Skip to content

Microsoft Says MAI Code 2 Hits 98% of GPT-6.1 Sol at 40% of the Price, but That Price Is Already Half Off

Microsoft's MAI Code 2 claims 98% of GPT-6.1 Sol at 40% of the price. The 40% uses a 50% promotional rate, and its cyber benchmark dropped 92 of 905 tasks.

Microsoft AI used its Ignite 2026 keynote on 17 November to launch six new MAI models and one licensed weights programme, all aimed at token bills that outrun the value teams get from agents. The headline number was for coding. MAI Code 2 reaches 98% of GPT-6.1 Sol’s quality at 40% of the price, Microsoft said, and 96% of Claude Opus 5.5 at 5% of the price.

The 40% is not a list price. Microsoft’s own footnote says the Code 2 cost is shown “at its 50% promotional price”, and the model page publishes no per-token rate for the model at all. At full price the same comparison would land near 80% of Sol’s cost, not 40%.

The rest of the launch has the same tight framing. The cybersecurity model wins a benchmark from which Microsoft removed 92 of 905 tasks before publishing. The image editing model is number one by 4 Elo points, with confidence intervals that overlap the model it beats. None of that makes the models fake. It means the launch deck is doing more work than the shipping dates.

Quick answer: Microsoft really did ship MAI Code 2, Code 2 Flash, MAI Cyber 2 Flash, MAI Image 2.7, Image 2.7 Flash and MAI Realtime 1 on 17 November, and Code 2 is live inside GitHub Copilot. But the 98% at 40% claim depends on a 50% promotional rate that Microsoft has not said it will keep, the coding scores come mostly from Microsoft’s internal test harness, and the cyber model is not sold as a standalone API. Treat the numbers as a negotiation opening, not a budget line.

What Microsoft actually shipped

The shipping list is narrower than the announcement. MAI Code 2 and Code 2 Flash are in GitHub Copilot today, with Code 2 Flash in limited-access preview in Microsoft Foundry. The others arrive with caveats: Image in public preview, Realtime in a Foundry preview that will not be billed until public preview, and Cyber inside Microsoft security products from mid-December.

The “seven new world class models” headline also counts things that are not new models. The list includes Dragon Copilot Transcribe, a product built on MAI Transcribe 2, Decision 1.1, a routing model already generally available, and Maia 200, a chip. Six models shipped.

The 40% number rests on a discount and an internal harness

The coding claim averages four benchmarks: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. Three of those ran on MSBench, Microsoft’s internal evaluation platform, driven through the GitHub Copilot CLI with subagents switched off, at maximum reasoning effort. Only FrontierCode scores and costs came from Artificial Analysis, an independent evaluator.

Microsoft also says these models are “explicitly not yet tuned” for fundamental science research or very long horizon agent work, narrowing the claim to everyday enterprise coding. The rival is moving too: GPT-6.1 Sol remains the cost benchmark Microsoft measures against.

The independent record is thinner than the press release suggests. Artificial Analysis’s main LLM leaderboard, fetched on 18 November, lists 114 models and does not include MAI Code 2 at all. That is not proof of weakness. But buyers cannot cross-check the 98% against the same leaderboard they use for every other model.

The one human number is more useful. In blind frontend tests with 258 annotators, developers preferred MAI Code 2’s output over GPT-6 Sol 64% of the time. That is a signal teams can retest on their own repository. Microsoft does not publish the raw vote counts.

Pricing is the weakest part of the package. Code 2 ships through Copilot seats rather than a published token rate, so 40% is a modelled cost, not a rate card. A promotional discount that expires can double the bill overnight, and Microsoft has not published the expiry date. Time limited pricing is routine, as Anthropic’s one week Fable 5.2 launch showed.

Model Microsoft’s claim Price Microsoft published The catch
MAI Code 2 98% of GPT-6.1 Sol at 40% of cost; 96% of Opus 5.5 at 5% No per-token rate; sold through GitHub Copilot Cost modelled at a 50% promotional price
MAI Code 2 Flash Matches GPT-6 Luna at 65% of cost, Haiku 5.5 at 20% No published rate yet Also uses the 50% promotional price
MAI Cyber 2 Flash 92.3% on CyberGym-E2E vs 91.5% for GPT-6 Astra $0.10 input, $0.01 cached, $0.50 output per 1M tokens Not sold standalone; Microsoft security products only
MAI Image 2.7 Number 1 on AA-Image-Editing v2.1 About $35 per 1,000 images 4 Elo ahead, intervals overlap
MAI Image 2.7 Flash Frontier quality at high volume About $10 per 1,000 images Ranks 7th on the same board
MAI Realtime 1 Beats GPT Realtime 2.1 on quality, speed and price $0.01 per audio minute, plus $1 and $5 per 1M reasoning tokens Preview only; reasoning billed separately
Decision 1.1 50x faster and 79x cheaper than GPT-Sol 6.1 Not published on the launch page Already available before Ignite

The cyber result lost 92 tasks before it was published

MAI Cyber 2 Flash scores 92.3% on CyberGym-E2E against 91.5% for GPT-6 Astra, at less than 2% of the cost per task. Those are close scores on a benchmark where Microsoft also ran its model at half the reasoning budget of the rival, which is a genuine efficiency point rather than a quality one.

Microsoft then disclosed something most vendors leave in a footnote. While reviewing results, it found that 92 of the benchmark’s 905 tasks were built from the same open source projects and vulnerability reports as some of its training data. It excluded those 92 tasks for every model and reported on the remaining 813. The reference solutions were never in training, Microsoft says, but the exclusion is still the honest move and it still shrinks the test.

The cost comparison is easy to check because both prices are public. MAI Cyber 2 Flash lists at $0.10 input, $0.01 cached input and $0.50 output per million tokens. GPT-6 Astra on Azure OpenAI Global lists at $10, $1 and $50. That 100x rate gap is where “less than 2%” comes from.

The model itself may be hard to buy. It is coming “exclusively to Microsoft security products”, starting with Codename MDASH. A security team outside that stack cannot benchmark it against its own incident queue.

Image editing is a real number one, by four Elo

MAI Image 2.7 tops AA-Image-Editing v2.1 with 1,215 Elo against 1,211 for GPT Image 2.6 Sunburst at max settings, across 5,478 votes. The ranking is real and the price gap is real: Artificial Analysis lists $34.82 per 1,000 images against $210.72 for the OpenAI model.

The gap is also inside the error bars. Microsoft’s model has a 95% interval of 1,201 to 1,229; OpenAI’s runs 1,198 to 1,224. On this sample the two are statistically tied for first, which is a different statement from “the best image editing model in the world”. The price claim survives that scrutiny better than the quality claim does.

The Flash variant is the more interesting product for Indian teams shipping at volume. It ranks 7th overall at $10.24 per 1,000 images, roughly a twentieth of GPT Image 2.6.

Confirmed versus not confirmed

Claim Status
Six MAI models were announced on 17 November at Ignite Confirmed on Microsoft’s blog and model pages
MAI Code 2 is available in GitHub Copilot now Confirmed by Microsoft; not tested in a Copilot seat
98% of GPT-6.1 Sol at 40% of cost Vendor claim, cost modelled at a 50% promotional price
MAI Cyber 2 Flash at 92.3% on CyberGym-E2E Vendor result on 813 of 905 tasks after exclusions
MAI Image 2.7 is number 1 on AA-Image-Editing Confirmed independently, by 4 Elo with overlapping intervals
MAI Image 2.7 is number 1 on text-to-image Not confirmed; Microsoft says it ranks 4th
List price of MAI Code 2 per million tokens Not published
Length of the 50% Code 2 promotion Not published
MAI Cyber 2 Flash as a standalone API Not confirmed; Microsoft says security products only
Performance on Indian language code comments and docs Not tested or published

What it means for Indian developers

The practical question is not whether 40% is real. It is whether your workload looks like Microsoft’s benchmark. Four everyday coding tasks, max reasoning, no subagents, run through Copilot’s CLI, is a decent proxy for enterprise refactoring and a poor one for long agent runs that touch ten services.

Teams already paying for Copilot seats get Code 2 without a new contract, which is the cheapest way to test the claim. Run your own eval on twenty tickets from your backlog and compare merge-ready diffs, not benchmark scores. If the model needs three retries to match Sol, the 40% disappears in retry cost, and the promotional clock makes that worse.

Data residency is the second issue. Microsoft has Azure regions in India, but Foundry previews usually land there later than in the US. Confirm the region before designing around Image 2.7 or Realtime 1, because both are preview services with limited footprints.

MAI Licensed Weights deserves attention beyond the launch noise. Microsoft will customise a frontier model on your data and let you run it in your own Azure environment, with your data and adaptations staying yours. That is not open weights. You cannot download the file or run it on your own GPUs.

The quiet winner is the voice model. At $0.01 per audio minute plus $1 and $5 per million reasoning tokens, MAI Realtime 1 is priced for Indian support desks handling tens of thousands of calls a month. It is also compatible with the OpenAI Realtime API, so migration is mostly an endpoint change. That matters more to a Gurugram contact centre than a four Elo image ranking matters to anyone.

Frequently asked questions

Is MAI Code 2 actually cheaper than GPT-6.1 Sol?

On Microsoft’s own evaluation, yes, at about 40% of the cost for 98% of the quality. The catch is that 40% is calculated at a 50% promotional price. At list price the gap narrows to roughly 80%, and Microsoft has published neither rate.

Where can I use MAI Code 2 today?

Inside GitHub Copilot. MAI Code 2 Flash is also in limited-access preview in Microsoft Foundry. Microsoft says both will appear on OpenRouter, Vercel, Hermes, Open Code and Kilo soon, but gives no date, and direct per-token access is not available yet.

Can I download the MAI Code 2 weights?

No. Microsoft’s licensed weights programme lets you customise a model and run it inside your own Azure environment, but the weights stay under Microsoft’s control. This is not an open weights release and cannot be self-hosted outside Azure.

Is MAI Image 2.7 really the best image editing model?

It is first on AA-Image-Editing v2.1 with 1,215 Elo, four points ahead of GPT Image 2.6 Sunburst. The confidence intervals overlap, so the honest description is a statistical tie at the top. On text-to-image it ranks fourth.

Sources

Share this article

Leave a Reply

Your email address will not be published. Required fields are marked *

Loading the next article…

Continue reading