How to Read Microsoft's MAI Code 2 Cost Claims and What Licensed Weights Mean
Microsoft launched MAI Code 2, MAI Code 2 Flash and a licensed-weights programme at Ignite. The headline numbers are averages, measured on promotional prices. Here is how to read them, and what a third way to distribute model weights means for teams.
- Reading time
- 16 min
- Sections
- 7
- Code examples
- 1
- Tables
- 3
- Sources
- 10
- Published
- 20 Nov 2026
- What changed
- Microsoft announces MAI Code 2 and MAI Code 2 Flash
- Why it matters
- For most developers, MAI Code 2 is one more option in the Copilot model picker, which the changelog lists in VS Code, Visual Studio, JetBrains, Xcode, the Copilot CLI and more.
- What to do
- Build a small private test set
On 17 November 2026, at its Ignite conference, Microsoft AI announced two in-house coding models, MAI Code 2 and MAI Code 2 Flash, and a new way to get them: MAI Licensed Weights. Microsoft says MAI Code 2 delivers 98% of the performance of OpenAI's GPT-6.1 Sol at 40% of the cost. It also says the model reaches 96% of Claude Opus 5.5's performance at 5% of the cost. Both models are already in GitHub Copilot.
These are company claims. Each is an average across four benchmarks Microsoft chose, at maximum reasoning effort, with MAI's cost counted at a 50% promotional price. That does not make them false, but you should know how such a number is built before using it in a budget.
Below: how to take apart a "percent of performance at percent of cost" claim, why a cheap tier like Flash matters for high-volume coding, and what licensed weights are.
What happened
Microsoft announces MAI Code 2 and MAI Code 2 Flash (17 November 2026)
Microsoft AI describes MAI Code 2 as built for everyday enterprise coding and MAI Code 2 Flash as built for high-volume work. Its claims, all as stated by Microsoft:
- MAI Code 2 reaches 98% of GPT-6.1 Sol's performance at 40% of the cost, and 96% of Claude Opus 5.5's performance at 5% of the cost. These figures are averaged over DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1.
- MAI Code 2 Flash performs comparably to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their cost, across three coding benchmarks.
- In blind frontend evaluations, developers preferred MAI Code 2's results over GPT-6 Sol's 64% of the time, based on 258 annotators.
Microsoft's footnotes say all models ran at max reasoning effort. Tests ran on MSBench, Microsoft's internal evaluation platform, using the GitHub Copilot CLI without subagents. FrontierCode scores and costs come from Artificial Analysis. Both MAI models' costs are shown at their 50% promotional price.
Microsoft also states a limit in plain words: the models are "explicitly not yet tuned for harder problems, like fundamental science research and very long-horizon agentic work."
Per the GitHub changelog, MAI Code 2 is on Copilot Pro, Pro+, Max, Business and Enterprise plans, and Flash is also on Free and Student plans, under usage-based billing. Microsoft says Flash is in limited-access preview in Microsoft Foundry, and that OpenRouter, Vercel, Hermes, OpenCode and Kilo are coming soon.
The Next Web reports that the promotional rate runs until 17 May 2027, citing GitHub's pricing page. It reports that MAI Code 2's price then rises from $0.50 to $1 per million input tokens and from $2.50 to $5 per million output tokens. When we fetched GitHub's public models-and-pricing page, it did not yet list MAI Code 2, so treat those figures as reported.
Microsoft launches MAI Licensed Weights (17 November 2026)
In a companion post, Microsoft describes MAI Licensed Weights as a way for organisations to deploy and customise Microsoft AI models in their own Azure environments. Microsoft says customers control how models are deployed, who can access them and when they are upgraded. It says "Your data and adaptations remain yours," and that it uses customer data to train its models only with the customer's express agreement.
The programme is not broadly available yet. Microsoft says it will launch through "Frontier Company engagements," and tells interested organisations to contact their account team. According to The Next Web, Microsoft AI's Suleyman told Axios the approach would "feel like open source" with more security and control.
Microsoft cites early customer results, all from customer-specific evaluations:
- AT&T used customised MAI Code 2 Flash and Code 2 for 5G network analysis. Microsoft says it beat GPT-6.1 Sol on standard tasks at about 1/50th of the cost.
- Woven by Toyota fine-tuned MAI Code 2 Flash on historical MISRA coding-standard fixes. In 44% of cases where its existing frontier models needed human intervention, Microsoft says the model produced ready-to-merge fixes at under one-tenth of the cost.
- Alaska Airlines raised the accuracy of aircraft maintenance report classification from 81.2% to 91.1% after tuning. Qualified employees keep final decisions.
