Kimi K3 Review: Frontier Performance at Half the API Cost, but the Weights Are Not Here Yet
Kimi K3 reviewed with live API pricing, independent benchmarks, production limits, and the decision rule versus GPT-5.6 Sol and Claude Fable 5.
Kimi K3 is the model to trial for cached, long-horizon coding, not the model to standardize on yet. Its API is $3 per million uncached input tokens and $15 per million output tokens, about half GPT-5.6 Sol's mixed-token cost, but the promised weights and technical report are still not public and Kimi's own docs flag unstable session switching, over-eager actions, and a user-experience gap.
Kimi K3 launched on July 16, 2026 as Moonshot AI's flagship for coding, agents, and end-to-end knowledge work. You can use it now through Kimi.com, Kimi Work, Kimi Code, or the kimi-k3 API.
The launch matters because K3 reaches the frontier price-performance conversation without pretending to win every dimension. Kimi itself says the model still trails Claude Fable 5 and GPT-5.6 Sol overall. That candor makes the narrower win more credible: K3 offers serious long-horizon capability at a materially lower API rate.
The Kimi K3 verdict in one table
Kimi K3 earns a controlled production trial, not a default-model migration. The model is cheap enough to test against real work and strong enough to win selected routes, but its launch-state limitations make a wholesale switch premature.
The explicit decision rule is simple: trial K3 when cost is constraining a long-context agent route; keep Sol or Fable when an error costs more than the token savings. If self-hosting is the requirement, wait for the actual weights and license.
What Kimi K3 actually is
Kimi K3 is an infrastructure-scale mixture-of-experts model, not a giant dense model that activates all 2.8 trillion parameters for every token. A mixture-of-experts model routes each token through a small set of specialist subnetworks. K3 uses Stable LatentMoE to activate 16 of 896 experts, which is how a model this large can serve responses at an API price below the closed frontier leaders.
Its official quickstart locks the core specification:
- 2.8 trillion parameters in total.
- 1,048,576 tokens of context, with flat token pricing rather than a long-context tier.
- Native visual understanding, accepting text and image input and returning text.
- Kimi Delta Attention, a hybrid attention design intended to carry information efficiently across long sequences.
- Attention Residuals, a method for retrieving useful representations across model depth instead of accumulating every layer uniformly.
- Automatic context caching, with no cache ID, time-to-live setting, or extra request parameter required.
Kimi attributes an approximate 2.5x scaling-efficiency improvement over K2 to those architecture and training changes. Treat that as a vendor architecture claim, not an application benchmark. The useful buyer-level consequence is easier to verify: K3 exposes a one-million-token workspace and charges a tenth of its uncached input rate when the prefix hits cache.
K3 is already available in four forms. Kimi.com is the chat surface. Kimi Work version 3.1.0 or later brings it to Windows and Apple silicon Macs. Kimi Code selects it from the terminal with /model. Developers call the kimi-k3 model through Kimi's API.
The API surface covers tool calls, JSON Mode, structured output, tool-choice constraints, and dynamically loaded tools. Vision input has one awkward limit: public image URLs are not accepted, so applications must send base64 data or a Kimi file reference.
The benchmarks say frontier-competitive, not frontier-best
Independent measurement puts Kimi K3 near the top, while also exposing the cost hidden by a low rate card: it is slow and verbose. Artificial Analysis scores K3 at 57, ranking it #4 of 189 models. It measures 62.0 output tokens per second, #91 of 189, and records 130M output tokens across its Intelligence Index compared with a 63M average.
That verbosity matters because output is the expensive side of K3's bill. Artificial Analysis reports spending $2,709.75 to evaluate the model. A model can be cheap per token and still become expensive per accepted result if it produces roughly twice the comparison-set output volume.
Kimi's own benchmark table shows the more nuanced capability pattern:
These are serious results, but they are not one clean head-to-head. Kimi ran K3 at max reasoning with temperature 1.0 and top-p 1.0. The evaluation runner changes across rows between Kimi Code, Claude Code, and Codex. Some Fable 5 results can include fallback to Opus 4.8. The table supports calling K3 frontier-competitive; it does not support calling K3 the universal winner.
The independent and vendor evidence agree on the useful conclusion. K3 is capable enough to challenge a production route. It is not proven enough to replace every route.
- K3 costs $3 input and $15 output per million uncached tokens, well below Sol and Fable 5.
- Cache-hit input falls to $0.30 per million tokens with automatic prefix caching.
- The 1,048,576-token context is not subject to a separate long-context price tier.
- Native vision, structured output, tool calls, and dynamic tool loading cover the core agent surface.
- Independent measurement places K3 #4 of 189 models on a broad intelligence index.
- Full weights, the final license, and the technical report are not public yet.
- Only max reasoning effort is available, so there is no cheaper or faster effort setting to route simple work toward.
- Kimi warns that missing thinking history or switching models mid-session can destabilize quality.
- Kimi's web-search tool is being updated and is not recommended in the near term.
