Kimi K3 is worth a production pilot for agentic and coding workloads, and it is not worth switching to as a general-purpose default. Moonshot AI’s 2.8-trillion-parameter model ranks second of 47 models in independent testing, leads every model on long-horizon coding, and fabricates answers more often than the model it replaces.
Moonshot AI released Kimi K3 on July 16, 2026 through the kimi.com website and API, priced at $3 per million input tokens and $15 per million output tokens, with cache hits at $0.30 per million. Moonshot AI has scheduled the downloadable weights for July 27, 2026. Our launch report covers the announcement itself: Kimi K3 Is Here: Open-Weight Model That Rivals Claude Fable 5.
Kimi K3 Review 2026: The Verdict in Numbers
This Kimi K3 review scores the model on the 5 attributes that decide a business deployment: coding capability, agentic capability, general intelligence, factual reliability, and cost. Kimi K3 leads on 2 of the 5, places third on 1, and regresses on 1 against its own predecessor.
| Attribute | Result | Verdict |
|---|---|---|
| Long-horizon coding | SWE Marathon 42.0 — above Claude Fable 5 and GPT-5.6 Sol | Best available |
| Agentic capability | BrowseComp 91.2%, MCP Atlas 84.2% | Frontier tier |
| General intelligence | GDPval-AA v2 1,687 — third overall | Behind Fable 5 and GPT-5.6 Sol |
| Factual reliability | 51% hallucination rate, up from 39% on Kimi K2.6 | Regression |
| Cost | $3 / $15 per M tokens, $0.94 per agentic task | Undercuts Fable 5 by 70% |
Kimi K3 placed second of 47 models in independent testing at release, behind only GPT-5.6 Sol. That ranking measures capability alone and excludes the hallucination regression, which is why capability rank and deployment recommendation diverge in this review.
Where Kimi K3 Beats Claude Fable 5 and Grok 4.5
Kimi K3 wins on 3 measured axes: long-horizon coding, agentic browsing, and front-end code generation. It scores 42.0 on SWE Marathon, above both Claude Fable 5 and GPT-5.6 Sol, and it ranks first on Arena’s Frontend Code Arena. Its BrowseComp score of 91.2% measures autonomous web-browsing performance, and its MCP Atlas score of 84.2% measures tool use through the Model Context Protocol.
Those 2 agentic scores decide whether a model can finish multi-step work rather than answer a single prompt, which is the criterion that separates a chat model from an agent runtime. Teams building on this class of model should measure it against the shipped alternatives in our guide to the 12 Best AI Agents for Business Tasks, where the same agentic capabilities are compared across commercial platforms.
Where Kimi K3 Loses
Kimi K3 loses on general intelligence and on factual reliability. On GDPval-AA v2 — a benchmark scoring real-world tasks across 44 occupations and 9 major industries — Kimi K3 scored 1,687, third behind Claude Fable 5 Max at 1,815 and GPT-5.6 Sol Max at 1,747.8. On general intelligence it sits close to Claude Opus 4.8 and GPT-5.5 rather than ahead of them.
The reliability number is the one that should govern deployment decisions. Independent testing measured Kimi K3’s hallucination rate at 51%, up from 39% for Kimi K2.6 on the same evaluation. Kimi K3 gets more answers right than its predecessor and fabricates more of the answers it gets wrong — a profile that suits supervised agentic work with verifiable outputs, such as code that compiles and tests that pass, and disqualifies it from unsupervised customer-facing text.
Kimi K3 vs Grok 4.5 vs Claude Fable 5: The Business Comparison
The 3 models occupy 3 distinct price-performance positions, and no single model wins on every axis.
| Attribute | Kimi K3 | Grok 4.5 | Claude Fable 5 |
|---|---|---|---|
| API price (in/out per M tokens) | $3 / $15 | $2 / $6 | $10 / $50 |
| Artificial Analysis Intelligence Index | 57.11 | 54 | Ranked 1st |
| GDPval-AA v2 (real-world tasks) | 1,687 | Not published | 1,815 (Max) |
| Cost per agentic task | $0.94 | $0.31 | Not published |
| Hallucination rate (independent) | 51% | 54% | Not published |
| Weights downloadable | Scheduled July 27, 2026 | No | No |
| Context window | 1M tokens | Not published | Not published |
Prices verified July 2026. Artificial Analysis figures as reported at each model’s release.
Grok 4.5 is the cheapest per task at $0.31 and carries the highest measured hallucination rate at 54%, up from 25% on Grok 4.3. Claude Fable 5 is the most capable and the most expensive, at 3.3× Kimi K3’s input price and 3.3× its output price. Kimi K3 sits between them and is the only one of the 3 whose weights a company can host itself after July 27.