How to read "98% of the performance at 40% of the cost"
Think of a car advert promising "98% of the speed of a sports car at 40% of the price." You would ask which roads were tested and whether the price is a launch discount. The same questions apply here.
Half one: the performance ratio is a ratio of averages
Microsoft averages scores across four benchmarks and then compares the averages. That is a ratio of averages. It is not the same as asking "how close is MAI on each benchmark?"
Here is a made-up example to show the difference. Suppose benchmark A is easy and both models score 80. Benchmark B is hard; model X scores 20 and model Y scores 14.
| Measure | Model X | Model Y | Y as share of X |
|---|---|---|---|
| Benchmark A (easy) | 80 | 80 | 100% |
| Benchmark B (hard) | 20 | 14 | 70% |
| Average score | 50 | 47 | 94% |
| Average of per-benchmark ratios | — | — | 85% |
The ratio of averages says 94%; the average of ratios says 85%. The easy benchmark's bigger numbers dominate, so a model can look close overall while trailing badly on the hardest test.
Three more things hide inside an average:
- The benchmark mix. Microsoft chose the four tests. They differ a lot. DeepSWE 1.1, from Datacurve, has 113 tasks written from scratch across 91 repositories; Epoch AI reviewed it in September 2026, found issues in at least 23 tasks and rated it "Flawed." SWE Atlas, from Scale, measures codebase Q&A, test writing and refactoring across 284 tasks; Scale says top systems score in the 40s. FrontierCode, from Cognition, asks whether a change is good enough for a maintainer to merge, and scores any attempt that fails a blocking check as zero.
- The version. Microsoft reports Terminal Bench 2.1, while the Terminal-Bench site, hosted by Stanford, Harbor and the Laude Institute, now headlines version 4.1. Scores across versions are not comparable.
- The spread. The headline claim states a single average with no error bars, while Terminal-Bench reports 95% confidence intervals. With about 100 tasks, a two-point gap can be noise.
If you want to go deeper on building your own tests, see evals. For a reminder that benchmark gains do not always become developer speed, see METR's developer slowdown study.
Half two: cost per task is not price per token
Price per token is what the vendor charges for each unit of text. Cost per task is what you pay to finish one job. The second equals the first times the number of tokens the model uses, including hidden reasoning tokens. A cheap model that rambles can cost more per task than a pricey model that answers briefly, the way a cheap taxi that takes the long route can cost more than a dear one that knows the shortcut.
Microsoft's claim is about cost, which suggests cost per benchmark run. Artificial Analysis, which supplies the FrontierCode numbers, reports cost in US dollars per task, split by input, cache hit, cache write, reasoning and answer tokens. On its FrontierCode results, per-task costs ranged from $0.09 to $12.00 across models.
We can do a rough check with list prices. GitHub's pricing page lists GPT-6.1 Sol at $2 per million input tokens and $10 per million output tokens, and Claude Opus 5.5 at $4 and $20. Using The Next Web's reported MAI Code 2 promo price of $0.50 and $2.50:
| Comparison | MAI price per token as share of rival | Microsoft's claimed cost share | What the gap implies |
|---|---|---|---|
| vs GPT-6.1 Sol (promo price) | 25% | 40% | MAI used roughly 1.6 times as many tokens per task |
| vs Claude Opus 5.5 (promo price) | 12.5% | 5% | Opus used roughly 2.5 times as many tokens per task |
| vs GPT-6.1 Sol (after promo, $1 and $5) | 50% | not stated | about 80% if token use stays the same |
| vs Claude Opus 5.5 (after promo) | 25% | not stated | about 10% if token use stays the same |
This is rough arithmetic, not Microsoft's method; it ignores cache discounts and long-context tiers. Still, it shows that token efficiency moves the result as much as the price tag, and that the "40%" leans on the promotion.
The Flash claim makes the same point. On GitHub's page, GPT-6 Luna and Claude Haiku 5.5 (default tier) have the same list price: $0.10 per million input tokens and $0.50 per million output. Yet Microsoft says Flash costs 65% of one and 20% of the other. With equal prices, that difference can only come from how many tokens each rival used, or from Haiku's higher long-context tier above 100K tokens. We could not find Flash's own per-token price on the pages we fetched.
Half three: max reasoning effort is a setting, not a constant
Reasoning models let you choose how long they think. Microsoft ran every model at max effort, a fair like-for-like rule. But many teams run at medium or low effort, and each model loses quality at a different rate as its thinking budget shrinks. Test at the effort level you will actually pay for.
Small tiers and routing between them
Renaming a variable or fixing a lint error is high-volume and low-risk. Designing a migration across a large codebase is rare and risky. Paying top-tier prices for both is like sending every parcel by overnight courier.
