Quick answer. On the Artificial Analysis Intelligence Index, Kimi K3 scores about 57 — fourth overall, behind Claude Fable 5 (~60) and GPT-5.6 Sol (~59) and narrowly ahead of Claude Opus 4.8 (~56). It leads all of them on the Frontend Code Arena (#1) and posts the strongest open-weight GPQA Diamond result to date (93.5%). It trails the top two closed models on broad agentic tasks.
When Moonshot AI released Kimi K3 on July 16, 2026, the interesting question wasn't whether an open-weight model could be big — at 2.8 trillion parameters it's the largest ever — but whether it could actually compete with the closed frontier. The short version: on coding it wins, on general intelligence it lands a close fourth, and on price-per-task it can undercut both GPT-5.6 Sol and Claude Opus 4.8.
This post breaks down every published benchmark and what it means for real work. For the architecture and specs, see our Kimi K3 complete guide.
How does Kimi K3 score on the Artificial Analysis Intelligence Index?
The Artificial Analysis Intelligence Index (v4.1) is the closest thing to a single cross-model capability score. Here's where K3 lands against the current frontier:
| Model | Intelligence Index | Reported footprint |
|---|---|---|
| Claude Fable 5 | ~60 | ~3T |
| GPT-5.6 Sol | ~59 | ~3T |
| Kimi K3 | ~57 | 2.8T (open-weight) |
| Claude Opus 4.8 | ~56 | ~1.5T |
K3 is fourth of all models on this index — the first open-weight model to place this high. It sits within a few points of GPT-5.6 Sol and Fable 5, and it edges out Opus 4.8. Parameter footprints for the closed models are reported estimates, not official figures.
Where does Kimi K3 lead: coding benchmarks
Coding is K3's standout category. On LMArena's Frontend Code Arena, K3 took the #1 spot — a 17-place jump from Kimi K2.6's #18 — and placed first in six of seven frontend domains, passing Claude Fable 5. On the terminal-agent side it posted 88.3% on Terminal-Bench 2.1.
| Coding benchmark | Kimi K3 |
|---|---|
| Frontend Code Arena | #1 (1st in 6 of 7 domains) |
| Terminal-Bench 2.1 | 88.3% |
| GPQA Diamond | 93.5% (strongest open-weight result at launch) |
That 93.5% GPQA Diamond score was the best open-weight result published at launch, which matters if you want a self-hostable model for reasoning-heavy engineering work rather than an API-only closed system.
Where does Kimi K3 trail: agentic and general tasks
On broad agentic work, the top two closed models still lead. On the GDPval v2 agentic benchmark (Elo), K3 scores well but behind Fable 5:
| Model | GDPval v2 (Elo) |
|---|---|
| Claude Fable 5 | 1,760 |
| Kimi K3 | 1,668 |
| Claude Opus 4.8 | 1,600 |
| GLM-5.2 | 1,514 |
| GPT-5.5 | 1,494 |
K3 sits second here, ahead of Opus 4.8, GLM-5.2, and GPT-5.5 but clearly behind Fable 5. On the AA-Briefcase knowledge-work benchmark, K3 scores 1,547 Elo, with only Claude Fable 5 rated higher — a strong result for a model you can eventually run yourself.
How is Kimi K3 on long context?
K3 keeps its accuracy across the full window. Evaluated with a 1M-token context and no context-management tricks, it scored 90.4 — meaning the 1-million-token window is genuinely usable for repository-scale code and large-document analysis, not just a spec-sheet number.
Kimi K3 pricing vs the frontier
K3 is priced like a Western mid-range model, not a budget Chinese one — but its token efficiency pulls real per-task cost back down:
| Model | Input / M | Output / M | Cost per task* |
|---|---|---|---|
| Kimi K3 | $3.00 ($0.30 cache) | $15.00 | $0.94 |
| GPT-5.6 Sol | — | — | $1.04 |
| Claude Opus 4.8 | — | — | $1.80 |
| Kimi K2.6 (prior gen) | $0.95 ($0.16 cache) | $4.00 | — |
*Per-task cost as measured by Artificial Analysis. K3's per-token price is roughly 3x its predecessor's, but it comes in cheaper per task than both GPT-5.6 Sol and Opus 4.8 because it uses fewer tokens to reach an answer.
Kimi K3 vs GPT-5.6 Sol vs Claude Fable 5 vs Opus 4.8: which should you use?
- Best overall intelligence: Claude Fable 5, then GPT-5.6 Sol. K3 is a close fourth.
- Best for frontend / coding: Kimi K3 — it's #1 on the Frontend Code Arena and posts the top open-weight GPQA Diamond score.
- Best open-weight option: Kimi K3, with no real competition at this capability tier once the weights land.
- Best value per task: Kimi K3, which undercuts both Opus 4.8 and GPT-5.6 Sol on measured cost-per-task.
- Best for the highest-stakes agentic work: Claude Fable 5 still leads on GDPval and AA-Briefcase.
One caveat for factual workloads: independent testing found K3's accuracy rose to ~46% (from K2.6's 33%) but its hallucination rate also climbed to ~51%. Keep verification in the loop.
Part of our AI models series. Read the Kimi K3 complete guide for architecture and access, and the Kimi complete guide for the full Moonshot lineup.
Frequently asked questions
Is Kimi K3 better than GPT-5.6 Sol?
On the Artificial Analysis Intelligence Index, GPT-5.6 Sol (~59) edges out Kimi K3 (~57) on general intelligence. But K3 leads on the Frontend Code Arena and comes in cheaper per task ($0.94 vs $1.04), so for coding and cost-sensitive work K3 can be the better pick.
Is Kimi K3 better than Claude Opus 4.8?
Kimi K3 narrowly beats Opus 4.8 on the Intelligence Index (~57 vs ~56) and on the GDPval agentic benchmark (1,668 vs 1,600 Elo), and it's cheaper per task. Opus 4.8 remains a strong closed-model option, but K3 is competitive and open-weight.
What is the best open-weight model in 2026?
As of its July 2026 launch, Kimi K3 is the strongest open-weight model — the largest ever at 2.8T parameters and the first to place fourth overall on the Artificial Analysis Intelligence Index, ahead of every other open-weight release.
Does Kimi K3 beat Claude Fable 5?
Not on general intelligence — Fable 5 leads the Intelligence Index, GDPval, and AA-Briefcase. But Kimi K3 does beat Fable 5 on the Frontend Code Arena, where it took the #1 spot.
How does Kimi K3 compare to Kimi K2.6?
K3 jumped 17 places on the Frontend Code Arena (from K2.6's #18 to #1) and improved accuracy from 33% to 46%. It's a large capability jump, but priced roughly 3x higher per token than K2.6.
The bottom line
Kimi K3 doesn't top every leaderboard, but it doesn't need to. It wins the coding arena outright, posts the best open-weight reasoning score published at launch, holds a usable 1M-token context, and lands fourth overall on general intelligence — all as an open-weight model you'll be able to self-host. For teams building on open models, that combination didn't exist a week ago.
Shipping on frontier AI and need developers who keep pace with releases like this? Extend your engineering team with vetted remote developers through Codersera.