Kimi K3 vs Claude Opus 5.5: Benchmarks, Real Costs & When to Switch
September 25, 2026 · kimi k3 vs claude opus 5.5 · claude opus 5.5 · claude opus 5.5 pricing · claude opus 5.5 benchmarks · kimi k3 vs claude · kimi k3 pricing · ai model comparison
Claude Opus 5.5 is the stronger model, and Kimi K3 is the cheaper one per token. But since September 22 the gap has shrunk on price and widened on quality. Opus 5.5 cut Anthropic’s list price to $4/$20 per million input/output tokens and scores 58 on Artificial Analysis’s Intelligence Index to Kimi K3’s 44. K3 still costs 25% less per input and output token ($3/$15), ships open weights, and uses fewer output tokens per task. Opus 5.5’s $0.20 cache reads now cost less than K3’s $0.30.
This comparison is current as of September 25, 2026, three days after Anthropic released Claude Opus 5.5. It builds on our July Kimi K3 vs Claude Opus 5 breakdown rather than repeating it. Here’s what changed since then:
- The per-token price gap fell from 1.67x to 1.33x.
- The independent benchmark picture has moved against K3. Part of that is a better Claude, and part of it is that Artificial Analysis rebuilt its index with harder tests.
- K3 is now on Amazon Bedrock too.
Kimi K3 vs Claude Opus 5.5 at a glance
| Kimi K3 | Claude Opus 5.5 | |
|---|---|---|
| Maker | Moonshot AI | Anthropic |
| Released | July 2026 | September 22, 2026 |
| Architecture | 2.8T-parameter MoE, 104B active | Undisclosed |
| API model ID | kimi-k3 | claude-opus-5-5 |
| Context window | 1,048,576 tokens | 1M tokens, no long-context premium |
| Max output | No published cap | 128K (300K on Batch API, beta) |
| Thinking | Reasoning model, effort modes incl. low and max | Adaptive, always on; low / medium (default) / high / xhigh / max |
| Input / output (per 1M) | $3 / $15 | $4 / $20 |
| Cache read (per 1M) | $0.30 | $0.20 |
| Cache write, 5 min / 1 h | $3 / $6 | $5 / $8 |
| Bulk discount | None published by Moonshot; Bedrock Flex at 0.5x | Batch API 50% off ($2 / $10) |
| Input modalities | Text + images (native vision) | Text + images |
| Open weights | Yes | No |
| Where to run it | Moonshot API, 20+ third-party hosts, Amazon Bedrock | Claude API, Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS |
Sources: Anthropic’s Opus 5.5 model page and pricing docs, Moonshot’s pricing page, and the AWS Bedrock model card for Kimi K3. Full K3 hardware specs are in our specifications page.
What changed from Opus 5 to Opus 5.5
Opus 5.5 came out two months after Opus 5. TechCrunch framed it as Fable-level performance at a lower price. Anthropic’s pitch rests on three points: it costs less, it’s faster (output “more than 30% faster than Opus 5”), and it’s more efficient (“40% less to run than Opus 5” on typical workloads). Its launch post also puts Opus 5.5 ahead of the pricier Fable 5.1 on most of its own evals.
Anthropic’s reported numbers, which are vendor-run and use effort settings Anthropic chose:
| Benchmark (Anthropic-reported) | Opus 5 | Opus 5.5 | Fable 5.1 |
|---|---|---|---|
| Terminal-Bench 4.0 | 52.3% | 66.4% | 55.8% |
| CursorBench 4.0 | 46.6% | 57.8% | 51.8% |
| FrontierCode v1.1 (main) | 48.0% | 54.4% | 50.3% |
| GDPval-AA v2.1 (Elo) | 1708 | 1846 | 1735 |
| AutomationBench | 26.9% | 40.0% | 31.4% |
| Humanity’s Last Exam (tools) | 63.6% | 67.7% | 65.6% |
| OSWorld 2.0 (partial) | 74.0% | 81.8% | 80.7% |
Anthropic didn’t publish an SWE-bench Verified or SWE-bench Pro score for Opus 5.5. Kimi K3 doesn’t appear in its tables. (GPT-6 Astra does, and we cover that matchup in Kimi K3 vs GPT-6 Astra.)
On price, input and output are down 20%. Cache reads are down 60%, from $0.50 to $0.20. Five-minute cache writes fell from $6.25 to $5. Fast mode, a research preview on the first-party API only, went from $10/$50 to $8/$40.