That is the job of a flash tier. Microsoft also says it built MAI Work Flash, a model optimised from Code 2 Flash, for Microsoft 365 and FabricIQ, and reports that in Word it used about 42% fewer input tokens per call in production than its comparison point. The general idea is covered in small language models.
LLM routing sends each request to the cheapest model likely to handle it. A simple coding router might work like this:
- Classify the task. Use cheap signals: files touched, diff size, whether tests exist, whether the task is a rewrite or a lookup.
- Try the small tier first for tasks in the low-risk classes.
- Check the result automatically. Run tests, linters and type checks.
- Escalate on failure to the larger tier, passing along what failed.
- Log cost per completed task, not per call.
The catch: escalation costs twice, once for the failed small attempt and once for the retry. Routing pays off only when the small tier succeeds often enough. Caching shared context also matters for agent loops, as covered in context price and prefix caching for coding agents. For a related look at tiered pricing among OpenAI models, see GPT-6.2 Astra and Sol pricing and routing. Set ceilings with cost and latency budgets so a runaway agent loop cannot quietly erase the gains.
Continue learning
How this article connects
11 pages link to or from it. Drag to turn, click a dot to open it.
Licensed weights: a third way to get a model
Most teams have chosen between two options.
- Closed API. You call the vendor's servers. The vendor holds the weights and controls updates.
- Open weights. You download the weights and run them anywhere, carrying hardware, serving and security yourself. Our guide to self-hosted open-weight coding agents covers that path.
Licensed weights sit in between: per Microsoft, the customer licenses a MAI model, customises it and runs it in its own Azure environment. Think of leasing a car you may modify and keep in your own garage, as long as the garage is on the dealer's estate.
| Question | Closed API | MAI Licensed Weights (as described) | Open weights |
|---|---|---|---|
| Who holds the weights? | Vendor | Customer's Azure environment, under licence | Customer, anywhere |
| Where does it run? | Vendor cloud | Customer's Azure tenancy | Any cloud or on-premises |
| Who decides upgrades? | Vendor | Customer, per Microsoft | Customer |
| Customisation | Vendor tuning tools | Customer data, with Microsoft support | Full, with your own tools |
| Who can get it? | Anyone who signs up | Through Frontier Company engagements for now | Anyone who accepts the licence |
| Run outside Azure? | No | Not described | Yes |
What it means for fine-tuning
Microsoft's Custom Models page says customers can "customize one or more top MAI models on your business' data, trained safely inside your own secure environment." The only method it names is fine-tuning. If you are new to it, start with fine-tuning and LoRA, which trains small adapter weights instead of the whole model.
The customer examples share a pattern: narrow, repetitive tasks with a long history of correct answers. That is where a tuned small model often beats a general large one. It is not proof that a tuned Flash model will win on open-ended work.
What it means for data control
Training data and the tuned model stay in your environment, and Microsoft trains on customer data only with express agreement. That can shorten security reviews. But "your own Azure environment" still means one cloud and one licence. Microsoft's posts do not say what happens to a tuned model if a licence ends.
In short, open weights give portability but no vendor help; licensed weights, as described, give vendor help but no portability. Pick by whether your bigger risk is lock-in or operational burden.
What it means in practice
For most developers, MAI Code 2 is one more option in the Copilot model picker, which the changelog lists in VS Code, Visual Studio, JetBrains, Xcode, the Copilot CLI and more. New models are enabled by default unless a Business or Enterprise admin turns that off, so check that setting.
For engineering leaders, the headline is pricing pressure. The Next Web reports that Suleyman said on stage, "Right now, token spend is going through the roof." A cheaper credible model gives you leverage, but only once measured on your own repositories. For platform teams, licensed weights offer a tuned in-house model without building open-weight infrastructure, but it is early and gated. And if your agents run for hours across many files, keep a frontier model in the loop.
What to do now
- Build a small private test set. Pick 30 to 50 real tasks with tests that decide pass or fail, and run every candidate on them.
- Measure cost per completed task. Log total tokens, including reasoning tokens, and divide spend by tasks that passed, not by calls made.
- Test at your real effort level. If you run models at medium effort in production, benchmark at medium effort.
- Price the post-promotion rate. Model your budget at the reported full price after 17 May 2027, not the launch discount.
- Pilot Flash on high-volume chores. Start with lint fixes, docstrings and small refactors, with tests as the automatic gate.
- Add escalation and track it. Route failures to a larger model and watch the escalation rate; if it climbs, the savings shrink.
- Review Copilot model policy. Decide whether new models should switch on automatically for your organisation.
