Kimi K3 vs Claude Opus 5: Which Should You Use? (2026)

Two frontier models landed eight days apart in July 2026. Moonshot AI shipped Kimi K3 on 16 July — a 2.8-trillion-parameter mixture-of-experts model with downloadable weights. Anthropic announced Claude Opus 5 on 24 July. They now sit at almost the same point on the capability frontier, priced very differently, and licensed on opposite philosophies.

This page is the decision guide: which one you should actually put behind your product, and when. For the raw multi-model benchmark tables, see our Kimi K3 benchmarks comparison. For the full token-cost breakdown, see Kimi K3 pricing and API costs.

How do Kimi K3, Claude Opus 5 and Claude Opus 4.8 compare head to head?

Opus 4.8 is included because it is still the Anthropic model with the most independently verified agentic numbers, and because a lot of teams are running it today and deciding whether to move.

The scaffold caveat matters more than any single number here. Moonshot's own evaluation table runs Kimi K3 under the Kimi Code harness while comparison models use Claude Code, Codex or mini-SWE-agent. Anthropic evaluates with its own scaffolds. Harness choice alone can move agentic scores by double digits, which is why the only row above that compares like with like is the Vals AI SWE-bench line — every model there runs the same minimal bash-only agent.

Two absences are worth naming. Neither Kimi K3 nor Claude Opus 5 appears on the verified Terminal-Bench 2.1 leaderboard at the time of writing. The top verified entries there are Claude Fable 5 with Claude Code at 83.8%, GPT-5.5 with Codex at 83.1%, and Opus 4.8 at 78.9%. Kimi's self-reported 88.3% would top that board if submitted — but it has not been.

Which is better for coding and agentic work?

Opus 5, but the margin is narrower than the marketing implies and it costs you.

On the one benchmark where both models run identical scaffolding — Vals AI's SWE-bench Verified with a bash-only agent — Opus 5 posts 97.0% against Kimi K3's 93.4%. That is a real, consistent lead, and it lands in the region where the remaining headroom is about three points. Opus 4.8, by comparison, sits at 88.6%, so the generational jump inside Anthropic's own line is larger than the gap between Opus 5 and K3.

Anthropic's launch claims for Opus 5 concentrate on long-horizon agent work rather than single-patch coding: it tops Frontier-Bench v0.1 with more than double Opus 4.8's score, lands within 0.5% of Claude Fable 5 on CursorBench 3.2 at half the price, scores roughly three times the next-best model on ARC-AGI-3, and posts about 1.5× the next-best pass rate on Zapier's AutomationBench. Those are vendor evaluations, but they are consistent in direction: Opus 5's advantage is reliability across many steps, not raw code generation on a single file.

Kimi K3's counter-argument is agentic breadth at a lower price. Moonshot reports 94.5 on MCPMark-Verified, 91.2 on BrowseComp and 95.0 F1 on DeepSearchQA — tool-calling and research-agent territory where K3 is genuinely competitive rather than merely close. It is also natively multimodal across text, image and video, with a 401M-parameter MoonViT-V2 encoder, which Opus 5 does not match on video input.

Practical read: if a failed agent run is expensive — a production migration, a repo-wide refactor, an overnight autonomous job — Opus 5's extra reliability is worth the premium. If failures are cheap and retryable, K3 gets you most of the way at roughly 60% of the token cost.

Which handles long context and long documents better?

This is close to a tie, with one asterisk each.

Kimi K3's context window is 1,048,576 tokens, and Moonshot prices it flat — a 900k-token request costs the same per token as a 9k-token one. Its Kimi Delta Attention architecture (69 KDA layers plus 24 Gated MLA layers across 93 total) exists specifically to make long-context inference cheaper, and Moonshot claims a substantial reduction in long-context inference cost versus prior attention designs.

Claude Opus 5 also offers a 1M-token window, also at standard pricing across the whole window, with prompt caching and Batch API discounts applying at full rate. Max output is 128k tokens on the synchronous Messages API, rising to 300k on the Batch API behind the output-300k-2026-03-24 beta header.

The asterisk on Anthropic's side: the tokenizer introduced with Opus 4.7 and carried into Opus 5 produces roughly 30% more tokens for the same text than earlier Claude models. That figure is a within-Claude comparison, not a comparison against Moonshot's tokenizer — but it is a concrete reminder that "1M tokens" means a different amount of your document depending on whose tokenizer counts it. Benchmark your own corpus before assuming the windows are equivalent.

The asterisk on Moonshot's side: a 1M-token request against a 2.8T MoE model is slow, and K3's already-modest output throughput does not improve with a full window.