If you already use Opus 5, check the four breaking API changes before you swap the model ID:
- Thinking can’t be disabled.
- Forced tool use returns an error.
- Thinking blocks are tied to the model and conversation that produced them.
- The older
computer_20251124tool is rejected on the Claude API and Google Cloud.
Also, text between tool calls now comes back inside thinking blocks. So a UI that streams progress messages can go quiet between tool calls.
Pricing: every billing line compared
| Per 1M tokens | Kimi K3 (Moonshot) | Claude Opus 5.5 | Difference |
|---|---|---|---|
| Input | $3.00 | $4.00 | K3 25% cheaper |
| Output | $15.00 | $20.00 | K3 25% cheaper |
| Cache read / hit | $0.30 | $0.20 | Opus 33% cheaper |
| Cache write, 5-min TTL | $3.00 | $5.00 | K3 40% cheaper |
| Cache write, 1-hour TTL | $6.00 | $8.00 | K3 25% cheaper |
| Bulk / async | None published (Bedrock Flex: $1.50 / $7.50) | Batch: $2.00 / $10.00 | Depends on route |
| Speed tier | K3 Fast via Vercel AI Gateway, ~50% premium | Fast mode: $8 / $40 | Similar premium |
Two details on this table can change your bill.
Cache reads have flipped. Against Opus 5, K3 was cheaper on every tier. Anthropic now bills Opus 5.5 cache hits at 5% of base input, which is $0.20. That’s lower than K3’s $0.30. Moonshot still caches request prefixes automatically, and a 5-minute write costs the same as normal input ($3), so a first request carries no premium. Anthropic charges 1.25x input for that write. So K3 wins when your prompts change a lot, and Opus 5.5 wins when you re-read a huge, stable prefix many times. Example D below shows where the two lines cross.
Tokenizers are not the same. Anthropic’s pricing docs say its tokenizer since Opus 4.7 “produces approximately 30% more tokens for the same text” than its previous one. Nobody publishes a Claude-to-Kimi ratio, so our examples assume equal token counts. If Claude counts your text as more tokens, its effective price rises by that ratio, so measure it on your own prompts.
Independent benchmarks: the same harness for both models
Vendor tables compare a model against the rivals the vendor picked. Artificial Analysis runs both models through the same 10-evaluation suite, Intelligence Index v4.3.2:
| Evaluation (Artificial Analysis) | Kimi K3 (max) | Claude Opus 5.5 (max) | Winner |
|---|---|---|---|
| Intelligence Index v4.3.2 | 44 | 58 | Opus +14 |
| GDPval-AA v2.1 (Elo) | 1524 | 1846 | Opus +322 |
| AA-Briefcase v1.1 (Elo) | 1505 | 1822 | Opus +317 |
| AutomationBench-AA | 58% | 70% | Opus +12 |
| Terminal-Bench 4.0 | 13% | 60% | Opus +47 |
| Humanity’s Last Exam | 47% | 61% | Opus +14 |
| SciCode | 59% | 67% | Opus +8 |
| CritPt | 23% | 32% | Opus +9 |
| AA-Omniscience | 20 | 46 | Opus +26 |
| AA-LCR v1.1 (long context) | 89% | 85% | K3 +4 |
Opus 5.5 wins nine of ten evaluations here. K3 wins only long-context reasoning. That’s a very different picture from July, when K3 took three of five head-to-heads against Opus 5. Two things explain most of the shift:
The index changed. Artificial Analysis’s changelog shows two changes. On September 7 it swapped in Terminal-Bench 4.0, a harder 66-task set replacing v2.1, and replaced τ³-Banking with AutomationBench-AA. On September 19 it re-anchored the GDPval-AA Elo scale. K3 hasn’t gotten worse. Its lower number now (44, versus the 57 we quoted in July) comes from the new version, and scores from different versions can’t be compared. Old Opus 5 moved down too: on the same v4.3.2 index it scores 48 (high effort) against K3’s 44. K3 still beats Opus 5 there on AutomationBench-AA (58% vs 54%) and AA-LCR (89% vs 79%).
Opus 5.5 really is a big step up. On the identical harness at high effort, Opus 5.5 scores 54 to Opus 5’s 48, and it reaches 58 at max. Artificial Analysis calls it “the highest score we have measured by several points.”
