Kimi K3 vs Opus 5
Open weights against a closed model
Official comparison as of 2026-10-20: K3 is Moonshot's open flagship — 2.8T parameters, a 1M context window, native vision, and the full weights are already public. On the Claude side Opus 5 is now a legacy model, and the current Opus flagship is Opus 5.5, released 2026-09-22. K3's advantages are 2.8T open weights and native vision. Moonshot itself acknowledges the model still trails Fable 5 and GPT-5.6 Sol.
Updated 2026-10-20
Switch models on the same endpoint and measure the cost per token yourself; check /models for callable models.
- Billed per token — live rates on /models
- Pay by card (Visa / Mastercard), Apple Pay, Google Pay or crypto
- Self-service: top up and activate instantly
Highlights
K3 scale
16/896 MoE, KDA+AttnRes, native vision, 1M context.
K3 official rates
Moonshot's official rates: $3 input and $15 output per 1M tokens, $0.30 on a cache hit. Anthropic's official rates: Opus 5 at $5 input and $25 output, Opus 5.5 at $4 input and $20 output.
Anthropic's current flagship
Anthropic's official models overview says that if you are unsure which model to use, you should start with Claude Opus 5.5 for most workloads; Opus 5 now sits in the legacy list and is still available.
Moonshot says it still trails
In the official Kimi K3 tech blog, Moonshot writes that Kimi K3's overall performance still trails Claude Fable 5 and GPT 5.6 Sol, while it shows frontier-level performance in its own evaluation suite and consistently outperforms the other tested models.
The matchup
The roster in Moonshot's official tech blog is Claude Fable 5, Claude Opus 4.8, GPT 5.6 Sol and GPT 5.5: in the GPU kernel optimization test, K3 performed competitively with Fable 5 and substantially outperformed Opus 4.8, GPT 5.6 Sol and GPT 5.5. The MiniTriton compiler, the chip design and the game generation demos are Moonshot's own showcases too. They show what K3 is good at, not that Opus 5 loses on those tasks — no same-source head-to-head accompanies them.
Why it is trending
At launch, Kimi K3 uses max thinking effort by default; Moonshot says low and high effort modes come in later updates. Kimi K3 is available on Kimi.com, Kimi Work, Kimi Code and the Kimi API; Moonshot says it is the world's first open 3T-class model.
Timeline
Kimi K3 launches, and the official tech blog publishes the evaluations and the case studies with it.
Moonshot commits to releasing the full Kimi K3 weights by 2026-07-27; the official model guide now states that the full weights have been released.
Anthropic releases Claude Opus 5.5. The third-party vs Opus 5 writeups the old timeline listed for late Jul are no longer cited here.
Confirmed vs caution
Confirmed
Checked on 2026-10-20. The official K3 specs/price/limitations have been verified line by line: the K3 side comes from Moonshot's official tech blog and the Kimi API Platform docs, the Opus side from Anthropic's official model pages and pricing pages. Opus 5 was the closed-source flagship when this page was written — as of 2026-09-29 the newest Opus from Anthropic is Opus 5.5 (released 2026-09-22, $4 / $20); as of 2026-10-20 Opus 5 sits in the legacy list. Claude Opus 5 was released on 2026-07-24 with the model ID claude-opus-5.
Caution
On price, Claude Opus 5 lists officially at $5/M input and $25/M output, against Kimi K3's $3/M and $15/M. Moonshot lists three caveats for K3: it was trained in the preserved thinking history mode, so the harness has to pass back all the historical thinking content; it may take unexpected decisions on your behalf when it runs into small issues or ambiguous intent; and overall the user experience still trails Claude Fable 5 and GPT 5.6 Sol. The evaluations in the tech blog are Moonshot's own runs.
How to choose
Go with Claude
Claude stack, speed-sensitive coding. Claude side: Opus 5 is the model released on 2026-07-24 with the model ID claude-opus-5, a 1M context window and 128K max output; it takes text and images as input and returns text.
