The timeline treated Kimi K3 like a Claude Fable 5 funeral. A 2.8T open model, #1 on a coding arena, near-frontier intelligence scores. Cue the "China ends the frontier" posts.
So: is K3 actually a Fable 5 killer?
No. Not on overall capability. Moonshot itself says K3 still trails Claude Fable 5 and GPT-5.6 Sol. Independent boards put it close to Opus-class, not above Fable. On a few coding boards it punches up. On price, speed, and verbosity it is not free power.
What follows is the receipts: vendor vs independent scores, and where the hype overshoots the data. Bookmarkable live boards sit at the end.
What Kimi K3 actually is
- 2.8T total parameters, open 3T-class positioning
- Stable LatentMoE: 896 experts, 16 active per token
- Kimi Delta Attention (KDA) and Attention Residuals for long sequences and deep stacks
- Native multimodal text + image input (video and screenshot workflows in product demos)
- 1,048,576 token context, flat API pricing across the full window
- Roughly 2.5x scaling efficiency vs Kimi K2, per Moonshot
- Training/serving notes include MXFP4 weights with MXFP8 activations and a recommended 64+ accelerator supernode for self-hosting
API pricing at launch: $0.30 / 1M cache-hit input, $3.00 / 1M cache-miss input, $15.00 / 1M output. Moonshot says official coding workloads often clear 90%+ cache hits on its own stack.
That is expensive for a Chinese lab release, and closer to mid-high Western API tiers than to older Kimi pricing. The bet is that long-horizon coding and knowledge work pay for the token bill.
Independent benchmarks (not vendor slides)
Artificial Analysis Intelligence Index
Artificial Analysis is the cleanest public third-party snapshot after launch.
| Metric | Reported value (AA) | Notes |
|---|---|---|
| Intelligence Index v4.1 | 57 | About #4 of 186 models in AA's class ranking |
| Output speed | ~36.9 tokens/s | Notably slow vs median peers |
| Input / output price | $3 / $15 per 1M tokens | Cache hit input $0.30 |
| Evaluation verbosity | ~130M output tokens on the index run | Far above a ~63M peer median |
| Context window | 1M tokens | Multimodal text + image input |
| Total eval cost | about $2.7k for the full index suite | High token use drives cost |
Index v4.1 blends nine evals, including GDPval-AA v2, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR. Full methodology lives on Artificial Analysis.
How to read the chart: open the Kimi K3 Intelligence section and the cost-per-task scatter on the same page. Capability is frontier-adjacent. Efficiency is not free.
Reporting around launch also places K3 roughly level with Claude Opus 4.8 and GPT-5.5 on that index, and behind Claude Fable 5 and GPT-5.6 Sol. That matches Moonshot's own "still trails Fable and Sol" line.
Vals Index
Vals lists Kimi K3 at 74.70% ± 0.96 on the Vals Index (release dated July 16, 2026), with a 1M context window. Vals is a separate enterprise-oriented suite, so do not mix the 74.7 figure with AA's 57 as if they share a scale.
LMArena Frontend Code Arena
After unblinding, K3 debuted at #1 on LMArena's Frontend Code Arena at 1679 Elo, ahead of Claude Fable 5 on that board, with a large jump from Kimi K2.6. Arena posts and recaps are summarized in coverage such as BenchLM's K3 note and Simon Willison's write-up.
One preference board is not the whole frontier. It is still a strong public signal for frontend and UI coding, which lines up with Moonshot's "vision in the loop" demos.
Moonshot coding scores (vendor-reported)
These numbers come from the official full benchmark table. Most K3 rows were run with Moonshot's Kimi Code harness at max reasoning effort. Other models often use their own best harness (Claude Code, Codex, Terminus). Treat this as product-level evidence, not a pure model bake-off.
Official launch charts (self-reported suite):
Source: Moonshot AI, Kimi K3: Open Frontier Intelligence
Source: Moonshot AI, Kimi K3 launch post
| Benchmark | Kimi K3 | What it measures | Public board / notes |
|---|---|---|---|
| Terminal-Bench 2.1 | 88.3 | Agentic terminal tasks | Cross-check AA Terminal-Bench |
| FrontierSWE | 81.2 dominance | Hard SWE / research tasks | frontierswe.com (K3 may lag public listing) |
| ProgramBench | 77.8 raw pass rate | Rebuild programs from binary + docs | programbench.com / Vals ProgramBench |
| DeepSWE | 67.5 (Kimi Code); 67.3 mini-SWE-agent | Fresh multi-repo issue fixing | deepswe.datacurve.ai |
| SWE Marathon | 42.0 | Multi-hour project work | swe-marathon.org |
Caveats that matter if you are buying this for a team:
- Raw pass rate ≠ fully resolved. ProgramBench's 77.8 is test-level pass rate, not "77.8% of programs fully rebuilt."