K3’s 13% on Terminal-Bench 4.0 stands out (Opus 5 scores 46% on the same test). It comes from one harness, one effort setting and Moonshot’s first-party API, on a brand-new task set where tool-calling quirks hurt some models more than others. Treat it as a reason to test K3 on your own terminal workflows, not as the final word on its coding ability.
On coding agents specifically, Artificial Analysis’s Coding Agent Index v1.5 puts Opus 5.5 in Claude Code at 66, its highest ever (as reported by OrcaRouter). K3 in Kimi Code CLI scored 57 on an earlier version of that index, so we don’t compare the two numbers directly.
Token efficiency: cost per token vs cost per task
K3 costs less per token. What you actually pay for, though, is a finished task. On that measure the result depends on effort level:
| Artificial Analysis Intelligence Index run | Score | Cost per task | Output tokens per task |
|---|---|---|---|
| Kimi K3 (max) | 44 | $2.00 | ~48k |
| Claude Opus 5.5 (low) | n/a | $0.55 | n/a |
| Claude Opus 5.5 (high) | 54 | $1.82 | n/a |
| Claude Opus 5.5 (max) | 58 | $5.98 | ~119k |
Sources: Artificial Analysis comparison and Opus 5.5 high pages; “n/a” means not published. These figures use list prices under Artificial Analysis’s methodology.
The key comparison: Opus 5.5 at high effort scored 10 points higher than K3 at max, for about 9% less per task. At max effort, Opus 5.5 produces about 2.5x as many output tokens as K3 and costs about 3x as much per task. The same effect undercuts Anthropic’s “40% cheaper than Opus 5” claim: at max, Artificial Analysis found Opus 5.5 roughly “level with Opus 5 on cost per task,” and on the Coding Agent Index it cost $13.04 per task versus $10.79 for Opus 5. The price cut reaches your bill only at medium or high effort.
Worked cost examples
All figures use the list prices above, the Moonshot official API for K3 and the first-party Claude API for Opus 5.5. We assume equal token counts on both sides (see the tokenizer caveat). Output includes thinking tokens, and both models always think.
A. A coding-agent session
Assume a long session that processes 6M input tokens in total. 90% of them are cache hits, which Moonshot reports as typical for coding. That leaves 600K fresh tokens written to a 5-minute cache, plus 250K output tokens.
| Line | Kimi K3 | Claude Opus 5.5 |
|---|---|---|
| 5.4M cached input | $1.62 | $1.08 |
| 0.6M fresh input (cache write) | $1.80 | $3.00 |
| 0.25M output | $3.75 | $5.00 |
| Session total | $7.17 | $9.08 |
With the same token usage, Opus 5.5 costs 27% more. But Anthropic’s launch customers report finishing tasks in “a third to half fewer steps” (Lovable) and “about half the turns” (Optiver). If Opus 5.5 needs 40% fewer tokens, its session drops to about $5.45, cheaper than K3. At max effort, with the ~2.5x output tokens Artificial Analysis measured, it rises to about $16.48. Effort level moves the bill more than list price does.
B. A RAG endpoint
1,000 queries a day. Each carries 20K tokens of freshly retrieved context (not cacheable), a 2K cached system prompt, and 600 output tokens.
| Per 1,000 queries | Kimi K3 | Claude Opus 5.5 |
|---|---|---|
| 20M uncached input | $60.00 | $80.00 |
| 2M cached input | $0.60 | $0.40 |
| 0.6M output | $9.00 | $12.00 |
| Per day | $69.60 | $92.40 |
| Per 30 days | $2,088 | $2,772 |
This workload suits K3 best: mostly fresh input, short answers. K3 saves about $684 a month, a 25% saving at the same quality. Given K3’s AA-LCR lead, quality on retrieval-heavy prompts may well be equal.
C. A batch job
10,000 documents, each with 8K input and 1.5K output tokens, for 80M in and 15M out. Results can wait overnight.
| Route | Cost |
|---|---|
| Claude Opus 5.5, standard | $620.00 |
| Kimi K3, Moonshot standard | $465.00 |
| Claude Opus 5.5, Batch API (50% off) | $310.00 |
| Kimi K3, Bedrock Flex tier (0.5x) | $232.50 |
The batch result surprises people. Moonshot’s pricing page lists no batch discount, so Opus 5.5 on Anthropic’s Batch API is cheaper than K3 at Moonshot’s standard rate. You get the savings back by running K3 through Bedrock’s Flex tier, which AWS prices at half of standard. Flex is a lower-priority real-time tier, not an async batch queue, so check its latency and throttling before you commit to it.
