Moonshot's open-weights agentic model undercuts Claude Opus 5.5 by a quarter on the rate card (Moonshot now presents Kimi K3.1 Pro as its flagship). Caching and batch discounts make the real gap smaller than that on agent workloads.
Kimi K3 is 25% cheaper than Claude Opus 5.5 on input and output. But Opus 5.5's cache hit ($0.20) is cheaper than Kimi's ($0.30), so on cache-heavy coding agents the difference narrows. Kimi K3's open weights add an option Opus doesn't have: running it yourself.
| Kimi K3 | Claude Opus 5.5 | |
|---|---|---|
| Provider | Moonshot AI | Anthropic |
| Tier | Flagship | Flagship |
| Input / 1M tokens | $3.00 | $4.00 |
| Output / 1M tokens | $15.00 | $20.00 |
| Cache hit / 1M | $0.30 | $0.20 |
| Cache write / 1M | $3.00 | $5.00 |
| Context window | 1M | 1M |
| Notes | Previous-generation flagship, superseded by Kimi K3.1 Pro but still sold; open weights; cache write $3 (5-min TTL) or $6 (1-h TTL) | New leading Opus model (Sep 22, 2026); Fable 5.1-level performance at 20% less than Opus 5; included on Pro/Max/Team at no extra cost |
Kimi K3 is 25% cheaper on input tokens ($3.00 vs $4.00 per 1M). Kimi K3 is 25% cheaper on output tokens ($15.00 vs $20.00 per 1M).
| Workload | Kimi K3 | Claude Opus 5.5 | Difference |
|---|---|---|---|
| Support chatbot 10,000 conversations × (2,000 input + 400 output tokens) | $120 | $160 | Kimi K3 saves $40.00 |
| Document Q&A 1,000 questions × (30,000 input + 800 output tokens) | $102 | $136 | Kimi K3 saves $34.00 |
| Coding agent session 4M input tokens (90% cache hits) + 150,000 output tokens | $4.53 | $5.32 | Kimi K3 saves $0.79 |
Standard rates per 1M tokens from data on this site. The coding-agent row assumes 90% of input is read from the prompt cache (full input price where a model has no cache rate). Cache-write fees, batch discounts and long-context surcharges listed in the notes are not included. Updated November 2, 2026.
you want open weights you can self-host or fine-tune, or you run uncached, high-volume chat and generation where its lower input and output rates apply to every token.
most of your input is a cached context (Opus 5.5 charges $0.20 per cached 1M tokens), or you need Anthropic's platform, Batch API and enterprise deployment options.
Anthropic bills Opus 5.5 cache hits at 5% of input. Kimi K3's cache hit is 10% of its input, which works out higher per token. Both bill cache writes: Anthropic at $5.00 for a 5-minute cache, Moonshot separately by TTL (5 minutes or 1 hour), with caching applied automatically to repeated prefixes. For an agent that is 90% cached, the 25% headline gap shrinks noticeably.
Moonshot lists batch inference for kimi-k3 at 60% of standard price. Anthropic's Batch API is 50% of standard. On overnight jobs Opus 5.5's discount is deeper, which again narrows the gap — batched Opus 5.5 lands close to batched Kimi K3.
Because Kimi K3 is an open-weights model, very high-volume teams can compare the API price against hosting it on their own GPUs or a third-party inference provider. That trade only pays off at sustained volume and with ops capacity; for most teams the hosted API is simpler. Both models offer a 1M-class window — Moonshot lists 1,048,576 tokens for the K3 series.
On standard input and output, yes: $3.00 vs $4.00 and $15.00 vs $20.00 per 1M tokens. Cache hits are cheaper on Opus 5.5.
Kimi K3 is an open-weights model, so you can run it yourself as well as through Moonshot's API. Check the licence terms before commercial self-hosting.
Yes. Moonshot lists batch inference for kimi-k3 at 60% of the standard price. Anthropic's Batch API is 50% off.
How we choose: prices come from each vendor's official pricing page and are re-checked every week. We take no affiliate fees, so rankings aren't for sale.