Who Should Pilot Kimi K3
Kimi K3 fits 3 buyer profiles: teams building coding agents, teams with data-residency requirements, and teams that need to fine-tune on proprietary data. The first profile is served by the SWE Marathon and Frontend Code Arena results. The second and third depend entirely on the July 27 weight release, because self-hosting is what removes the vendor from the data path.
Kimi K3 does not fit teams that need verified factual output without a review step, and it does not fit teams that require a US-jurisdiction vendor. Moonshot AI is a Chinese lab, and its models entered US production stacks through the open-weight route rather than through enterprise procurement, a shift we documented in Chinese AI Models Are Running US Enterprise Workloads.
How to Pilot Kimi K3 in 30 Days
A Kimi K3 pilot answers 1 question: does the model finish supervised multi-step work more cheaply than the incumbent, at an acceptable review cost? Run the pilot in 4 stages across 30 days.
- Select a verifiable workload. Choose tasks whose output validates automatically — code that compiles, tests that pass, data that reconciles. The 51% hallucination rate makes unverifiable output economically unattractive regardless of price.
- Baseline the incumbent. Record cost per completed task and human review minutes per task on the current model for 2 weeks before switching anything.
- Run both models in parallel. Route identical tasks to Kimi K3 and the incumbent through the API for 2 weeks, and compare completion rate, cost per completed task, and review minutes — not benchmark scores.
- Re-evaluate after the weight release. Moonshot AI has scheduled downloadable weights for July 27, 2026, which changes the cost model from per-token to per-GPU-hour and changes the data-residency answer entirely.
Two operational limits apply during the pilot. Kimi K3’s benchmark results at launch came predominantly from Moonshot AI’s own harness, with independent reproduction arriving afterwards from Artificial Analysis, so treat vendor tables as hypotheses. Moonshot AI also serves the model from a single vendor API today, which means the pilot carries the same concentration risk as any hosted frontier model until the weights ship.
- Kimi K3 Is Here: Open-Weight Model That Rivals Claude Fable 5 — our July 16 launch coverage and the full benchmark table
- GLM-5.2: Zhipu AI’s 1M-Context Open-Source Model — the release that set the open-source context-length bar
- LM Studio Bionic: The AI Agent Built for Open Models — the tooling for running downloaded weights in-house
The honest verdict is that Kimi K3 is a specialist, not a replacement. It is the best model available for long-horizon coding work and among the best for agentic browsing, at a fifth of Claude Fable 5’s price — that combination justifies a pilot on any codebase-facing workload this quarter. It is also a model whose hallucination rate went up, not down, and that is not a footnote: a 51% fabrication rate on the same evaluation where its predecessor scored 39% means every unverified output costs more review time than it saves. Run it where the output verifies itself. The date that changes the calculation is July 27 — if Moonshot AI ships the weights, the question stops being “is the API worth $3 per million tokens” and becomes “is 2.8 trillion parameters worth the GPUs”, which is a different and much more interesting conversation.
Related Coverage
- How to Read AI Coding Benchmarks — why self-reported scores and independent scores diverge
- 5 Agentic AI Security Risks Every Team Must Know in 2026 — the controls to apply before any agent model reaches production
Frequently Asked Questions
Is Kimi K3 worth it in 2026?
Yes, for coding and agentic workloads. Kimi K3 leads all models on SWE Marathon at 42.0 and ranks first on Arena’s Frontend Code Arena, at $3 per million input tokens against Claude Fable 5’s $10. No, for unsupervised factual output: independent testing measured a 51% hallucination rate.
Is Kimi K3 better than Claude Fable 5?
No, not overall. Claude Fable 5 Max scored 1,815 on GDPval-AA v2 against Kimi K3’s 1,687, and ranks first on the Artificial Analysis Intelligence Index. Kimi K3 beats Claude Fable 5 specifically on SWE Marathon long-horizon coding, where it scores 42.0.
How much does Kimi K3 cost?
Kimi K3 costs $3 per million input tokens and $15 per million output tokens through Moonshot AI’s API, with cache hits at $0.30 per million. Artificial Analysis measured its cost per agentic task at $0.94.
Can I self-host Kimi K3?
Not yet. Moonshot AI has scheduled Kimi K3’s downloadable weights for July 27, 2026. Until that release, the model is accessible only through the kimi.com website and the Moonshot AI API.
Is Kimi K3 the largest open AI model?
Yes. Kimi K3 carries 2.8 trillion parameters, ahead of the previous open-weight leader DeepSeek V4 Pro at 1.6 trillion and ahead of Alibaba’s announced Qwen3.8 at 2.4 trillion.