Which one is cheaper to run in production?

Kimi K3 on list price, comfortably — but the honest answer has three layers.

Layer one — sticker price. K3 is 40% cheaper on both input and output. Both vendors price cache hits at 10% of input, so heavy prompt caching preserves the ratio rather than closing it.

Layer two — the batch inversion. Anthropic's Batch API takes 50% off both directions, which puts Opus 5 at $2.50 input and $12.50 output. That undercuts Kimi K3's list price. If your workload is asynchronous — nightly evaluation runs, bulk document processing, offline code migration — the cheaper model is Opus 5, not K3. This is the single most commonly missed fact in this comparison.

Layer three — tokens are not comparable units. Per the tokenizer note above, the same source text bills differently on each platform. Run 10,000 representative requests through both before you build a spreadsheet on list prices.

Self-hosting changes the math only at serious scale. K3's MXFP4 weights are roughly 1.4 TB before KV cache, activations and routing state; Moonshot recommends a supernode configuration of 64 or more accelerators for production serving, and vLLM's published recipe targets eight GB300-class GPUs as a floor. That is a real cluster with a real on-call rotation — not a cost saving unless you are already burning tens of millions of output tokens a day.

Is Kimi K3 actually open source?

No. It is open weight under a bespoke licence, and the distinction has teeth.

Despite widespread reporting that K3 shipped under a "Modified MIT" licence, the file on Hugging Face is a custom document titled the Kimi K3 License. It grants you the right to use, copy, modify, distribute and create derivative works, including commercially. It then attaches two conditions:

• Model-as-a-Service revenue trigger. If you run a MaaS business that, with its affiliates, exceeds $20 million in aggregate revenue over any consecutive 12 months, you must enter a separate agreement with Moonshot AI before commercial use.
• Attribution trigger. Commercial products above 100 million monthly active users or $20 million in monthly revenue must prominently display "Kimi K3" in the user interface.

Both carve out purely internal use — if you are not exposing the model's capabilities to third parties, neither condition applies — and use through Moonshot's official products or certified inference partners. A revenue-triggered obligation is a field-of-use restriction, which is why Hugging Face carries the model as license:other rather than as an OSI-approved licence.

For the overwhelming majority of teams reading this, that is a distinction without a practical difference: you can download the weights, run them on your own hardware, in your own region, under your own data-handling policy, and fine-tune them. That capability simply does not exist for Claude Opus 5 at any price. If your blocker is a regulator, a data-residency clause, or an air-gapped environment, this section is the entire comparison.

Anthropic's answer to residency is narrower: an inference_geo: "us" parameter that pins inference to the United States at a 1.1× price multiplier, plus distribution through Amazon Bedrock, Google Cloud and Microsoft Foundry. Real, but it is a routing guarantee, not possession of the model.

Which is faster in real use?

Depends entirely on whether a human is watching the cursor.

Artificial Analysis measures Claude Opus 5 at high effort generating around 55 tokens per second against Kimi K3 (max) at about 37 — Opus 5 is roughly 50% faster once it starts writing. But time to first token runs the other way and by a wider margin: 3.09 seconds for K3 against 13.62 seconds for Opus 5, because Opus 5 at high effort spends real time on adaptive thinking before emitting anything.

For a chat UI or an inline IDE completion, a 13-second silence is a usability problem and K3 feels dramatically more responsive. For a batch job or a long agent run, throughput wins and Opus 5 finishes first. Note that Opus 5's effort parameter defaults to high on the Claude API and in Claude Code; dropping it lowers latency and token spend, and is the first knob to reach for if Opus 5 feels sluggish.

Where is each model genuinely worse?

Where Kimi K3 loses:

• Roughly 3.6 points behind on the only like-for-like coding benchmark available (Vals AI SWE-bench Verified).
• Its flagship Terminal-Bench 2.1 figure of 88.3% is a vendor self-report on Moonshot's own harness, unverified on the public leaderboard.
• ~33% lower output throughput.
• Self-hosting is theoretical for most teams: ~1.4 TB of weights and a 64-accelerator recommendation.
• Thinner enterprise distribution — no Bedrock/Vertex/Foundry-equivalent procurement path.
• The $20M MaaS revenue trigger becomes a live negotiation if you succeed.

