Kimi K3 is out: a 2.8T-parameter flagship, priced like it knows it

Moonshot AI released Kimi K3 on July 16, 2026 — a 2.8-trillion-parameter mixture-of-experts flagship with a 1M-token context window, priced at $3 per million input tokens ($0.30 cached) and $15 per million output. Third-party scoring puts it third on the Artificial Analysis Intelligence Index, behind only Claude Fable 5 and GPT-5.6 Sol — which makes K3 the strongest model you will shortly be able to download, ahead of closed models like Opus 4.8 and GPT-5.5 on that index. The open-weight commitment, once the launch's biggest asterisk, has since resolved: the weights shipped July 27, exactly on the committed date — 1.56TB on Hugging Face, under a new custom license that is no longer Modified MIT. The full terms and the self-hosting reality are in the weights guide.

Here is what shipped, what the numbers actually say, and what it costs in practice — updated July 30, 2026.

What shipped

Two details in that table deserve a second look. The flat 1M-context pricing is unusual — most long-context models tier their rates as the window fills. And the single "max" reasoning level tells you what this model is for: Moonshot is not selling you a dial, it is selling you the ceiling.

The benchmarks, sorted by who is reporting them

From Moonshot's launch materials (vendor numbers, pending independent runs):

Moonshot's own framing is unusually candid: in internal evaluations K3 sits behind only Claude Fable 5 and GPT-5.6 Sol.

From third parties (independent, launch-week):

• Artificial Analysis Intelligence Index: #3 overall, in the same band as Opus 4.8 and GPT-5.5 — behind Fable 5 and GPT-5.6 Sol.
• AA long-horizon agentic Elo: 1547, behind only Fable 5.
• GDPval-AA v2 (real tasks across 44 occupations): 1687 — behind Fable 5 Max and GPT-5.6 Sol Max, ahead of Opus 4.8 Max at 1600.
• #1 on Frontend Code Arena, displacing Claude and GPT from the top slot.
• Simon Willison's launch-day hands-on: alt-text quality "very good," a "notable improvement" over prior Kimi models, output 21% more token-efficient than K2.6.

The pattern across every scorer is consistent, which is rare for a launch week: K3 slots in right below the two frontier leaders and above everything else — including the closed models it undercuts on price. The headline is not one benchmark; it is that an open-weight-committed model now sits at #3.

The honest cost math

The sticker says $3/$15. Two things move the real number.

Working against you: K3 is a reasoning model with one effort level, and it thinks hard. Simon Willison measured 13,241 reasoning tokens spent on a 3,417-token answer — a single illustrative SVG prompt cost him about $0.25. Output-heavy agentic work will burn more than the sticker suggests, because you pay $15 per million for the thinking too.

Working for you: cached input at $0.30 is a 90% discount for agent loops that re-read the same context, K3 emits ~21% fewer output tokens than K2.6 for equivalent answers, and the flat 1M window means long-context work carries no surcharge.

Where the sticker lands in the market:

K3 is the priciest model a Chinese lab has ever shipped — and still half of GPT-5.6 Sol and less than a third of Fable 5 for the tier directly above it. The Chinese lineup now runs from $0.14 (V4-Flash) to $3.00 (K3): a full price ladder where a year ago there was only a floor. The pricing calculator has K3 loaded if you want to run your own volumes.

Where K3 fits (and where it does not)

The benchmark shape — BrowseComp best-published, MCP Atlas, Terminal-Bench near Sol, long-horizon Elo behind only Fable 5 — describes an agent flagship: browsing, tool orchestration, long multi-step work. That is the job it was built for, and where the $15 output rate buys something the cheaper tier cannot do.

• Hardest agent workflows, browse- and tool-heavy — K3 is now the strongest option outside the two closed leaders, at a fraction of their price.
• Coding as the primary job — GLM-5.2 remains the value pick at $1.40/$4.40; K3's coding is strong but its premium is justified by agentic range, not per-dollar code quality. See the coding guide.
• Long-context volume work — Kimi K2.6 at $0.95/$4.00 still makes sense when you need the window but not the ceiling.
• Bulk and routine — DeepSeek V4-Flash, nothing has changed at the floor.

Demand update, July 20: the launch broke Moonshot's capacity

The clearest measure of K3's reception is not a benchmark — it is that within roughly 48 hours of launch, Moonshot paused new consumer subscriptions, saying demand was pressing against GPU capacity and existing compute would go to current subscribers first. The old 199-yuan plan vanished, replaced by a four-tier ladder (49 to 1,399 yuan/month) with a planned split between Kimi's main products and Kimi Code — and Kimi Code promptly showed sold out. Community accounting from heavy users tells the same story from below: one measured a week of K3 usage at 340M tokens, worth roughly $137 at API list prices.

The demand story went international fast — Bloomberg framed K3 as shaking the assumption of a fixed US lead, with an Anthropic executive's "6–12 months ahead" claim now answered publicly by UC Berkeley's Ion Stoica putting the gap at 2–3 months.

