Last updated: July 17, 2026
Key Takeaways
- Kimi K3 debuted at number one on Arena.ai's Frontend Code Arena with 1,679 points, ahead of Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618, but it trails both models on broader intelligence rankings.
- The headline scores hide messy details: benchmarks were run through different software harnesses, several gaps are smaller than the margin of error, and K3's hallucination rate went up compared to its predecessor.
- K3's real advantages are price (roughly one third the output cost of Fable 5) and open weights arriving July 27, which the closed flagships cannot match at any price.
What Is Kimi K3?
Kimi K3 is the new flagship AI model from Moonshot AI, a Beijing-based company. Released July 16, 2026, it has 2.8 trillion parameters, which makes it the largest open-weight model ever announced. It handles text, images, and video, and supports a context window of one million tokens, meaning it can process very long documents or coding sessions in one go.
The comparison everyone is making is against the two leading closed models: Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol. Unusually for an AI launch, Moonshot itself admits K3 still trails both on overall user experience. The benchmark charts tell a more complicated story than a simple third-place finish.
The Headline Numbers
| Model | AA Intelligence Index | Frontend Code Arena | API Cost (per 1M output tokens) |
|---|---|---|---|
| Kimi K3 | 57.1 | 1,679 (#1) | $15 |
| Claude Fable 5 | 59.9 | 1,631 | $50 |
| GPT-5.6 Sol | 58.9 | 1,618 | $15 |
Two leaderboards, two different winners. On Artificial Analysis's Intelligence Index, an independent composite of many tests, K3 lands third behind Fable 5 and GPT-5.6 Sol. On Arena.ai's Frontend Code Arena, where real developers vote blind on which model built the better web interface, K3 took first place outright, jumping 17 spots from where the previous Kimi model sat.
Why Do the Leaderboards Disagree?
Because they measure different things. Frontend Code Arena is a human preference contest focused on one skill: building web interfaces. K3 ranked first in six of seven frontend categories there. Composite indexes like Artificial Analysis average dozens of tests across reasoning, knowledge, and agent tasks, where the closed flagships still hold the edge.
The coding benchmarks split the same way. K3 leads on SWE Marathon, a test of long sustained coding sessions, at 42.0. It edges Fable 5 on Terminal Bench 2.1 (88.3 vs 84.6) while trailing GPT-5.6 Sol there by half a point. On DeepSWE, the order flips: Sol leads at 73.0, Fable 5 follows at 70.0, and K3 sits at 67.5. No model wins everywhere.
Reasons to Read the Numbers Skeptically
We took the same approach in our Is Fable 5 Nerfed? analysis, and it applies here too. Several caveats matter:
- Mixed test setups. Depending on the benchmark, models ran through different software harnesses (KimiCode, Claude Code, Codex, and others). That limits clean model-to-model conclusions.
- Tiny gaps. Artificial Analysis puts the margin of error on its index at roughly one point. K3's gap to GPT-5.6 Sol is about half a point, too small to call decisive.
- Max effort settings. All of Moonshot's published K3 scores use maximum reasoning effort, which is the most expensive and slowest way to run any model.
- Disputed scoring. The author of one coding benchmark publicly noted that Moonshot used a scoring method his team does not recommend, one that can overstate results.
- Hallucinations went up. Independent testing measured a 50.9 percent hallucination rate for K3 on one factual-reliability test, up from 39.3 percent for its predecessor.
The Cost Story Is Clearer Than the Capability Story
K3's API pricing is $3 per million input tokens and $15 per million output tokens, with cached input at $0.30. Fable 5 runs $10 and $50 for the same units, roughly 3.3 times more on output. One wrinkle: K3 used significantly more tokens than its rivals to finish the same evaluations, about 1.9 times Sol's total, so cheap per-token prices do not always mean a proportionally cheaper bill. Even so, most per-task cost comparisons still favor K3 by a wide margin.
What "Open" Buys You That Closed Models Cannot
The full model weights are scheduled for release on July 27, 2026 under a modified MIT license. That matters in ways no benchmark captures. Open weights mean no deprecation risk (the model cannot be shut off or quietly changed), a choice of hosting providers competing on price, and the option to run it in whatever legal jurisdiction you prefer.
The catch is scale. At 2.8 trillion parameters, K3 is far beyond consumer hardware, a topic we covered in detail in Can You Run Kimi K3 Locally? The 2.8T Hardware Reality. Until the weights actually ship, using K3 means using Moonshot's API, and that carries the data-handling tradeoffs we documented in Is Kimi K3 Free?. Your prompts route through servers you do not control, which is the same data-sovereignty question that applies to every hosted AI service, closed or open.
Bottom Line
Kimi K3 is a real frontier-class model with standout results in frontend coding and long coding sessions, at a fraction of Fable 5's price. It is not an across-the-board replacement for Claude Fable 5 or GPT-5.6 Sol, and Moonshot says as much itself. The sensible read: the gap between open and closed models has narrowed to the point where the right choice depends on your task, your budget, and how much control you want over where your data lives.
Frequently Asked Questions
Is Kimi K3 better than Claude?
It depends on the task. Kimi K3 beat Claude Fable 5 on the Frontend Code Arena leaderboard and on some coding tests, but Fable 5 scores higher on overall intelligence rankings and won the most individual benchmarks across Moonshot's own published tests.
Is Kimi K3 open source?
It is open weight, not fully open source. The trained model files are scheduled for public release on July 27, 2026 under a modified MIT license, but the training data and code remain private.
How much does Kimi K3 cost?
The API costs $3 per million input tokens and $15 per million output tokens, with cached input discounted to $0.30. That is about one third of Claude Fable 5's output pricing. The downloadable weights will be free.
Who makes Kimi K3?
Moonshot AI, an artificial intelligence company based in Beijing, China. It launched its first Kimi chatbot in October 2023 and has released some of the largest open-weight models available.
What is the Frontend Code Arena?
A live leaderboard run by Arena.ai where developers compare two AI-built web interfaces side by side without knowing which model made them, then vote for the better one. Kimi K3 debuted there at number one with 1,679 points.
Can you run Kimi K3 on your own computer?
Not realistically. Even heavily compressed, the 2.8 trillion parameter model needs hundreds of gigabytes to over a terabyte of memory, which is beyond any consumer machine. It requires server-class hardware or a hosted API.
When do the Kimi K3 weights come out?
Moonshot says the full model weights will be released by July 27, 2026. Until then, K3 is only available through the Kimi app and the Moonshot API.
