Generated scoreboard card showing MAI Code 2 at $0.50 and $2.50 per million tokens with no independent index row, against GPT-6.1 Sol at $2.00 and $10.00 with an Artificial Analysis Intelligence Index of 55 and a deprecated status note.
Guides & Insights

MAI Code 2 vs GPT-6.1 Sol: Microsoft Benchmarked Itself Against a Deprecated Model

Author

Gideon Frost

Date Published

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

The most quoted line in Microsoft's Ignite announcement is that MAI Code 2 delivers 98% of GPT-6.1 Sol's coding performance at 40% of the cost. It is a good line. It also names a model that Artificial Analysis has already marked deprecated, superseded inside OpenAI's own line by GPT-6.2 Sol, and that is the detail that reframes the whole comparison: Microsoft picked the reference point, and the reference point is no longer the one to beat. MAI Code 2 shipped on 17 November 2026 with a promotional rate of $0.50 per million input tokens and $2.50 per million output. GPT-6.1 Sol costs $2.00 and $10.00 and posts 55 on the independent index. Neither of those facts makes the 98% claim false. They make it a comparison against a model on its way out, which is a different thing from a comparison against the frontier.

The rate cards and the deprecation notice

• Price per million tokens — GPT-6.1 Sol charges $2.00 in and $10.00 out with no expiry; MAI Code 2 charges $0.50 and $2.50 until 17 May 2027, then $1.00 and $5.00

• Cached input — $0.10 against $0.025, a 4× gap that survives both price changes

• Long prompts — no published tier on MAI Code 2; GPT-6.1 Sol reprices everything above 272,000 input tokens at 2× input and 1.5× output, so a 300K-token request costs more per token than its headline rate suggests

• Context — 1,050,000 tokens on both, with 128,000 maximum output on both

• Input modality — GPT-6.1 Sol reads files as well as text and images; MAI Code 2's model card claims text and images and stops there

• Lifecycle status — MAI Code 2 brand new; GPT-6.1 Sol marked deprecated by Artificial Analysis, which now benchmarks it every 18 hours instead of every 6

That last line is the one a buyer should weigh hardest. A deprecated model is not discontinued — it still answers and it still bills at list — but its evaluation cadence has slowed and its vendor has moved on. Building a cost model around a model in that state means building it around a number that will not be updated.

What 98% is measured on, and why the reference point matters

Card breaking down Microsoft's 98% claim into its four benchmarks, noting that three ran on the internal MSBench harness and only FrontierCode came from Artificial Analysis, with the MAI Code 2 pass rates and the deprecated status of GPT-6.1 Sol.

Strip the framing and 98% is an average of four pass rates, three of which Microsoft generated itself. DeepSWE 1.1 and Terminal Bench 2.1 and SWE Atlas went through MSBench, Microsoft's internal platform, via the GitHub Copilot CLI at 1.0.102, subagents off, every model at max effort. FrontierCode's scores and costs came from Artificial Analysis. What Microsoft does not publish is the split: the card gives MAI Code 2's own numbers — 73.2%, 84.6%, 43.1% — and one average against GPT-6.1 Sol, with no line showing which benchmark carried the average and which dragged it.

An average hides the shape of a result, and the shape matters. SWE Atlas, at 43.1%, is the weakest of the three public numbers, and it is also the one that tests codebase question answering, test writing and refactoring rather than issue resolution. FrontierCode, the only independently sourced component, is one of four. A model that is at parity on bug-fixing benchmarks and behind on refactoring can average to 98% and still be the wrong tool for a refactor.

Microsoft's other headline is a preference study rather than a pass rate: in blind side-by-side evaluation of front-end work across interfaces, games, data visualisations, 3D scenes and SVG graphics, 258 annotators preferred MAI Code 2's output over GPT-6 Sol's 64% of the time. Note the model named there — GPT-6 Sol, not 6.1 — and note that a preference for how generated UI looks is not a measure of whether it works.

What the independent record says about GPT-6.1 Sol

GPT-6.1 Sol has something MAI Code 2 does not: a third-party profile. Artificial Analysis scores it 55 on the Intelligence Index v5.0.1, the same nine-evaluation composite everyone else in this bracket is measured on, and records $1.27 per index task, 23.3 minutes per task, 145 seconds to first answer token and 49 tokens per second of output.

Screenshot of the Artificial Analysis page for GPT-6.1 Sol showing a deprecation notice recommending GPT-6.2 Sol, an Intelligence Index of 55, 145.15 seconds to first answer token, 49 output tokens per second and 23.3 minutes per task.