- Independent measurement finds K3 slower and much more verbose than the comparison average.
The API math is the real reason to test it
Kimi K3 wins its first production argument on cost, especially when a coding session reuses a stable repository prefix. The official rate card charges $0.30 per million cache-hit input tokens, $3 per million cache-miss input tokens, and $15 per million output tokens.
For one million uncached input tokens plus one million output tokens, the direct comparison is:
- Kimi K3: $18
- GPT-5.6 Sol: $35
- Claude Fable 5: $60
K3 is 48.6% cheaper than Sol and 70% cheaper than Fable on that deliberately simple workload.
The better coding-agent example is a month with 10M input tokens and 1M output tokens, where 9M input tokens are stable enough to hit cache. Kimi says its official API exceeds a 90% cache-hit rate on coding workloads, though your repository layout and prefix stability decide whether your application reproduces it.
- K3: 9 x $0.30 cached input + 1 x $3 uncached input + 1 x $15 output = $20.70.
- Sol: 9 x $0.50 cached input + 1 x $5 uncached input + 1 x $30 output = $39.50.
- Fable: 9 x $1 cached input + 1 x $10 uncached input + 1 x $50 output = $69.00.
With no cache hits, the same workload costs $45 on K3, $80 on Sol, and $150 on Fable. Cache does not create K3's price advantage, but it preserves it during long sessions.
The trap is optimizing token price instead of cost per accepted outcome. Artificial Analysis saw K3 emit 130M output tokens where the comparison average was 63M. If your route needs more review, longer generations, or repeated correction, the $15 output rate can erase part of the headline saving.
Consumer membership is a separate purchase path from the API. Kimi's official membership pricing lists a free Adagio plan with 6 Agent credits and one concurrent Agent task, but no Kimi Code credits. Moderato starts at $19 per month, or $15 per month when billed annually, with 60 Agent credits and 1x Kimi Code credits. Higher monthly tiers are Allegretto at $39, Allegro at $99, and Vivace at $199.
That makes the free plan suitable for seeing K3's interface and output style. It does not replace an API evaluation for a production workload, because membership credits and API token billing are different meters.
The production walls are more important than the demo
Kimi K3's production risk is not raw intelligence. It is the interaction between session state, fixed reasoning effort, tool permissions, and an immature product surface.
Preserve the full thinking history
K3 was trained to receive preserved thinking history across a session. Kimi warns that quality can become highly unstable if an agent runtime drops that history or switches an existing session from another model to K3. This is more than a prompt-format detail: it means a model-router experiment can look worse than K3 really is if the route changes mid-conversation.
Start a K3 evaluation with a fresh session and keep the complete assistant history intact. Do not compare it by swapping only the model ID halfway through a Sol or Fable trajectory.
Max effort is the only effort
K3 always reasons, and reasoning_effort currently accepts only max. Lower and higher controls are planned, but they are not available now. That removes a useful production lever: simple work cannot be sent through the same endpoint at a cheaper, faster effort setting.
For a funded founder debugging a hard migration, forced max effort may be fine. For a mid-market CTO routing thousands of ordinary classification jobs, it is wasteful. Keep a cheaper model beside K3 instead of making K3 carry every request.
Fence actions more tightly than prompts
Kimi says K3 can make unexpected decisions when instructions are ambiguous because it was trained to push through long-horizon work. The practical fix is not a longer personality prompt. It is explicit action policy: allowlisted tools, confirmation before external writes, spending caps, bounded file paths, and rollback.
An agent that writes a strong patch but also changes an unrelated configuration is not autonomous in a useful sense. Treat proactive behavior as a capability that requires a smaller blast radius.
Do not build around the web-search tool yet
Kimi's API docs say web search is being updated and is not recommended in the near term. For research agents, that means bringing a separate search or retrieval layer instead of assuming K3's official tool is ready.
Kimi also discloses a noticeable user-experience gap versus Fable 5 and Sol. That phrase is broad, but the surrounding limitations make it concrete: session compatibility is brittle, effort routing is unfinished, and one official tool is in transition. None of those defects invalidate the model. They do set the scope of a sensible trial.
Open-weight is a promise until July 27
Kimi K3 is not an available self-hosting choice on July 17, 2026. Kimi promises the full model weights by July 27, with the technical report to follow alongside them. Until those artifacts and the final license are public, "open" describes the announced destination, not a downloadable production asset.
The scale also changes what open weight means. At 2.8 trillion parameters, raw weights occupy roughly 1.4 TB at 4 bits per parameter or 5.6 TB at 16 bits per parameter, before runtime overhead, cache, activations, and redundancy. Kimi recommends supernode deployments with 64 or more accelerators.
That is data-center infrastructure, not a workstation model. A solo technical builder should rent the API. A mid-market company considering private inference should price the cluster, networking, serving staff, and capacity headroom before treating self-hosting as a savings plan.
The distinction matters:
- Open weight means the trained parameters can be downloaded.