- Harness is part of the score. Moonshot warns K3 is sensitive to preserved thinking history. Drop that history and quality can wobble.
- Dominance is pairwise win probability, not percent of tasks completed.
- Several K3 rows were vendor-run against public harnesses before every owner board refreshed. Prefer independent pages once they update.
Cost, speed, and token hunger
Independent measurements and early field reports converge:
- Strong intelligence score, slow decode (~37 t/s class on AA)
- High verbosity on hard evals (more output tokens than median peers)
- Cost per Intelligence Index task competitive with GPT-5.6 Sol in some recaps, cheaper than heavier Claude max tiers, still not "budget open model" pricing
- Cache hits matter a lot: $0.30 vs $3.00 input is a 10x swing on repeated repo context
If you run short chat prompts, K3 can feel expensive and over-eager. If you run multi-hour coding agents with stable prefixes and high cache reuse, the bill looks different.
Native vision and long-horizon work
K3 is multimodal by design. Moonshot's case studies push kernel optimization, compiler work, game/UI loops with live screenshots, chip design sandboxes, and research-to-code pipelines. Those are demos, not leaderboards. Still, they explain the product pitch: long sessions, tools, vision feedback, and less babysitting.
For product surface area, Moonshot lists kimi.com, Kimi Work, Kimi Code, and the Kimi API (kimi-k3).
Who should care right now
Worth a pilot if you:
- Need long-horizon coding agents on large repos
- Want near-frontier quality with a path to self-host weights after July 27
- Care about frontend / UI work where Arena already ranked K3 first
- Can measure cost per accepted PR, not cost per million tokens in isolation
Probably not your default if you:
- Need low latency chat or high QPS cheap completions
- Cannot tolerate proactive agents that improvise on ambiguous tasks
- Need fully independent, same-harness coding boards for procurement this week
The honest summary
Kimi K3 did not crown a new global #1 model. It did something more interesting for builders: it compressed the open vs closed gap into months, put open 3T-class weights on a calendar, and showed competitive coding agent results while independent labs still place it just behind Fable 5 and GPT-5.6 Sol.
Watch three concrete markers next:
- July 27 weight release (license, checkpoints, serving stack)
- Independent boards posting K3 with traces, costs, and confidence intervals
- Your own bake-off: fixed harness, fixed budget, real repos
Live charts and boards
Bookmark these when you want the latest numbers, not a screenshot frozen in this post. Key charts are also linked inline above next to the scores they support.
| Chart / board | What it shows | Link |
|---|---|---|
| Artificial Analysis model page | Intelligence Index, speed, price, cost per task, token use | artificialanalysis.ai/models/kimi-k3 |
| Artificial Analysis homepage | Cross-model quality, price, and output speed charts | artificialanalysis.ai |
| Terminal-Bench 2.1 (AA) | Independent agentic terminal coding results | artificialanalysis.ai/evaluations/terminalbench-v2-1 |
| Vals model page | Vals Index and suite of enterprise / coding evals | vals.ai/models/kimi_kimi-k3 |
| DeepSWE leaderboard | Fresh repository repair tasks | deepswe.datacurve.ai |
| FrontierSWE | Hard implementation / performance / research tasks | frontierswe.com |
| ProgramBench | Rebuild programs from behavior + docs | programbench.com |
| SWE Marathon | Multi-hour whole-project engineering | swe-marathon.org |
| Moonshot launch post | Official architecture write-up and full score table | kimi.com/blog/kimi-k3 |
| OpenRouter route | Live API pricing and provider routing | openrouter.ai/moonshotai/kimi-k3 |
Sources and credits
Primary and independent sources used for numbers and charts in this post:
- Kimi K3: Open Frontier Intelligence : Moonshot AI launch post, architecture, pricing, full benchmark table, official charts
- Kimi K3 on Artificial Analysis : Intelligence Index, speed, price, verbosity, cost charts
- Artificial Analysis Intelligence Index methodology
- Terminal-Bench 2.1 on Artificial Analysis
- Kimi K3 on Vals : Vals Index and related evals
- DeepSWE leaderboard
- ProgramBench
- FrontierSWE
- SWE Marathon
- OpenRouter: moonshotai/kimi-k3
- Simon Willison on Kimi K3 : launch context, AA highlights, practical API notes
- BenchLM: Kimi K3 release note : Arena Elo, independent index framing
Scores change as boards refresh. Prefer the live chart links in the section above when you need current rankings.