Pick K3
K3 side: Moonshot's open flagship with 2.8T parameters and native vision, official rates $3/$15, self-host if you can feed a supernode — Moonshot recommends deploying it on supernode configurations with 64 or more accelerators.
How to use it
On the K3 side: call kimi-k3 on the Kimi API Platform, use Kimi Work on the desktop (3.1.0 or later) and switch to K3 in the terminal with /model in Kimi Code. The Claude side keeps the existing routes: Claude Code and the Claude API work as before.
On QCode
The QCode /models directory lists both kimi-k3 and claude-opus-5; one key goes through the standard API interface, so you can compare the two models with the same key. Billing is per token, with per-model rates at /models.
FAQ
Did K3 beat Opus 5?
Moonshot's official evaluation lines up against Claude Fable 5 and GPT 5.6 Sol, and K3 still trails both closed models; Opus 5 is not on the blog's comparison roster. On the Anthropic side, Opus 5 is now a legacy model and the current Opus flagship is Opus 5.5, released 2026-09-22.
Price gap?
Official unit prices side by side: Kimi K3 at $3 input and $15 output, $0.30 on a cache hit, per 1M tokens; Claude Opus 5 at $5 input and $25 output; Claude Opus 5.5 at $4 input and $20 output.
Local K3?
Self-hosting works: Moonshot states that the full model weights have been released and recommends deploying Kimi K3 on supernode configurations with 64 or more accelerators. Quantization-aware training uses MXFP4 weights with MXFP8 activations for broad hardware compatibility.
Thinking history?
K3 was trained with thinking history preserved: Moonshot writes that if the agent harness fails to pass back all the historical thinking content, or if an ongoing session with another model is switched over to K3, generation quality can become highly unstable. The official recommendation is a harness with verified compatibility, such as Kimi Code, and avoiding switching to K3 in the middle of a session.
Vs Fable 5 instead?
In its own Kimi K3 tech blog, Moonshot writes that Kimi K3 "still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol", and in the same paragraph says it showed frontier-level performance across its evaluation suite, consistently outperforming other tested models. Anthropic's official models overview puts Claude Fable 5.1 on demanding reasoning and long-horizon agentic work.
DeepSeek 0813 in this fight?
Yes. DeepSeek's same-source 0813 grid on Hugging Face includes a Kimi K3 row too; that grid is published by a third party and its scores are not an official measure, while the full table is on /deepseek-v4-pro-0813. To compare the two models directly on QCode, send the same prompt to kimi-k3 and claude-opus-5 and compare the answers and the usage.
Sources
The official Kimi K3 tech blog (kimi.com/blog/kimi-k3), the Kimi API Platform docs (platform.kimi.ai/docs), Anthropic's official models overview and pricing pages (platform.claude.com/docs, cla.com/pricing), the Anthropic newsroom post on Opus 5.5, and the QCode /models directory.
K3 and Opus, item by item
Out of quota? With an active personal plan buy a booster pack, without one buy Pay-as-you-go Packs, which start at $4 for a one-time quota you use within 7 days (top-ups start at $5).
Related
Kimi K3.1 Pro API pricing and launch status (2026-10-23)
Checked 2026-10-26: kimi-k3.1-pro official price $2.00 input, $10.00 output, $0.20 cached; Highspeed $4.00 / $20.00 / $0.20; CNY prices listed separately; weights by 2026-11-04.
Kimi K3.1 Pro vs Claude Opus 5.5 (2026-11)
Official comparison of Kimi K3.1 Pro and Claude Opus 5.5: release dates, context, official max output, thinking behavior, official prices, open weights and vendor-reported benchmarks, plus switching between them on one QCode key.
Opus 5 guide
Closed-side hero page.
Crypto payment guide
Nine coins, 56 networks, zero platform fee for AI coding API credits.
Not affiliated with Moonshot or Anthropic.