TUE, NOVEMBER 17, 2026
Independent · In‑Depth · Practitioner‑Tested
Large Language Models

Qwen4-Max vs Kimi K3.1 Pro vs Claude Opus 5.5 (2026): The Coding Scores Disagree

Terminal-Bench 4.0 has two published scores for Claude Opus 5.5 - 59.6 on Alibaba's chart, 66.4 on Moonshot's. Moonshot reports its own model losing to Opus by 0.7 while Alibaba has it winning by 7.6. Prices run $1.50, $2 and $4 per million input, and Kimi's highspeed tier costs exactly what Opus does.

🕐 9 min read 👁 36 views 📅 Nov 18, 2026
Straight answer
● The finding that matters Terminal-Bench 4.0 has two different published scores for Claude Opus 5.5 - 59.6 on Alibaba's chart and 66.4 on Moonshot's. Which model "wins coding" depends entirely on whose chart you read.
● Moonshot reports its own model losing On Moonshot's figures Kimi K3.1 Pro scores 65.7 against Opus 5.5 at 66.4. Alibaba's chart has Kimi ahead by 7.6 points. Same benchmark, same two models.
● Prices, which are not what you would guess Qwen4-Max $1.50/$4.50 · Kimi K3.1 Pro $2/$10 · Opus 5.5 $4/$20. Kimi's highspeed tier is $4/$20 - identical to Opus.
● One is open weights Kimi K3.1 Pro, on Hugging Face. The other two are closed.

Three flagship models, three price points, and a benchmark table that does not agree with itself. This page works through where the numbers come from, because on coding the answer changes depending on which vendor published the chart - and that is more useful to know than any single score.

The specifications, which are not in dispute

Qwen4-MaxKimi K3.1 ProClaude Opus 5.5
Released16 November 2026API 23 October, weights 2 November 20262026
Input price / 1M$1.50$2.00$4.00
Output price / 1M$4.50$10.00$20.00
Cached input / 1M$0.075$0.20Not stated here
ArchitectureSparse MoE, 2.4T backbone, 71B activeMoE, 2.8T total, 104B active, 896 routed expertsNot published
Context window1M (max input 991K)1,048,5761M
Open weightsNo, closedYes, Kimi K3.1 Pro LicenseNo, closed
InputsText, image, videoText, image; video via Files APIText, image

Two things in that table are worth pausing on. Qwen4-Max at $1.50/$4.50 is genuinely the cheapest frontier model available. And Kimi K3.1 Pro is the only one of the three whose weights you can download - ungated on Hugging Face, under a licence that is MIT-style for almost everyone. Two conditions attach: a model-as-a-service business doing more than $20M over any twelve months needs a separate agreement with Moonshot, and a product with over 100M monthly actives or $20M monthly revenue has to display the Kimi name in its UI. Internal use is exempt from both.

Now the part that does not agree

Terminal-Bench 4.0 is the benchmark that matters most here, because it measures multi-step shell and filesystem work - what a coding agent actually does. Three sources, three different pictures:

SourceQwen4-MaxKimi K3.1 ProOpus 5.5
Alibaba's launch chart48.567.259.6
Moonshot's figures (Claude Code harness)not shown65.766.4
Artificial Analysis, independent48.4867not shown

Read the Opus 5.5 column. Alibaba puts it at 59.6; Moonshot puts it at 66.4. That is a 6.8-point spread on one benchmark for one model, published by two companies that both compete with it. On Alibaba's numbers Kimi beats Opus by 7.6 points. On Moonshot's own numbers, Kimi loses to Opus by 0.7 - Moonshot is reporting its own model in second place.

A vendor publishing its own model losing is a strong signal that the figure is honest. It also means the widely-repeated line that Kimi tops Terminal-Bench rests on a competitor's chart rather than its maker's. Artificial Analysis measured Kimi at 67 independently, which supports Alibaba's number for Kimi - but nobody independent has published an Opus 5.5 figure, so the gap is unresolved.

The lesson generalises past these three models: a benchmark number is only comparable to another number produced on the same harness, by the same evaluator, on the same day. Cross-vendor charts are marketing with arithmetic attached.

The scores that do agree

BenchmarkQwen4-MaxKimi K3.1 ProOpus 5.5
DeepSWE v1.175.775.574.2
ProgramBench31.531.565.0
Artificial Analysis Index43 (9th)46 (top open-weights)58

DeepSWE separates the three by 1.5 points across the board, which is noise - nobody should choose on it. ProgramBench does the opposite: Opus 5.5 at 65.0 against 31.5 for both challengers is more than double, and it is the widest gap anywhere in this comparison.

Qwen4-Max and Kimi K3.1 Pro both scoring exactly 31.5 is worth noting. It may be coincidence; it may mean the two were run through the same harness configuration. We cannot tell from the published material, and neither vendor comments on it.

