the model to beat. Our fuller writeup is in the Claude Fable 5 breakdown, and if the price is the sticking point, Fable 5 alternatives covers the field.

On Moonshot's own numbers: the vendor-reported benchmarks show K3 trailing Fable 5 and GPT-5.6 Sol on overall performance while outperforming Opus 4.8 and GPT-5.5 across the rest of Moonshot's evaluation suite. We are labelling those vendor-reported and leaving them there — we are not going to invent specific SWE-bench figures for either model, and neither should anyone else.

Spec, price and benchmark table

Every number below carries its provenance, because that is the difference between a comparison and a vibe. V = vendor-reported, I = independent third party, FP = DataLLM Lab first-party executed run.

Three rows deserve a second look. The blended cost figures are a model, not a list price — Artificial Analysis assumes a 7:2:1 mix of cache-hit input, fresh input and output. Change that ratio and the gap changes with it. The output-speed row is nearly a tie, but the time-to-first-token row is not: roughly two seconds against roughly two minutes, which is the difference that actually shows up in an interactive product. And Fable 5 is absent from the first-party rows on purpose: it was not in our 13-model executed run, so attributing an executed score to it would be fabrication.

Two axes, two winners

The cleanest way to see the split is to stop ranking and start plotting. Horizontal axis: general intelligence. Vertical axis: blind human preference on frontend code. Fable 5 wins going right; K3 wins going up.

Claude Opus 4.8 sits at 55.7 on the index and is not plotted here, because we do not have a Frontend Arena Elo for it from the same source. Same discipline as the table: no axis, no dot.

Our launch-day run: the cost nobody prints

Leaderboards score answers. Our benchmark runs them. In our executed coding benchmark each model writes code that is then executed against hidden tests, and cost is computed from real token usage times list price, normalized to 1,000 tasks. Thirteen models, nine tasks.

The headline from that run is that correctness is table stakes: 10 of 13 models scored a perfect 9/9, and the spread that actually mattered was cost — 88x between the cheapest and the priciest. GPT-5.5 hit 9/9 at $8.83 per 1,000 tasks. Claude Opus 4.8 hit 9/9 at $4.05. GLM 5.2 did it for $1.99, Kimi K2.7-Code for $1.34, and DeepSeek V4-Flash for $0.13. Qwen3 Coder Next came in cheapest at $0.10. Grok 4.3 was one of the misses at 8/9, failing a nested-dict flatten.

We added Kimi K3 on its launch day, 16–17 July 2026. Two things came out of it.

One: K3 passed all 8 tasks we could execute; the 9th could not be evaluated. That ninth task, parse_csv_line, never completed, because the upstream returned 429 rate-limit responses repeatedly. That is a launch-day capacity signal, not a correctness miss, and we are not going to round it up. It is not a 9/9. It is eight for eight on what we could run, with one task unevaluated.

Two, and this is the finding: K3 cost about $4 per 1,000 tasks — one run came in at $4.37. That is Opus-4.8 territory, sitting almost exactly on Opus at $4.05, and roughly 3x GLM 5.2, 30x DeepSeek V4-Flash, 40x Qwen3 Coder Next. Worth noting that K3's per-token list price is lower than Opus 4.8's $5/$25, so landing on the same cost per 1,000 tasks means K3 is spending more tokens to get there — which is exactly what Artificial Analysis found from a different direction, flagging K3 as very verbose at 130M output tokens across the Intelligence Index against a 63M median.

If your mental model of K3 is the cheap open Chinese model that beat Anthropic, that $4 figure is the correction. It executed like a frontier model and it billed like one.

The pricing myth, corrected

K3 lists at $3.00 per 1M input and $15.00 per 1M output, with cache-hit input around $0.30 — a 90% discount. Two things follow that most coverage skips.

First, that is not a cheap Chinese price. It is Anthropic's standard Sonnet 5 rate, matched exactly — and Sonnet 5 is currently running below that, on introductory pricing of $2/$10 through 31 August 2026, which makes Anthropic's mid-tier model cheaper than K3 today. K2.6 listed at $0.95 and $4. K3 more than tripled input and nearly quadrupled output. Coverage framed the launch, correctly, as the end of super-cheap Chinese AI. The era where the open Chinese model was a rounding error on your invoice ended on 16 July.

Second, the gap to Fable 5 is about 3.3x, not 10x. Fable 5 lists at $10 input and $50 output — with a 90% prompt-cache discount of its own, US-only inference at a 1.1x multiplier, and a Batch API that halves the bill. Run Artificial Analysis's blended model at 7:2:1 cache-to-input-to-output and you get roughly $2.31 per 1M for K3 against roughly $7.70 for Fable 5. Real, meaningful, and nowhere near the order-of-magnitude readers assume.

The corollary is that cache discipline moves the needle more than model choice here. Both models discount cached input by 90%. A workload with a heavy stable prefix collapses most of the difference; a workload that regenerates its context every call pays full freight on both.

Two further wrinkles from independent testing. Artificial Analysis's own summary is that K3 is among the leading models in intelligence but somewhat expensive compared with other models of similar price, and also slower than average and very verbose — which is exactly what our $4 per 1,000 tasks showed from a different direction. And Simon Willison's SVG test found K3 exposes only a single max reasoning effort, burning 13,241 reasoning tokens to emit a 3,417-token response, while producing strong vision and alt-text quality. No effort dial means no cheap mode.

A decision rule you can actually use

Skip the rankings. Answer four questions.

1. Is the output a user interface that a human will judge on sight? Lean K3. That is precisely what the Frontend Arena measures, and K3 won 6 of 7 domains in a blind vote. Gaming is the exception — Fable 5 took that one.

2. Is the hard part the reasoning, not the rendering? Lean Fable 5. It leads the Intelligence Index at 59.9, and the Arena result does not transfer to multi-step logic. Winning a preference vote on frontend code is not the same as being the best coder, and treating it as such is how people get burned.

3. Does self-hosting change your economics or your compliance posture? Then K3 is the only one of the two with a path — conditional on the 27 July weight drop actually landing under the promised Modified MIT license. Verify before you plan around it. Broader field in the best open-source LLMs of 2026.

4. Is this shipping to production this month? Weight the 429s. Our launch-day run could not complete one task because of repeated upstream rate-limiting. That will settle as capacity comes online, but a day-one model on a critical path needs a fallback route configured, not assumed.

And the question that is not on the list: which is cheaper. At $3/$15 versus $10/$50 with identical cache discounting, and $4 per 1,000 executed tasks in our own run, K3 is a value pick relative to a flagship — not a budget model. If cost is the actual constraint, neither of these is your answer; GLM 5.2 at $1.99 and DeepSeek V4-Flash at $0.13 per 1,000 tasks both hit 9/9 in the same run. For how K3 stacks up against the other frontier release of this cycle, see Kimi K3 vs GPT-5.6.

Everything here is as of July 2026, and the top of both leaderboards is a moving, tightly-clustered target. Re-check before you commit a quarter of engineering to either answer.

FAQ