- Ask the hard licence questions. If licensed weights interest you, ask about upgrades, tuned weights at licence end, and data handling.
Risks and open questions
The benchmarks are vendor-chosen and run in-house. Microsoft picked the tests and ran most on its own platform. Independent reruns will tell us more.
One of the four benchmarks has known flaws. Epoch AI rated DeepSWE 1.1 "Flawed" after finding issues in at least 23 of its 113 tasks. Scores on it carry extra uncertainty.
The cost advantage depends on a temporary price. If the reported promotion ends as scheduled, the cost ratios roughly double at unchanged token use.
Long-horizon work is out of scope. Microsoft says the models are not yet tuned for very long-horizon agentic work or fundamental science research.
Licensed-weight terms are unclear. Eligibility, pricing, portability and end-of-licence rules are not public yet.
Customer results are self-reported. The AT&T, Toyota and Alaska Airlines numbers come from customer-specific evaluations that Microsoft summarises.
Watch next: independent leaderboard results for MAI Code 2 on Artificial Analysis and Terminal-Bench, Flash's official per-token price in GitHub's pricing table, the arrival of the models on OpenRouter and other third-party platforms, and the first public terms for MAI Licensed Weights.
Frequently asked questions
Is MAI Code 2 really 98% as good as GPT-6.1 Sol?
That is Microsoft's claim, averaged across four benchmarks it selected at max reasoning effort. A ratio of averages can hide large gaps on individual tests, so test both models on your own tasks.
Why does the cost comparison mention a promotion?
Microsoft shows MAI costs at a 50% promotional price. The Next Web reports it runs until 17 May 2027, after which MAI Code 2's per-token price doubles, and so would its cost share at unchanged token use.
What is the difference between cost per task and price per token?
Price per token is the vendor's rate for each unit of input or output. Cost per task is that rate multiplied by all tokens used to finish a job, including reasoning tokens. A model that uses more tokens can cost more per task even with a lower per-token price.
How are licensed weights different from open weights?
Open weights can be downloaded and run anywhere. MAI Licensed Weights, as Microsoft describes it, are licensed, customised and run in your own Azure environment. You gain vendor support but give up portability.
Can I use MAI Code 2 Flash for free?
The GitHub changelog lists MAI Code 2 Flash on Copilot Free and Student plans, as well as paid plans. Usage counts against each plan's included allowance under GitHub's usage-based billing. Larger MAI Code 2 requires a paid plan.
Summary
Microsoft's MAI Code 2 claims 98% of GPT-6.1 Sol's performance at 40% of the cost. Those numbers are ratios of averages across four Microsoft-chosen benchmarks, at max reasoning effort and at a 50% promotional price, and simple arithmetic shows that token use per task matters as much as price per token. Flash makes routing between a cheap tier and a strong tier more practical for high-volume coding chores, as long as you track escalations and cost per completed task. MAI Licensed Weights adds a middle path between closed APIs and open weights, with customisation and data control inside your own Azure environment, but access is gated and the terms are not public. Test on your own tasks, price the post-promotion rate, and respect Microsoft's own warning that these models are not tuned for very long-horizon agentic work.
Sources
- Official Microsoft AI: Introducing MAI Code 2 and MAI Code 2 Flash
- Official Microsoft AI: MAI Licensed Weights: Frontier performance you control
- Official Microsoft AI: Custom Models
- Web GitHub Changelog: MAI Code 2 and MAI Code 2 Flash in GitHub Copilot
- Code GitHub Docs: Models and pricing for GitHub Copilot
- Web The Next Web: Microsoft MAI Code 2 and licensed weights at Ignite
- Web Epoch AI: DeepSWE benchmark review
- Web Scale: SWE Atlas is Complete
- Web Artificial Analysis: Announcing FrontierCode on Artificial Analysis
- Web Terminal-Bench: Terminal-Bench leaderboard
Cite this article
Kelahmet, Ö. C. (2026, November 20). How to Read Microsoft's MAI Code 2 Cost Claims and What Licensed Weights Mean. Özer Cem KELAHMET. https://www.ozercemkelahmet.com/articles/mai-code-2-licensed-weights
BibTeX
@online{kelahmet2026code,
author = {Kelahmet, {\"O}zer Cem},
title = {{How to Read Microsoft's MAI Code 2 Cost Claims and What Licensed Weights Mean}},
year = {2026},
month = {nov},
url = {https://www.ozercemkelahmet.com/articles/mai-code-2-licensed-weights}
}Version historyfirst version
- Published
Factual corrections are listed here with what changed. Wording and formatting edits are not listed.