The Artificial Analysis Intelligence Index is the cleanest row in this whole page, because it is one evaluator scoring all three. Opus 5.5 at 58 against 46 and 43 is a clear, consistently-measured lead. Note the index version matters - Kimi's launch-day score of 55 was on an earlier version and is not comparable to 46 on v5.

Price per unit of capability, which inverts the obvious answer

The assumption going in is that the open-weights Chinese model is the budget option and Opus is the premium one. The pricing does not work that way. Kimi K3.1 Pro has two API tiers, and the highspeed tier is $4 input and $20 output - exactly Opus 5.5's price. At equal latency, the open model costs the same as the closed one.

So the real choice is three-way rather than cheap-versus-expensive. Qwen4-Max is materially cheaper than both and measurably weaker at coding. Kimi at its standard tier is half Opus's price and within a point of it on Terminal-Bench by its own reckoning. Opus wins the only independently-scored comparison and wins ProgramBench outright.

Which one you should actually use

You are driving a coding agent and budget is not the constraint → Claude Opus 5.5. It wins the one row scored by a neutral party, wins ProgramBench by more than double, and is the only one with a published first-party price from the vendor's own page.
You want the best coding result per dollar → Kimi K3.1 Pro on the standard tier. Half the price of Opus, within a point on its maker's own Terminal-Bench figure, and $0.20 cached input is the cheapest cache on the board.
You need to self-host, air-gap, or avoid vendor lock-in → Kimi K3.1 Pro is the only option here. Ungated weights on Hugging Face, and the licence conditions only bite above $20M revenue.
You are running long-horizon agents rather than writing code → Qwen4-Max. It leads Agents' Last Exam at 46.7 and the HLAB legal-agent benchmark 21.3 against 9.2 for Opus - and at $1.50/$4.50 the economics of a long-running agent change completely. Full Qwen4-Max breakdown.
You are optimising purely for input cost at scale → Qwen4-Max at $1.50, or $0.075 cached - less than half Kimi's cached rate and the lowest figure in this comparison.
You want the tools rather than the raw models → Our Codex versus Claude Code comparison covers what most people actually reach for, which is a harness rather than an API.

What we would want before trusting any of this further

An independent Terminal-Bench 4.0 run on Opus 5.5. It is the single missing number, it is the one both vendors disagree about, and until someone neutral publishes it the coding ranking between Kimi and Opus is genuinely unresolved rather than merely contested. Artificial Analysis has run Kimi and Qwen4-Max; the Opus figure would settle it.

Everything above was read off vendor and evaluator pages on 17 November 2026. Model pricing in this category has moved twice in the last quarter, so check before you commit a budget - and if you find a figure here that has changed, the correction is free and does not require you to be a customer.

FAQ

Which is best for coding: Qwen4-Max, Kimi K3.1 Pro or Claude Opus 5.5?
On the only benchmark scored by a neutral evaluator - the Artificial Analysis Intelligence Index - Opus 5.5 leads at 58 against 46 for Kimi and 43 for Qwen4-Max. On Terminal-Bench the answer depends on whose chart you read: Alibaba has Kimi beating Opus by 7.6 points, Moonshot has Kimi losing to Opus by 0.7.
Why do the Terminal-Bench scores differ between sources?
Alibaba's launch chart reports Opus 5.5 at 59.6; Moonshot reports 66.4 on the Claude Code harness. Different harnesses, different runs, both self-published. No independent Opus 5.5 figure has been released.
Is Kimi K3.1 the same as Kimi K3.1 Pro?
Yes. K3.1 Pro is the first and only model in the K3.1 series. The API offers kimi-k3.1-pro and an API-only kimi-k3.1-pro-highspeed variant.
How much do they cost?
Per million tokens: Qwen4-Max $1.50 in / $4.50 out, Kimi K3.1 Pro $2 / $10, Claude Opus 5.5 $4 / $20. Kimi's highspeed tier is $4 / $20, the same as Opus.
Which has open weights?
Only Kimi K3.1 Pro, ungated on Hugging Face under the Kimi K3.1 Pro License. It is MIT-style unless you run a model-as-a-service business over $20M in twelve months, or a product above 100M monthly users or $20M monthly revenue, which triggers an agreement or an attribution requirement.
Is Qwen4-Max worth it at $1.50?
For long-horizon agent work, yes - it leads Agents' Last Exam and the HLAB legal-agent benchmark. For coding, it scores 48.5 on Terminal-Bench and 31.5 on ProgramBench against 65.0 for Opus, so the saving comes with a real capability cost.
What context window does each have?
All three are around one million tokens: Qwen4-Max 1M with a 991K maximum input, Kimi K3.1 Pro 1,048,576, and Opus 5.5 1M.
⚖ Our Verdict

Opus 5.5 for coding when budget allows - it leads the only independently scored index at 58 and wins ProgramBench 65.0 to 31.5. Kimi K3.1 Pro for the best result per dollar and the only open weights. Qwen4-Max for long-horizon agents at $1.50, not for code. No winner declared on Terminal-Bench, because the two published figures for Opus differ by 6.8 points and no neutral run exists.