- Open source also requires code and license terms that permit the intended use.
- Practical self-hosting means the model fits an infrastructure and operating budget you can sustain.
The current open-weight model landscape already includes smaller options that are easier to deploy. K3 may join it at the capability frontier, but the release files and license have to arrive first.
Who should pick Kimi K3, GPT-5.6 Sol, or Claude Fable 5
Pick by the cost of failure, not the model's best benchmark row.
Kimi K3 is the price-performance trial
Kimi K3 is the right experiment when long context, cached repository work, or native vision creates a real workload advantage. At $3 input, $0.30 cached input, and $15 output per million tokens, it has room to lose some efficiency and still beat the frontier leaders on cost. Skip it when the session layer cannot preserve history, when official web search is central, or when a premature action could be expensive.
GPT-5.6 Sol is the production default to beat
GPT-5.6 Sol is the stronger default when the route depends on a mature tool surface and proven behavior across different jobs. It supports function calling, structured output, web and file search, code execution, hosted shell, computer use, MCP, and tool search. Its $5 input, $0.50 cached input, and $30 output rates are higher than K3's, but a lower review burden can justify that premium. The current GPT-5.6 routing guide explains when Sol, Terra, and Luna should split the work.
Skip Sol when a stable, repeatable route produces the same accepted result on K3. Paying the frontier premium after the cheaper model passes your evaluation is habit, not risk management.
Claude Fable 5 is the judgment premium
Claude Fable 5 is Anthropic's most capable generally available model for ambitious coding and knowledge work. It costs $10 input and $50 output per million tokens, with a 90% prompt-cache discount on input. Anthropic positions it for long-running projects that plan, delegate, test, and check their own work.
The wall is price and routing predictability. Sensitive cybersecurity and biology requests can fall back to Opus 4.8. Fable earns its premium only when fewer corrections, better judgment, or longer reliable autonomy produces more accepted work than K3 or Sol.
A reversible Kimi K3 evaluation plan
The safest K3 adoption path changes one route at a time and keeps the old model ready. A model launch should be treated like a dependency upgrade: observable, reversible, and narrow enough to diagnose.
Choose one expensive route
Start with a job where Sol or Fable cost is meaningful and K3 has a plausible advantage, such as repository-scale debugging, long-document analysis, or vision-guided frontend repair. Do not start with every AI call in the product.
Freeze the comparison conditions
Use the same prompts, files, tools, permission boundaries, and acceptance rubric. Hold the agent runtime constant wherever possible so a model difference is not confused with a tooling difference.
Start fresh and preserve history
Create new K3 sessions, retain the full assistant thinking history required by the API, and keep stable prefixes byte-consistent enough to earn cache hits. Never switch an in-flight session from another model to K3.
Constrain actions outside the model
Allowlist tools, require approval for external writes and destructive changes, cap spend, and preserve rollback. Put the boundary in code and permissions, not only in the system prompt.
Measure accepted outcomes
Record input, cached input, output, elapsed time, retries, rejected work, and review effort. Divide total spend by accepted outcomes. That number captures K3's low rate and its tendency toward longer output in one metric.
Promote only the routes K3 wins
Move a route when K3 meets the quality bar at lower total cost. Keep Sol or Fable as an explicit escalation path for difficult cases, and leave simple high-volume work on a cheaper model.
This plan does not require believing K3 is better in general. It requires proving K3 is better for one job. That is the only level at which a production model decision stays true for long.
What is Kimi K3?
Kimi K3 is Moonshot AI's 2.8-trillion-parameter flagship model for long-horizon coding, knowledge work, and reasoning. It accepts text and images, exposes a 1,048,576-token context window, and is available through Kimi's apps, coding tool, and API.
Will Kimi K3 be open source?
Kimi says full model weights will be released by July 27, 2026. As of July 17, the weights, final license, and technical report are not public, so commercial rights and reproducible self-hosting should be assessed when those artifacts arrive.
Can I run Kimi AI locally?
K3 is not a normal local model. Its 2.8-trillion-parameter weights are roughly 1.4 TB even at 4 bits per parameter before runtime overhead, and Kimi recommends supernodes with 64 or more accelerators. Most teams should evaluate K3 through the API.
Is Kimi entirely free?
No. The Adagio membership is free and includes 6 Agent credits, but it includes no Kimi Code credits. Paid memberships start with Moderato at $19 per month, while the Kimi API is billed separately by input and output tokens.
Is Kimi K3 better than GPT-5.6 Sol or Claude Fable 5?
Not overall. K3 leads selected vendor-reported rows and costs less, but Kimi says it still trails Sol and Fable overall. The practical question is whether K3 meets your route's acceptance bar at a lower total cost.
Want the current model picks mapped to the work they actually fit? Get the AI tools map for business owners, free.
Get the AI tools map for business owners
A plain-English map of which AI models and tools fit which job, updated as prices and capabilities change. Free to subscribers.
Jul 17, 2026