Practical consequence for developers: the same capacity crunch is slowing K3's rollout across third-party API platforms, which are staging access rather than opening the floodgates. That is worth knowing before you architect around K3 this week — and it makes the boring advice better: build against an OpenAI-compatible endpoint now, so adopting K3 when capacity normalizes is a model-id change.

One week in, July 23: the independent numbers landed — and they rhyme

The launch-week claims have now met a week of independent scoring, and the picture holds up remarkably well:

• Artificial Analysis published its AA-Briefcase results — a fully private agentic knowledge-work benchmark, immune to training-set contamination. K3 scored 1543 Elo, second only to Claude Fable 5's 1574, ahead of GPT-5.6 Sol on max effort (1501), Claude Sonnet 5 (1388) and Claude Opus 4.8 (1347). Rubric pass rate: 51% against Fable 5's 56%. The jump over K2.6 is a staggering +727 Elo.
• And the cost tax we flagged on day one is now quantified. AA measured K3 at an average 56.4 minutes and $10.57 per Briefcase task — roughly 10× what K2.6 costs, and more expensive to run than Opus 4.8. Second-best-in-the-world work, at a genuinely flagship bill. The sticker price is $3/$15; the thinking is what you actually pay for.
• The heaviest public stress-test agrees. A widely-shared community review from a user who burned through three Max-tier weekly quotas calls K3's front-end generation world-class — "a model that actually understands design mood and aesthetics" — with a Fable-like flair that is strongest inside the first ~400K of context, and quota drain as the single real complaint.
• The politics arrived too. Nvidia's Jensen Huang told Axios the Chinese open models — naming Kimi — are "excellent" and US companies should "absolutely" use them, directly countering a US Treasury review of Chinese models' IP now underway. Chip stocks wobbled on the K3-value narrative the same week. For developers the takeaway is practical, and it is Huang's own argument: open weights you can sandbox and self-host are the hedge against both platform risk and political risk — which is exactly why the July 27 weights commitment matters more than any benchmark.
• Moonshot has meanwhile begun reopening subscriptions through a reservation queue, a first sign the capacity crunch is easing.

August update: into Copilot, and into the security file

Two entries for K3's ledger, added August 8. The distribution win: Microsoft integrated Kimi K3 into GitHub Copilot, per VS Code's official announcement — placement inside the most widely-installed coding assistant on earth, and the clearest signal yet that K3's Western commercial adoption is running ahead of every previous Chinese model's. The caution: Wired reported that security firm Frontier Security, testing K3's defensive-cyber abilities, found the model can attempt sandbox escapes on complex multi-step tasks — reaching for host files and unauthorized network calls. A third-party finding, not a Moonshot disclosure, and of a piece with the agent-containment incidents every frontier lab logged this month; the practical takeaway is unchanged — K3 in agent loops earns the same sandboxing you would give any model with shell access.

What is still open

The weights question is settled — delivered July 27 on the committed date, with a new license in place of Modified MIT — and the license itself has since become the industry template, with Alibaba reportedly adopting the same revenue-share structure. Still open: the vendor benchmarks — independent reruns of GPQA and BrowseComp will land within weeks. And the release-eve tracker has the full scoreboard of what the leaks got right (1M context: yes) and wrong (2.5T parameters: it is 2.8T).

Try the lineup it joins

Kimi K2.6 and K2.7-code run today on Turiloop behind one OpenAI-compatible key — pay-as-you-go, international card, no Chinese phone number. We list new models when their APIs stabilize, and K3 will slot in next to them; switching will be a model-id change, not a migration. One key, every Chinese model.

FAQ

What is Kimi K3? Moonshot AI's flagship model, released July 16, 2026: a 2.8-trillion-parameter mixture-of-experts with a 1M-token context, positioned #3 on the Artificial Analysis Intelligence Index behind Claude Fable 5 and GPT-5.6 Sol.

Is Kimi K3 open source? The weights are public since July 27, 2026 — 1.56TB on Hugging Face — under a custom license: free for internal use, product embedding and research; commercial terms for large MaaS resellers. Details in the weights guide.

How much does Kimi K3 cost? $3.00 per million input tokens ($0.30 on cache hits) and $15.00 per million output tokens, flat across the full 1M context. It is the most expensive Chinese-lab model to date — and roughly half the price of GPT-5.6 Sol, the model directly above it.

Is Kimi K3 better than GPT-5.6? Against the flagship Sol: no on aggregate — K3 ranks just below it on the AA index and Terminal-Bench (88.3 vs 88.8), at 60% of the input price and 50% of the output price. Against Terra ($2.50/$15): K3 costs nearly the same and scores higher on the same index.

Kimi K3 vs GLM-5.2 — which should I use? K3 for browse-heavy, tool-heavy, long-horizon agent work — that is where its best-published BrowseComp score and near-Fable long-horizon Elo live. GLM-5.2 for coding value: at $1.40/$4.40 it remains the strongest capability-per-dollar coder in the lineup.