Kimi K3.1 Pro vs Claude Opus 5.5: Half the Rate Card, 0.7 Points Behind on One Benchmark

Kimi K3.1 Pro and Claude Opus 5.5 are the same shape of model sold at exactly half the price. Moonshot's new flagship costs $2.00 per million input tokens and $10.00 per million output; Anthropic's costs $4.00 and $20.00. Both accept a million tokens of context. Both always think before they answer. And on the one benchmark both companies print, they are 0.7 points apart: 65.7% for Kimi K3.1 Pro against 66.4% for Claude Opus 5.5 on Terminal-Bench 4.0. That is the entire case for the challenger in three numbers, and it is a strong one — with two caveats that decide whether it holds for your workload: the two scores were produced at different settings, and the 2× is a price per token rather than a price per finished task.

Kimi K3.1 Pro went live on 24 October 2026, a day after the teaser video that named it. Claude Opus 5.5 has been generally available since 22 September. Neither is a fresh lab experiment: Opus 5.5 is the current Opus-tier model Anthropic has been shipping since late September, and Kimi K3.1 Pro is a post-trained version of Kimi K3 rather than a new architecture. That makes this comparison unusually clean for a model-vs-model page — both rate cards are published, both vendors print a benchmark table, and neither model is hiding behind a waitlist.

The two rate cards, line by line

Five lines matter, and one of them breaks the pattern.

• Input — $2.00 per million tokens on Kimi K3.1 Pro against $4.00 on Claude Opus 5.5. Exactly half.

• Output — $10.00 against $20.00. Exactly half, and output is the line that dominates agentic and reasoning work, where the model writes far more than it reads.

• Cache read — $0.20 on both. This is the anomaly: the two vendors agree to the cent on the price of reading a cached prefix, which is the line that dominates a long-running agent re-reading a repository every turn.

• Cache write — Kimi charges $2.00 for a five-minute cache entry and $4.00 for an hour; Anthropic charges $5.00 and $8.00. Kimi is cheaper here too, and by more than half.

• Fast variants — Kimi K3.1 Pro Highspeed is $4.00 and $20.00, roughly two to three times the output speed. Anthropic's fast mode on Opus 5.5 is $8.00 and $40.00 at up to 2.5×. Same relationship, same doubling.

The arithmetic on a mixed request is therefore trivial. A call that reads a million input tokens and writes 200,000 output tokens costs $4.00 on Kimi K3.1 Pro and $8.00 on Opus 5.5. An agent that re-reads a million-token context with a 70% cache hit rate pays $0.74 for input on Kimi against $1.34 on Opus, or $2.74 against $5.34 once the output is included. The saving comes from fresh input and output, not from caching, because caching costs the same on both.

What the rate card does not tell you is how many tokens each model spends. Kimi K3.1 Pro defaults to its highest reasoning setting, max. Claude Opus 5.5 defaults to medium. A model that thinks at max by default will, on average, write more reasoning tokens per answer than one that thinks at medium — so the advertised 2× is a price per token, not a price per finished task, and the real gap depends on how much of Kimi's extra thinking your work actually needs. Moonshot does not publish a tokens-per-task figure that would settle it.

The one benchmark both labs print, and why it is not a measurement

Terminal-Bench 4.0 is the closest thing to a shared yardstick in these two launches, and even here the two numbers are not produced the same way.

Anthropic's 66.4% was run at xhigh effort, one level above the model's default of medium, on Anthropic's own harness, with its safety routing active. The company publishes the table and says so. Kimi's 65.7% appears in Moonshot's launch material at the model's default max setting, on a harness the company also controls; the launch table cites Opus 5.5 at 66.4%, the same number Anthropic published, without saying whether Moonshot re-ran it or quoted it. A 0.7-point gap between two vendor-run tables, at different effort settings, on different scaffolding, is not a measurement of the difference between the models. It is a statement that the two are close enough for harness choices to decide the order.

Two more data points point the same way. The independent evaluation of Opus 5.5 on the same benchmark, from Artificial Analysis as reported at its launch, is 59.6% — seven points below Anthropic's own number, in a configuration that includes server-side safeguard routing. And Moonshot's own documentation was still carrying literal "[xx.x%]" placeholders for its Terminal-Bench and science-benchmark scores on launch day, while the marketing post carried the numbers. Neither fact proves anything about relative quality. Both show that launch-day benchmark tables are marketing artifacts first.

Where the vendors do agree, the story is consistent: Kimi K3.1 Pro is close to Opus 5.5 on tasks both measure, and ahead of its own predecessor by a wide margin. Moonshot's long-horizon coding suite moved from 55.95% to 61.2%, its no-tools CAD score is 72.8 against Opus 73.0, and its hallucination rate on the vendor's own hard set fell from 23.4% to 8.3% with tools — a number that, if it survives outside evaluation, matters more to document-heavy work than any benchmark point.