D. Repeated questions over an 800K-token codebase
You load an 800K-token repository once into a 1-hour cache, then ask n questions within the hour. Each question produces about 5K output tokens.
| Questions in the hour | Kimi K3 | Claude Opus 5.5 |
|---|---|---|
| 10 | $7.71 | $8.84 |
| 20 | $10.86 | $11.44 |
| 40 | $17.16 | $16.64 |
K3’s cheaper cache write ($6 vs $8) gives it an early lead; Opus 5.5’s cheaper reads close it by about 5.5 cents per question. They break even at roughly 31 questions per hour, after which Opus 5.5 is cheaper.
Availability, speed and limits
Where you can run them. Opus 5.5 is available on the Claude API, Amazon Bedrock (anthropic.claude-opus-5-5), Google Cloud, Microsoft Foundry and Claude Platform on AWS. It’s also in Claude Code and on the Pro, Max, Team and Enterprise plans. K3 got a big distribution boost when AWS added it to Bedrock on September 18:
- Model IDs are
global.moonshotai.kimi-k3andus.moonshotai.kimi-k3. - Global pricing is $3/$15/$0.30. The US geo profile costs 10% more.
- Prompt caching retains entries for at least 30 minutes; video input isn’t supported.
- AWS recommends the OpenAI-compatible APIs over Converse, which has known multi-turn bugs with this model.
Add the independent hosts in our API providers guide, and “we can only buy through AWS” no longer rules K3 out.
Speed. Artificial Analysis measures Opus 5.5 at roughly 88 to 91 output tokens per second at high and xhigh effort. K3 on Moonshot’s first-party API runs at about 35 tokens per second. That’s roughly 2.5x, though some K3 hosts advertise far higher throughput than Moonshot’s own endpoint.
Output ceiling. Opus 5.5 guarantees 128K output, or 300K on batch with a beta header. Moonshot publishes no K3 output cap, which leaves more room but guarantees less.
Self-hosting. Only K3 offers this. The weights are about 1.56TB, and serving them needs about 1.7 to 1.84TB of accelerator memory, for example 8x B300 or 16x H200. Our VRAM requirements guide has the details, and the open-weights breakdown covers the license.
When to switch: a decision guide
| Your situation | Pick | Why |
|---|---|---|
| Hard agentic coding in a terminal | Opus 5.5 | Terminal-Bench 4.0: 60% vs 13% on the same harness |
| Knowledge work, analysis, reports | Opus 5.5 | GDPval-AA 1846 vs 1524 Elo |
| Failed runs are expensive (migrations, audits) | Opus 5.5 at high effort | Higher score for slightly less money per task |
| High-volume RAG with fresh context | Kimi K3 | 25% cheaper input and output; wins AA-LCR |
| Output-heavy generation at scale | Kimi K3 | $15 vs $20 output; fewer tokens per task than Opus at max |
| Overnight batch jobs | K3 on Bedrock Flex, or Opus 5.5 Batch | $232.50 vs $310 in example C |
| All-day Q&A over one huge cached prefix | Opus 5.5 | $0.20 cache reads win past ~31 queries/hour |
| Self-hosting, fine-tuning, air-gapped | Kimi K3 | Only one with open weights |
| Latency-sensitive interactive UX | Opus 5.5 | ~2.5x faster output on first-party APIs |
| Already on AWS, need a non-Anthropic fallback | Kimi K3 | Now on Bedrock with the same IAM and logging |
Opus 5 to Opus 5.5: easy to justify. Budget a day for the breaking changes, and don’t default to max effort, or your bill can go up rather than down.
K3 to Opus 5.5: worth it when a failed agent run costs more than the tokens (engineer time, a broken deploy, a wrong report). At high effort it’s no more expensive per task.
Opus 5.5 to K3: worth it for fresh-input, output-heavy work, for weights you control, or for a second vendor. Compare on your own prompts first; “same quality, 25% cheaper” is no longer a safe assumption. For the flagship tier, see Kimi K3 vs Claude Fable 5.1.
Honest caveats
- Opus 5.5 is three days old. Independent numbers exist only from Artificial Analysis so far.
