Kimi K3.1 Pro is the open-weight model (its trained weights are published) that scores closest to Claude Opus 5.5 on an independent quality index. It scores 46, just under the 48 that Opus 5.5 reaches at its default setting. Its tokens cost half as much as Opus's.
What you save depends on the cost of a finished task, not the price per token. On the index, Kimi K3.1 Pro cost $3.68 per task. That is less than Opus 5.5 at its highest setting ($10.04) and more than at its default ($2.45).
Which open-weight model can replace Claude Opus, and how much would we save? It comes up when Opus runs in your product and its line on the invoice grows every month.
Which open-weight model comes closest to Claude Opus 5.5?
There is no open version of Claude Opus. Claude Opus 5.5, released in September 2026, is a proprietary model, so you cannot download it or run it yourself. The useful question is which published model comes closest.
Because its weights are published, any provider can run an open-weight model, and it behaves the same wherever it runs. Kimi K3.1 Pro from Moonshot was released as an open-weights model in October 2026.
The gap to closed models has narrowed. Since January 2026, the most capable open-weight models have trailed frontier closed models by an average of four months, according to Epoch AI. So the best open model today sits about where the best closed model was four months ago.
We are Lyceum, a sovereign European AI inference provider for open-source models: built and run in Europe, on GPUs in European data centres. Our serverless models run in Paris and Finland, and you call them through one OpenAI-compatible API. Your prompts and outputs are processed, not stored, and never used for training.
The Artificial Analysis Intelligence Index combines 9 evaluations into one score; the current version is v5.0.1. Opus 5.5 can reason before it answers, and effort is how long it may do so: low, medium, high, xhigh or max. Opus 5.5 uses medium unless you set another level.
| Model | Index score | Input | Cached input | Output |
|---|---|---|---|---|
| Kimi K3.1 Pro | 46 | $2.00 | $0.20 | $10.00 |
| GLM-5.3 | 32 (max effort) | $1.40 | $0.26 | $4.40 |
| DeepSeek V4.1 Flash | 29 (high effort) | $0.50 | $0.13 | $1.50 |
Index scores from Artificial Analysis, prices in US dollars per 1M tokens from our model pages, both read on 10 November 2026. The open models also have a reasoning setting: the index ran Kimi K3.1 Pro and GLM-5.3 at max, DeepSeek V4.1 Flash at high. For reference, Claude Opus 5.5 scores 59 at max, 57 at xhigh, 52 at high and 48 at medium effort.
Of our models, only Kimi K3.1 Pro comes within a few points of Opus 5.5. GLM-5.3 scores 32 at max effort, and DeepSeek V4.1 Flash 29 at high effort.
One combined score hides where a model is strong, as our guide on reading open-model benchmarks explains. Two of the index's own tests show it.
Terminal-Bench 4.0 gives an agent 66 tasks to solve in a terminal, from software to security and hardware. Kimi K3.1 Pro has the top score of every model tested, 67.2%.
GDPval-AA covers real work from 44 occupations, such as documents and spreadsheets, rated by Elo: a higher number won more blind comparisons. Opus 5.5 at max effort leads with 1866, and Kimi K3.1 Pro follows at 1861. Opus 5.5 at medium effort rates 1593.
Where Opus 5.5 still leads
Opus 5.5 still leads on overall quality at high effort and above. It scores 52 at high, 57 at xhigh and 59 at max, against 46 for Kimi K3.1 Pro.
It is also much faster per task. In Artificial Analysis's runs, Opus 5.5 at medium effort took 7.1 minutes per index task. Kimi K3.1 Pro took 37.5 minutes.
If your product needs Opus's top score on hard reasoning, no open model we serve has a published index score that high. If you run Opus at its default, Kimi K3.1 Pro scores 2 points lower and answers more slowly.
Why half the token price is not half the bill
Anthropic charges $4 per 1M input tokens and $20 per 1M output tokens for Opus 5.5. Our price for Kimi K3.1 Pro is $2.00 and $10.00, exactly half. Cached input, prompt text the model has already read, costs $0.20 per 1M tokens on both.
The catch is the reasoning. A reasoning model writes out its working before it answers, and those reasoning tokens are billed as output. A model that thinks longer pays for it on every task.
On the index, Opus 5.5 wrote 42.5M output tokens at medium effort, 56.1M at high, 101M at xhigh and 218M at max. Kimi K3.1 Pro wrote 113M. That is 2.7 times Opus at its default. Per task, that made Kimi K3.1 Pro about 1.5 times as expensive: $3.68 against $2.45.
We saw the same spread in our own test calls on 10 November 2026. We sent one arithmetic question to three of our models, twice each, at temperature 0. The question asked for the largest order that boxes of 6, 9 and 20 items cannot pack exactly; the answer is 43.
- GLM-5.3 answered correctly both times with 375 and 395 output tokens. The longer answer cost about $0.0018: 60 input tokens at $1.40 plus 395 output tokens at $4.40 per 1M.
- DeepSeek V4.1 Flash answered correctly with 3,479 and 1,260 output tokens. The longer answer cost about $0.0053: 78 input tokens at $0.50 plus 3,479 output tokens at $1.50 per 1M.
- Kimi K3 is the Kimi model before K3.1 Pro, with higher token prices. We show it as the nearest Kimi model we could call. It answered correctly with 1,509 and 2,373 output tokens. The longer answer cost about $0.036: 133 input tokens at $3.00 plus 2,373 output tokens at $15.00 per 1M.
The longest answer used about 9 times the tokens of the shortest. This is a smoke test with one question, not a benchmark. It shows what a price list does not: how much a model writes before it answers.
What switching saves at each Opus effort level
Artificial Analysis publishes the average cost of one index task for every model and setting. That figure already counts each model's tokens, so it is the fairer basis for savings. We use it for a hypothetical workload: 1,000 agent tasks a month, like the ones in the index.
monthly cost = tasks per month x average cost per task
Opus 5.5 at max effort: 1,000 x $10.04 = $10,040
Kimi K3.1 Pro: 1,000 x $3.68 = $3,680| Model and setting | Index score | Cost per task | 1,000 tasks a month | Kimi K3.1 Pro instead |
|---|---|---|---|---|
| Claude Opus 5.5, max | 59 | $10.04 | $10,040 | saves $6,360 (63%) |
| Claude Opus 5.5, xhigh | 57 | $6.05 | $6,050 | saves $2,370 (39%) |
| Claude Opus 5.5, high | 52 | $3.46 | $3,460 | costs $220 more |
| Claude Opus 5.5, medium (default) | 48 | $2.45 | $2,450 | costs $1,230 more |
| Kimi K3.1 Pro | 46 | $3.68 | $3,680 |
Hypothetical workload, calculated from Artificial Analysis's average cost per index task, read on 10 November 2026.
If your team runs Opus at xhigh or max effort, switching to Kimi K3.1 Pro saves 39% to 63% on tasks like these. It gives up 11 to 13 index points. If you run Opus at its default, the switch costs more on these tasks, for a score 2 points lower.
The example favours the model that writes fewer tokens, because index tasks are long and reasoning-heavy. It does not show your own tasks, short answers with little reasoning, savings from caching, or speed under load. For per-token prices across our catalogue, see what open models cost per million tokens.
How to test the switch on your own tasks
Measure the cost per task on your own work before you switch. Pick a set of real requests with known good answers, including the hard ones. Run them on Opus at the effort setting you use today, then on Kimi K3.1 Pro.
From your own code, point an OpenAI client at our base URL and log the output tokens of every answer.
from openai import OpenAI
client = OpenAI(base_url="https://api.lyceum.technology/openai/v1", api_key="lk_...")
resp = client.chat.completions.create(
model="moonshotai/kimi-k3.1-pro",
messages=[{"role": "user", "content": "Your real task here"}],
)
print(resp.choices[0].message.content)
print(resp.usage.completion_tokens) # output tokens, reasoning includedFor each request, record three things: whether the answer is right, how long it took, and how many output tokens it used. Multiply the tokens by each model's price to get your own cost per task.
If you work in Claude Code, our Anthropic-compatible endpoint answers its requests with our open models. Claude Code asks for a model tier with each request, such as Opus, Sonnet or Haiku. Our endpoint maps each tier to one of our open models. By default it sends the Opus tier to GLM-5.3. To test Kimi K3.1 Pro in Opus's place, name it directly.
/model moonshotai/kimi-k3.1-proOur guide on open models in Claude Code covers the setup, and testing an open model before you switch gives a fuller test plan.
Switch when Kimi K3.1 Pro gets your hard cases right at a lower cost per task than your current Opus setting. Serverless inference on our API is pay as you go, with no commitment. To set up the comparison with us, book a call to benchmark Kimi K3.1 Pro against Opus on your own workload.
Frequently asked questions
Can I run Claude Opus on my own servers?
No. Claude Opus is a proprietary model and its weights are not published. To run a model you can move between providers, use an open-weight model such as Kimi K3.1 Pro.
Does Kimi K3.1 Pro read images and long documents?
Yes. On our API it accepts text and images. One request can hold up to 1M tokens, its context window.
Which model answers the Opus tier in Claude Code on Lyceum?
GLM-5.3, by default. You can choose another of our models at any time by passing its model ID to the model command.