the short version
- Microsoft states its new coding models' costs only as percentages of competitors, so per-million-token budgeting is not yet possible from the announcements.
- The headline 98%-at-40% figure is an average over four benchmarks Microsoft picked to mirror GitHub Copilot workloads, and the scoring setup behind it is not in the post text.
- Microsoft says the models are not yet tuned for very long-horizon agentic work, which puts long autonomous agent runs outside what it claims for them.
- Neither announcement's available text names Azure AI Foundry, GitHub Copilot model-picker availability, rate limits or context length.
Microsoft released MAI Code 2 and MAI Code 2 Flash on November 17, 2026. Both are coding models aimed at everyday enterprise work rather than the hardest problems. Microsoft says MAI Code 2 delivers 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, averaged over four coding benchmarks. Code 2 Flash is claimed to match GPT-6 Luna at 65% of the cost and Claude Haiku 5.5 at 20%.
On November 18 the pair was included in a Microsoft AI post introducing seven new models. The others were MAI Cyber 2 Flash, MAI Image 2.7, MAI Realtime 1, Dragon Copilot Transcribe on MAI Transcribe 2, and Decision 1.1. For engineers deciding whether to route coding traffic to the new models, the most important numbers are missing. The published text of both posts gives no per-token prices, no rate limits and no named endpoint.
What MAI Code 2 costs per million tokens
Neither announcement gives a per-million-token input or output price for either model. Every cost claim is relative: 40% of GPT-6.1 Sol, 5% of Opus 5.5, 65% of GPT-6 Luna and 20% of Haiku 5.5. The model blog says "cost", while the seven-model post says "about 40% of the price". Neither post states whether these figures are list price per token, cost per completed task or cost per benchmark run.
That distinction matters for anyone modelling spend. A model that is cheaper per token but uses more tokens or retries per task can land at a different ratio than the headline. One commenter on Hacker News said the comparison showed MAI Code 2 at a 50% promotional discount, against Opus 5.5 and GPT-6.1 Sol run at Max reasoning effort. The available text of either post neither confirms nor rules out that claim, and the announcement's footnotes 1 through 3 are referenced but not reproduced. Until list prices appear, treat the ratios as Microsoft's own framing.
98% of GPT-6.1 Sol on four Copilot-style benchmarks
The flagship comparison averages four benchmarks that Microsoft says "target the everyday enterprise coding tasks we see in GitHub Copilot". They are DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. Per-benchmark scores are not given in the post text. A team therefore cannot tell whether the 98% average hides a weak spot on, say, terminal work.
MAI Code 2 gives you 98% of the performance of GPT-6.1 Sol at 40% of the cost, and 96% of Claude Opus 5.5’s performance at just 5% of the cost.
Microsoft also cites blind frontend app evaluations in which developers preferred MAI Code 2's results over GPT-6 Sol 64% of the time. That comparison is against GPT-6 Sol, not the GPT-6.1 Sol used in the headline average, and the post text does not give the sample size. Microsoft lists the model's strengths as taking code to production, deep debugging, multi-file implementation, repo-wide reasoning and UI. It also says the model is tuned for dashboards and visualizations.
Code 2 Flash targets Haiku 5.5 and Luna workloads
The smaller model is pitched at high-volume work: code explanations, targeted edits, routine fixes and rapid iterations. Microsoft says it offers comparable performance to GPT-6 Luna and Claude Haiku 5.5 across three coding benchmarks, at 65% and 20% of their cost respectively. Unlike the flagship comparison, the three benchmarks are not named in the post text.
This tier is where the cost claim is easiest to test, because routine edits are cheap to evaluate against a team's own repository. One Hacker News commenter said a lot of their coding workload already goes to Haiku and Luna and that they are interested in trying the new model on it. That is the substitution Microsoft's Flash comparison is aimed at.
Not yet tuned for long-horizon agentic work
Microsoft is explicit about scope. It says it built these models to prioritize everyday coding work and that they are "explicitly not yet tuned for harder problems, like fundamental science research and very long-horizon agentic work." The seven-model post frames the lineup around a price point that lets customers run agents 24-7. The coding announcement itself, however, limits its claims to everyday tasks.
For teams running autonomous coding agents, this suggests routing some work to the models rather than switching everything. Well-scoped tasks such as targeted edits, debugging and multi-file changes fit the stated strengths, while very long-horizon agent runs are outside what Microsoft claims. One commenter on Hacker News asked whether the models are Microsoft's own or reinforcement-learned on top of open models. Neither source answers that, and neither names a licence or a weights release.
Where can engineers call MAI Code 2?
The available text of both announcements does not say where engineers can call the models. Azure AI Foundry is not mentioned. GitHub Copilot appears only as the source of the task mix the benchmarks were chosen to reflect, not as a stated deployment surface. Neither source gives a rate limit, context window, region list or SLA.
- Per-million-token input and output prices: not published.
- API endpoint, Foundry listing or model ID: not published.
- GitHub Copilot model-picker availability: not stated.
- Rate limits, context length and regions: not stated.
- Serving hardware: the seven-model post says Maia 200 is in production running Microsoft's largest models at 1.7x more tokens per watt, but does not say it serves these coding models.
Practitioner reaction on Hacker News reflects the access gap. One commenter said they had wanted to try MAI models since v1 but Microsoft locked down Copilot subscriptions, and that they would consider it if offered a plan comparable to Claude Max. Another said enterprises without negotiated Anthropic API discounts are better served by GitHub Copilot, which already offers Anthropic and OpenAI models. A third noted that large companies already reach those models through GitHub and Copilot. Until Microsoft publishes prices and an endpoint, the practical step is to prepare an internal eval set of routine tasks and hold off on migration plans.
Questions this raises
How much does MAI Code 2 cost per million tokens?
Microsoft has not published a per-million-token input or output price for MAI Code 2 or Code 2 Flash. Every cost claim is relative, such as 40% of GPT-6.1 Sol or 5% of Claude Opus 5.5, and the posts do not say whether that means list price per token or cost per task.
Where can developers call MAI Code 2?
The announcements do not name an API endpoint, model ID or Azure AI Foundry listing. GitHub Copilot is mentioned only as the source of the task mix behind the benchmarks, not as a stated deployment surface.
Is MAI Code 2 suited to long-horizon agentic coding?
No, by Microsoft's own description. It says the models are explicitly not yet tuned for harder problems like fundamental science research and very long-horizon agentic work, and are built for everyday coding tasks such as debugging and multi-file changes.
Spotted an error? Email samir.s.sengupta@gmail.com and I will correct it and date the change at the top of the note. How these notes are researched, drafted and checked is in the editorial policy.