- Vendor numbers are vendor numbers. Anthropic’s “40% cheaper” claim didn’t show up at max effort in independent testing, and its own launch post admits “benchmark margins have become a less reliable guide to real-world differences.”
- Our cost examples are simplified. Equal tokenization, fixed cache-hit rates, no tool fees. Change any of those and the result can shift by more than the list-price gap.
- K3 varies by host. Speed, caching and even quality differ between Moonshot, Bedrock and third-party hosts.
FAQ
Is Kimi K3 better than Claude Opus 5.5? On independent benchmarks, no. Artificial Analysis scores Claude Opus 5.5 at 58 (max effort) and 54 (high effort) on its Intelligence Index v4.3.2, against 44 for Kimi K3 at max effort. Opus 5.5 leads on GDPval-AA, AutomationBench-AA, Terminal-Bench 4.0, Humanity’s Last Exam and SciCode. K3 wins long-context reasoning (AA-LCR 89% vs 85%), costs 25% less per input and output token, and is the only one of the two with open weights.
How much cheaper is Kimi K3 than Claude Opus 5.5? Per token, 25% cheaper on input and output: $3 vs $4 per million input tokens and $15 vs $20 per million output tokens. But Opus 5.5 is cheaper on cache reads ($0.20 vs $0.30) and has a 50% Batch API discount ($2/$10). The price gap shrank from 1.67x against Opus 5 to 1.33x against Opus 5.5, and on cache-heavy or batch workloads it can disappear or flip.
What changed from Claude Opus 5 to Opus 5.5? Opus 5.5 launched September 22, 2026 with a 20% price cut ($4/$20 from $5/$25), cache reads down 60% to $0.20, and output Anthropic says is over 30% faster. Anthropic’s own benchmarks show big jumps, such as Terminal-Bench 4.0 66.4% vs 52.3% and AutomationBench 40.0% vs 26.9%. Four API changes can break existing code, including thinking that can no longer be disabled and forced tool use returning an error.
Is Claude Opus 5.5 cheaper per task than Kimi K3? It depends on effort. Artificial Analysis measured $1.82 per Intelligence Index task for Opus 5.5 at high effort versus $2.00 for Kimi K3 at max, so Opus 5.5 at high scored 10 points higher for slightly less money. At max effort Opus 5.5 used about 119k output tokens per task versus K3’s 48k, which pushed its cost to $5.98 per task.
Is Kimi K3 available on Amazon Bedrock like Claude Opus 5.5? Yes, since September 18, 2026. AWS lists Kimi K3 on Bedrock at $3 input, $0.30 cache read and $15 output per million tokens on Global cross-Region inference, with a US geo profile at a 10% premium and a Flex tier at half price. Opus 5.5 is on Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. We could not confirm K3 on Google Cloud or Azure.
Does Claude Opus 5.5 have a 1M token context window? Yes. Opus 5.5 has a 1M-token context window billed at standard rates across the whole window, with no long-context premium. Max output is 128K tokens on the synchronous API and up to 300K on the Message Batches API with a beta header. Kimi K3’s window is 1,048,576 tokens, and Moonshot publishes no output cap.
Should I switch from Kimi K3 to Claude Opus 5.5? Switch if task success matters more than token price: hard agentic coding in a terminal, knowledge work, or anything where a failed run costs more than the tokens. Stay on K3 if you need open weights or self-hosting, run high volumes of output-heavy work, or care most about long-context retrieval. In both cases, test on your own tasks and compare cost per completed task, not cost per token.
The bottom line
Opus 5.5 is a better model than Kimi K3 right now, and K3’s price advantage is smaller and less consistent than it was against Opus 5. On the same independent harness Opus 5.5 leads by 14 index points, runs about 2.5x faster, and at high effort costs slightly less per task. K3’s case rests on specific strengths: 25% cheaper tokens for fresh-input and output-heavy work, the best long-context retrieval score of the two, a Bedrock Flex route that beats Anthropic’s batch pricing, and open weights that no Claude model will offer.
Our practical advice: use Opus 5.5 at medium or high effort for work where mistakes are costly, and route high-volume, retrieval-heavy and batch traffic to K3. Then track cost per completed task on both, not cost per token.
Keep reading: the July Kimi K3 vs Claude Opus 5 comparison, Kimi K3 vs Claude Fable 5.1, every Kimi K3 API provider, and what self-hosting K3 really takes.