Inside OpenAI's own line the picture has moved. GPT-6.2 Sol scores 57 at $1.01 per task, so the model Microsoft benchmarked against has been both beaten and undercut by its own successor. That is not an argument against MAI Code 2 — a cheaper model beating an older, pricier one is exactly what a price-performance launch is supposed to do. It is an argument against treating the 98% as a description of the current frontier. The honest framing is that Microsoft matched a strong model from the previous generation at 40% of its price, and has not yet shown what it does against the one shipping now.

Where the tokens go, which is where the real difference is

Microsoft's 40% is a cost-per-task figure, not a price ratio. The price ratio is 25% — $0.50 against $2.00 input, $2.50 against $10.00 output. For the cost ratio to sit at 40% while the price ratio sits at 25%, MAI Code 2 has to be spending more tokens per task than GPT-6.1 Sol does, roughly 1.6 times as many by my arithmetic on the two published ratios. That is an inference, not a Microsoft claim, and blended input/output rates make it approximate.

Two things follow. First, MAI Code 2 is cheaper per finished task but not cheaper per token in the way the headline implies, so any workload priced per token rather than per task will see a smaller saving than 60%. Second, token-hungry behaviour interacts badly with the reasoning-effort ladder: both models expose effort levels, both were benchmarked at maximum effort, and a team running at max effort on either side is paying for thinking tokens twice over. Dropping to a lower effort setting on routine work is usually a bigger saving than switching models.

Access: one is an API, the other is a subscription feature

GPT-6.1 Sol is callable on OpenAI's API with a model ID, a documented snapshot policy, a knowledge cut-off of 30 April 2026 and support for tools, structured outputs, response format and seed. It is also in ChatGPT Work and Codex on Plus, Pro, Business, Enterprise and Edu plans.

MAI Code 2 is a Copilot feature, not an API. You reach it through a GitHub subscription — Pro through Enterprise — and whichever client you already use, from VS Code to Xcode. It is not in Microsoft Foundry; only MAI Code 2 Flash has a Foundry preview, through an OpenAI-compatible Responses API. There is no published endpoint for MAI Code 2 itself, no date for one, and no weight distribution under any licence.

For a team whose code lives in GitHub and whose developers already work in Copilot, that is a fine answer and arguably a better one than an API key. For a CI bot, an internal agent framework, a code-review pipeline or anything with its own orchestration, GPT-6.1 Sol is a model you can call and MAI Code 2 is not.

Which one to reach for

• Copilot-native developer work, cost-sensitive, high volume — MAI Code 2 at $0.50 and $2.50, and MAI Code 2 Flash at $0.05 and $0.25 for the routine tier, are the cheaper answer by a wide margin.

• Application, agent or CI integration — GPT-6.1 Sol, because it has an API. MAI Code 2 has a subscription.

• Anything above 272,000 tokens per request — GPT-6.1 Sol's long-prompt tier doubles the input rate, which erodes its price position further; MAI Code 2 publishes no equivalent tier, which is a gap in the documentation rather than a promise of no tier.

• Writing code in something other than English — the model card rules it out at release for MAI Code 2; OpenAI attaches no such restriction to GPT-6.1 Sol

• Long-horizon autonomous work — Microsoft says explicitly that MAI Code 2 is not tuned for very long-horizon agentic work and is not intended for unsupervised consequential actions. GPT-6.1 Sol's positioning is broader, though its own deprecation is a reason to look at GPT-6.2 Sol instead.

Pricing the comparison on one key

Screenshot of OrcaRouter's GPT-6.1 Sol model page showing the $2.00 input and $10.00 output per million token rates, a 1M-token context, 128K maximum output and text, image and file input.

GPT-6.1 Sol is on OrcaRouter at provider list price with 0% markup — $2.00 input and $10.00 output per million tokens — and so are Claude Opus 5.5 and Gemini 4 Carbon, all behind one key. GPT-6.2 Sol is not on the catalogue yet, so the successor question cannot be settled here; what can be done is to price 6.1 Sol's deprecation honestly and watch the per-task bill rather than assuming the newer model is the more expensive one. It is not — on Artificial Analysis, 6.2 Sol costs $1.01 per index task against 6.1's $1.27.

MAI Code 2 is not on our catalogue; the only published route is GitHub Copilot. If Microsoft opens a Foundry endpoint for the large model, the useful first experiment is not a switch but a split: MAI Code 2 on the everyday repair traffic where its price advantage is largest, a measured frontier model on the refactors and long-horizon work, and automatic failover underneath so an unproven model never becomes a single point of failure. The 40% is worth testing. The 98% is worth testing separately, on your own repositories, before anyone plans a migration around it.