Title card for MAI Code 2 vs Kimi K3.1 Pro. MAI Code 2 is shown as Microsoft's GitHub Copilot model at $0.50 input and $2.50 output per million tokens with a 1M context window, no parameter count and no published score. Kimi K3.1 Pro is shown at $2.00 and $10.00 per million tokens with a 1,048,576-token context window, an Artificial Analysis index of 46 and weights released on 4 November 2026.
Guides & Insights

MAI Code 2 vs Kimi K3.1 Pro: One Publishes a Number, the Other Publishes a Discount

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

MAI Code 2 went on sale on 17 November 2026 with a price and no score. Kimi K3.1 Pro, which Moonshot AI announced on 24 October 2026 and opened up on 4 November, arrives with both: $2.00 per million input tokens and an independent Intelligence Index reading of 46. That asymmetry, not a benchmark table, is what this comparison is about. Microsoft's launch post names four coding benchmarks — DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1 — and then publishes ratios instead of results: 96% of Claude Opus 5.5's performance at 5% of the cost, 98% of GPT-6.1 Sol's at 40%, averaged across the four. Moonshot publishes an actual percentage on an actual benchmark, and Artificial Analysis runs its own harness on top of that. One side hands you a discount you can verify to the cent and a capability claim you cannot check at all. The other hands you a number you can argue with.

Both are hosted. Only one is downloadable. That turns out to matter more than the four-fold price gap.

What Microsoft published, and what it left out

The launch post is specific about method and silent about results, which is an unusual pairing and worth reading closely. The four benchmarks are named. Three of them — DeepSWE 1.1, Terminal Bench 2.1 and SWE-Atlas — were run on MSBench, Microsoft's internal evaluation platform, driven through the GitHub Copilot CLI with subagents disabled and every model at maximum reasoning effort. FrontierCode's scores and costs come from Artificial Analysis, which makes it the only component of the average Microsoft did not produce itself. Costs are quoted at the model's 50% promotional price.

Then the results: two ratios and a preference share. 96% of Claude Opus 5.5's performance at 5% of the cost. 98% of GPT-6.1 Sol's at 40%. In blind frontend evaluations, reviewers preferred MAI Code 2's output over GPT-6 Sol 64% of the time. Nowhere on the launch post or the model page is there a single pass rate for the model on any of those four benchmarks, a parameter count, or an output ceiling. As of 18 November 2026 there is no Artificial Analysis row for it either, so nothing independent exists to fill the gap.

That is not a scandal. It is a choice with a specific cost for the buyer: you cannot compute what a finished task costs, because you do not know how many attempts the model needs. A ratio averaged over four benchmarks measures the average, not the distribution, and coding work is where the distribution has a long tail.

The rate card, which is fully public and fully checkable

Pricing is where the MAI Code 2 story becomes concrete, because GitHub documents it to the tenth of a cent.

• Input — MAI Code 2 at $0.50 per million tokens against Kimi K3.1 Pro at $2.00; MAI is a quarter of the price

• Output — $2.50 against $10.00 per million; the same four-fold ratio, so the discount is not a mix artefact

• Cached input — $0.025 against $0.20 per million; MAI is an eighth, and GitHub publishes no cache-write charge on either side

• From 17 May 2027 — MAI Code 2 reprices to $1.00 / $0.05 / $5.00, halving its advantage to 2× on fresh tokens and 4× on cache reads

• Context — 1,048,576 tokens on Kimi K3.1 Pro through Moonshot's API, against a 1 million token window on MAI Code 2 that opens only inside Visual Studio Code and Copilot CLI

• Output ceiling — not published by Microsoft, against no separate output ceiling in Moonshot's published card either; the 1,048,576 figure is a context window, not a generation limit

• Weights — none, hosted only, against a 1.56 TB MXFP4 download published on 4 November 2026

The cached-input line is where an agentic workload actually spends its money, and it cuts in Microsoft's favour harder than the headline does. A coding agent replays a fixed system prompt, a stable tool schema and a repository prefix on every turn. At $0.025 per million cached tokens against $0.20, the gap on the lines that dominate the bill is wider than 4×. The expiry date then does most of the damage: from 17 May 2027 the cached rate becomes $0.05 while Moonshot's stays where Moonshot puts it, and the advantage narrows to 4× on caching and 2× on fresh tokens.

The one instrument the other side does publish

Kimi K3.1 Pro's evidence has a different shape — a vendor number you can check against an independent one, and a per-task cost that includes the time the model spends thinking.

Moonshot's own launch table puts Kimi K3.1 Pro at 65.7% on Terminal-Bench 4.0, second behind Claude Opus 5.5 at 66.4%. That is a real percentage on a current benchmark, published by the vendor on launch day, and it does not flatter the model. On the Intelligence Index, version 5.0.1, Kimi K3.1 Pro scores 46, well above the median of 23 that Artificial Analysis reports for comparable models.