Where Claude Opus 5 loses:

• 67% more expensive per token at list, and the tokenizer change makes the effective gap wider on the same text.
• 13.6-second time to first token at default effort is bad for interactive products.
• No weights: no on-prem, no air-gap, no fine-tuning, no protection against deprecation. Anthropic has already retired Opus 4 and 4.1 from first-party access.
• It is also absent from the verified Terminal-Bench 2.1 board, so its agentic terminal claims rest on Anthropic's own evaluations.
• No native video input.
• Vendor lock-in is structural — your only lever on price is Anthropic's price list.

Which should you choose?

Choose Claude Opus 5 if: a failed run costs more than the token bill; you are doing repository-scale refactors, production debugging, or long-horizon autonomous agents; you need procurement through Bedrock, Vertex or Foundry; your workload is asynchronous and can use the Batch API at $2.50/$12.50; or you are already on Claude Code and the integration is doing real work for you.

Choose Kimi K3 if: token cost is a line item you actually manage; you need weights on your own hardware for residency, air-gapping or fine-tuning; you need native video understanding; time to first token matters because a human is waiting; or your failure mode is cheap and retryable.

If you are on Opus 4.8 today, the upgrade decision is the easy one: Opus 5 costs exactly the same per token, and the measured coding gap is bigger than the gap between Opus 5 and K3. We break that down in Claude Opus 5 vs Opus 4.8.

The realistic answer for most engineering teams is not "one of them". Route the expensive, high-stakes, low-volume work to Opus 5 and the cheap, high-volume, retryable work to K3. Both expose OpenAI-compatible endpoints, so the routing layer is a config change rather than a rewrite — and it is the only strategy that survives the next model release, which on 2026's cadence is about six weeks away.

FAQ

Is Kimi K3 better than Claude Opus 5?

Not overall. On the only benchmark where both run identical scaffolding — Vals AI's SWE-bench Verified with a bash-only agent — Opus 5 scores 97.0% against K3's 93.4%, and Artificial Analysis puts Opus 5 one point ahead on its Intelligence Index (61 vs 60). Kimi K3 wins on price, open weights, native video input and time to first token.

Is Kimi K3 cheaper than Claude Opus 5?

Yes at list price: $3.00 per million input tokens and $15.00 output, against Anthropic's $5.00 and $25.00 — about 40% less. But Anthropic's Batch API halves that to $2.50 and $12.50, which undercuts K3 for asynchronous workloads. Both vendors price cache hits at 10% of input, so caching does not change the ratio.

Is Kimi K3 open source?

It is open weight, not open source. The weights ship under a bespoke "Kimi K3 License", not Modified MIT as widely reported. Commercial use is permitted, but Model-as-a-Service businesses above $20 million in revenue over any consecutive 12 months must sign a separate agreement with Moonshot, and products above 100 million MAU must display "Kimi K3" in the UI.

Which is better for coding, Kimi K3 or Opus 5?

Claude Opus 5, by roughly 3.6 points on like-for-like SWE-bench Verified. Anthropic's launch evaluations emphasise long-horizon agent reliability — topping Frontier-Bench v0.1 and posting about 1.5× the next-best pass rate on Zapier AutomationBench. Kimi K3 is competitive enough that for cheap, retryable coding tasks the cost saving usually wins.

How does Kimi K3 compare to Claude Opus 4.8?

Kimi K3 is ahead on independent coding evaluation — 93.4% against Opus 4.8's 88.6% on Vals AI's SWE-bench Verified — and costs 40% less per token. Opus 4.8 holds the strongest verified Anthropic entry on the Terminal-Bench 2.1 leaderboard at 78.9%. Since Opus 5 costs the same $5/$25 as 4.8, there is little reason to stay on 4.8.

Can I run Kimi K3 locally?

Only on a serious GPU cluster. The MXFP4 weights are roughly 1.4 TB before KV cache and activations; Moonshot recommends a supernode of 64 or more accelerators for production, and vLLM's recipe targets eight GB300-class GPUs as a floor. Supported engines are vLLM, SGLang and TokenSpeed. It is open weight, not laptop weight.

What context window do Kimi K3 and Claude Opus 5 have?

Both are around 1M tokens — K3 at exactly 1,048,576 and Opus 5 at 1M — and both price the full window at standard per-token rates. Opus 5 caps output at 128k tokens, rising to 300k through the Batch API beta header. Note that Anthropic's newer tokenizer produces roughly 30% more tokens for the same text than pre-4.7 Claude models.

All benchmark figures above were checked against Moonshot AI's Hugging Face model card and GitHub repository, Anthropic's Claude platform pricing and model documentation, the Vals AI SWE-bench Verified leaderboard, the Terminal-Bench 2.1 leaderboard and Artificial Analysis, on 23 August 2026. Model releases at this cadence date quickly — re-check before committing a budget.