The effort dial is the real difference between the two products

Both models expose a reasoning-effort field; they disagree about how many stops it has and where it rests.

Kimi K3.1 Pro has three: low, high and max, with max as the default. Claude Opus 5.5 has five: low, medium, high, xhigh and max, with medium as the default. Neither model lets the caller turn thinking off. Kimi fixes sampling outright — temperature 1.0, top-p 0.95, n at 1 — and rejects other values, so effort is the only behavioural dial it offers. Anthropic leaves temperature settable but keeps adaptive thinking always on, so effort is the only dial that changes how much the model thinks.

The defaults are doing real work. Kimi ships honest but expensive: every request that does not pass reasoning_effort pays for the deepest setting. Opus ships at the middle rung and reserves the top two for callers who ask. If you migrate a workload and change nothing else, you have changed both the model and its thinking budget, which is the single most common way a "cheaper model" turns out not to be. Anthropic's own table shows the size of that effect within one model: CursorBench 4.0 scores 57.8% at xhigh and 52.5% at medium, while FrontierCode moves by 0.2 points between the same two settings. Effort sensitivity is per workload, and no model card predicts yours.

One operational trap applies to both. Switching effort mid-conversation invalidates prefix-cache hits, so an agent that raises effort halfway through a long session pays to re-read everything it had cached. Decide the level before the session starts, not during it.

What each vendor does to the model when you are not looking

Both labs have now shipped a model whose behaviour on some inputs is not the behaviour you asked for, and both describe it openly.

Anthropic's Opus 5.5 launch documentation says most cyber-related tasks may be rerouted to Claude Opus 4.8, and that biology and frontier-model-development work may fall back to Opus 5. Those fallbacks happen server-side: the response comes back from a different model, and the caller's only signal is the content. For a security-adjacent workload this is a real capability change, not a policy footnote.

Moonshot took a different route. It removed cyber-related data from both mid-training and the reinforcement-learning stage while keeping general coding, and reported the cost of doing so: its adversarial coding score fell from 90.5% on the unmitigated checkpoint to 81.0% on the model it shipped. The company frames that as a deliberate cap on offensive cyber capability and says it will raise the cap gradually. Kimi K3.1 Pro does not reroute your request; it is simply weaker at that one category, by a margin its own vendor published.

Which is better depends on what you are building. If you need deterministic output from the model you selected, a silent fallback is worse than a disclosed weakness. If you need the strongest possible result on a task the vendor has gated, the fallback may still be better than a refusal — but you are now benchmarking two models while thinking you are using one.

The date that decides the open-weight question

Kimi K3.1 Pro is being sold as an open-weight model, and the weights are not out yet. Moonshot says the full weights will be released by 4 November 2026. Until then, the open-weight column of any comparison is a promise, and the Hugging Face organisation that hosts Kimi K3 has nothing newer than K3 as of 24 October.

Claude Opus 5.5 has no open-weight plan and never will. For anyone whose deployment cannot call a hosted API, the comparison ends there regardless of price: one side is a product you may be able to download in eleven days, the other is a rental. For everyone else, 4 November is the date the price comparison becomes a real architectural choice rather than a procurement one.

What to run today

Here is the part the rate cards do not cover: availability. Claude Opus 5.5 is callable through OrcaRouter today at $4.00 and $20.00 per million tokens with cache reads at $0.20 — the provider's list price passed through with no markup added — on the same key as the rest of your stack, with a median time to first token of 3.86 seconds over the last seven days on our route. Kimi K3.1 Pro joined the same catalogue on 24 October at $2.00 and $10.00, cache reads at $0.20, a 1,048,576-token window and a published output ceiling of 943,718 tokens. Both now answer on one key.

Having both behind one key is the thing neither vendor document mentions. At a 2× price gap, the decision that matters more than picking a winner is a routing policy: send the hard fifth of your traffic to Opus 5.5, send the rest to Kimi K3.1 Pro, and let automatic failover cover a stall on either side. The pass-through model is what keeps that arithmetic honest — each side of the split bills at the provider's own published price with nothing added on top, so the saving you calculate is the saving you get, and a price change at either vendor is live here the same day.

Which brings the comparison back to where it started. If your work is dominated by the tasks where the two models measure within a point of each other, the price gap is the whole story and the answer is the cheaper model. If it is dominated by long-horizon coding, where Opus has three months of independent evaluation behind it and Kimi has a vendor table published yesterday, the 0.7 points are not the number to trust — the number to trust is your own evaluation, run at the effort setting you will actually deploy. Both vendors have now made that easy to do, because both let you turn the dial and both publish what the dial costs.