The cost side is where an independent reading earns its keep, and it is not flattering either. Artificial Analysis puts Kimi K3.1 Pro at $3.68 per Intelligence Index task, running 38.4 minutes per task at 31 tokens per second of output and 157.8 seconds to the first answer token — slow against a median of 13.6 minutes per task. Those are the numbers you need to decide whether a model belongs in a loop rather than a chat window, and no equivalent exists for MAI Code 2.

Artificial Analysis page for Kimi K3.1 Pro showing an Intelligence Index of 46 against a median of 23, $2.00 per million input tokens and $10.00 output, a $3.68 cost per Intelligence Index task, 38.4 minutes per task, 31 tokens per second, 157.77 seconds to first token, a 1M token context window and a 90 percent cache discount.

What the four-fold discount is actually buying

Strip the marketing out and the trade is legible. MAI Code 2 costs a quarter of Kimi K3.1 Pro today and half of it after May 2027. In exchange you accept four things you cannot inspect: the pass rate, the attempts per finished task, the model's size, and its behaviour outside the four benchmarks Microsoft chose. Microsoft is unusually direct about the last one — the launch post says the models are "explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work." That sentence should be believed more readily than the 96%, because it is a limitation the vendor volunteered.

Kimi K3.1 Pro's trade runs the other way. You pay four times as much per token, wait 38 minutes for a task, and get a model you can download, fine-tune inside your own tenancy and keep running if Moonshot's API has a bad week. The weights are 1.56 TB across 96 shards in MXFP4, which is a real serving job rather than a laptop download, and the licence is a custom instrument with a revenue threshold — not Apache, not MIT, not OSI-approved. So "open" here means auditable and self-hostable, not unencumbered. For a security review that distinction is fine. For a product you intend to redistribute, read the licence first.

Neither model is a bad buy. They are priced for different buyers: one for a team that already pays for Copilot and wants cheaper tokens inside tooling it has adopted, one for a team that wants an asset it can move.

Calling both sides, and the one with no endpoint

Kimi K3.1 Pro is on OrcaRouter as kimi/kimi-k3.1-pro at $2.00 and $10.00 per million tokens, passed through at the provider's list rate with 0% markup — so if Moonshot changes its card, your bill changes the same day rather than after a reseller repricing cycle. It sits behind the same key as the rest of the catalogue, and automatic failover means a provider-side rate limit becomes a retry rather than an outage in the middle of a 38-minute agent run.

MAI Code 2 is not on our catalogue and cannot be. The only published route is a GitHub Copilot subscription, across Pro, Pro+, Max, Business and Enterprise; the smaller MAI Code 2 Flash is the only model in the family with an API surface, a limited-access preview in Microsoft Foundry through an OpenAI-compatible Responses API. If you want to compare the two on your own repositories, one side of the experiment has to run inside a Copilot client and the other against an HTTP endpoint. That is a real constraint on the design, not a footnote.

The arrangement that survives it is a split rather than a migration. Keep the measurable model on the path that has to be budgeted, where Kimi K3.1 Pro's $3.68 per task and 38-minute completion times are known quantities you can plan around. Put MAI Code 2 in front of the developers already working in Copilot, where trying it costs nothing beyond a seat they already hold. Then decide on finished-task cost from your own repositories rather than from a four-benchmark average.

Microsoft's MAI Models page announcing MAI Code 2 and Code 2 Flash under the heading 'Everyday coding at scalable prices', with the post tagged Nov 17, 2026 and Ignite.

Which one to pick

• You need to audit the model or run it yourself — Kimi K3.1 Pro, and there is no second candidate here. The weights are downloadable and the licence question is answerable.

• Your developers already live in Copilot and the work is routine repair — MAI Code 2 at $0.50 and $2.50, on the understanding that the 96% and 98% figures are Microsoft's own and unreproduced.

• You need a defensible number for procurement — Kimi K3.1 Pro, because an independent index reading of 46 exists and the other model has no row.

• You are budgeting past May 2027 — Kimi K3.1 Pro, or MAI Code 2 modelled at the doubled rate card rather than the promotional one.

• You want the cheapest token for a human-driven workflow — MAI Code 2, since a person waiting on a completion is not paying for wall-clock time the way an agent loop does.

Scoreboard comparing MAI Code 2 and Kimi K3.1 Pro across six dimensions. Input $0.50 versus $2.00 per million tokens, output $2.50 versus $10.00, cached input $0.025 versus $0.20, context 1M tokens versus 1,048,576 tokens, published score none versus an Artificial Analysis index of 46, and weights not distributed versus 1.56 TB available since 4 November 2026.

The open question is whether Microsoft ever publishes an absolute number. One Terminal-Bench 4.0 pass rate would settle more than any ratio, because it would put MAI Code 2 on the same axis as everything else. Until that exists, this is a comparison between a price list and a measured model — and the honest recommendation is the cheaper token for work you can watch, and the measured model for work you cannot.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily