LATEST
AI Tools

Microsoft MAI Code 2: 98% of GPT-6.1 Sol at 40% of the Price

Updated November 18, 2026
14 min read
Editor's Note: This article was human-reviewed for accuracy
Close-up of lines of code on a laptop screen in a dark room

Quick Answer & Key Takeaways

  • • Microsoft launched seven MAI models at Ignite on November 17, 2026, covering coding, cybersecurity, image editing, voice and decision-making.
  • • MAI Code 2 claims 98% of GPT-6.1 Sol's coding quality at 40% of the price, but Microsoft has published no per-token rate for it yet.
  • • The prices Microsoft did publish are striking: Cyber 2 Flash at $0.10 per million input tokens and Realtime 1 from $0.01 per audio minute.

Microsoft walked into its Ignite conference on November 17, 2026 with a blunt pitch: stop paying frontier prices for everyday AI work. The company launched seven new models under its Microsoft AI (MAI) banner, each built for one job (coding, cybersecurity, image editing, live voice, or making decisions) and priced, Microsoft says, at a fraction of what OpenAI and Anthropic charge for the same work.

The headline number belongs to MAI Code 2, which Microsoft says delivers 98% of GPT-6.1 Sol's coding quality at 40% of the price, and 96% of Claude Opus 5.5's at 5% of the cost. But the launch is also a lesson in reading vendor announcements carefully. Microsoft published hard per-token prices for only some of the seven models, the most eye-catching percentages come from its own benchmarks, and the rival it benchmarks against most, GPT-6.1 Sol, was superseded the same day, when OpenAI announced GPT-6.2 Sol and GPT-6.2 Astra. Here is what each model costs, what it can actually do, and where the numbers need a second look.

The seven models at a glance

Microsoft counts seven models in the launch. Axios counts six. The difference is MAI Image 2.7 Flash, which appears on the image model's own page rather than in the main announcement. Either way, this is the lineup:

ModelBuilt forHeadline priceAvailability
MAI Code 2Everyday enterprise codingNo list price published; 40% of GPT-6.1 Sol, per MicrosoftGitHub Copilot now
MAI Code 2 FlashHigh-volume routine codingNo list price published; 65% of GPT-6 Luna, per MicrosoftGitHub Copilot; limited-access Foundry preview
MAI Cyber 2 FlashSecurity defence and vulnerability hunting$0.10 input / $0.01 cached / $0.50 output per 1M tokensMicrosoft security products only; MDASH from mid-December
MAI Image 2.7Image editingAbout $35 per 1,000 imagesPublic preview, Microsoft Foundry
MAI Image 2.7 FlashHigh-volume image generationAbout $10 per 1,000 imagesPublic preview, Microsoft Foundry
MAI Realtime 1Live speech-to-speech voiceFrom $0.01 per audio minuteMAI Labs; limited-access Foundry preview
Decision 1.1Routing work to cheaper modelsNo separate price published; Decision-1 lists at $0.042 per 1M input tokensAvailability details not published

MAI Code 2: the coding pair, and the missing price

MAI Code 2 is the launch's centrepiece, and Microsoft's case for it rests on four benchmarks for everyday enterprise coding: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. Averaged across those four, Microsoft says Code 2 reaches 98% of GPT-6.1 Sol's performance at 40% of the price, and 96% of Claude Opus 5.5's at 5% of the cost. The smaller Code 2 Flash is aimed at explanations, targeted edits and routine fixes, and Microsoft puts it at 65% of GPT-6 Luna's price and 20% of Claude Haiku 5.5's across three of the same benchmarks.

Microsoft also says Code 2 is tuned for dashboards and visualisations, and that in a blind side-by-side test of frontend work, reviewers preferred its output over GPT-6 Sol most of the time. The two Microsoft pages disagree slightly on that figure: the main announcement says 68% of the time, the Code 2 page says 67%. The gap is small, but it is a reminder that these are vendor-run evaluations with vendor-chosen framings, and no independent head-to-head exists yet.

The bigger gap is the price itself. Microsoft has published no per-token list price for Code 2 or Code 2 Flash. Its Foundry pricing page lists no Code 2 row at all as of November 17. If you apply Microsoft's percentages to the rivals' published rate cards, the implied figures are about $0.80 input and $4.00 output per million tokens for Code 2 (40% of GPT-6.1 Sol's $2 and $10), and about $0.20 and $1.00 for the Opus comparison (5% of $4 and $20). That is our arithmetic, not a Microsoft price, and it should not be quoted as one. Until Microsoft publishes a rate card, "40% of the price" is a claim about a benchmark average, not a bill you can forecast.

Two limits are worth knowing before you switch. First, Microsoft says the models are built for the coding work its customers do every day and are explicitly not yet tuned for harder problems, such as fundamental science research and very long-horizon agentic work. Second, Code 2 is available in GitHub Copilot today, which is the fastest way to try it. But Copilot's own subscription, not a per-token meter, is what you pay there.

MAI Cyber 2 Flash: the one with a real price tag

Of the seven models, Cyber 2 Flash is the only text model with a published per-token rate. Microsoft lists it at $0.10 per million input tokens, $0.01 per million cached input tokens and $0.50 per million output tokens. At November 17, 2026 market rates (1 USD = 0.756 GBP, 1 USD = 1.429 CAD), that is roughly £0.08 and C$0.14 per million input tokens, and £0.38 and C$0.71 per million output tokens.

The benchmark behind it is CyberGym-E2E, which measures real-world vulnerability discovery. Microsoft reports 92.3% task success for Cyber 2 Flash against 91.5% for GPT-6 Astra, using half the reasoning budget, at less than 2% of the cost per task. It also reports 40.4% on WebExploitBench against 38.4% for GPT-5.6 Cyber at the same reasoning budget. The model carries a one-million-token context window, four times the size of MAI Cyber 1.1 Flash, and Microsoft says its output-token price is 86% lower than that earlier model.

Microsoft deserves credit for two honesty notes in its own footnotes. First, the "less than 2% of the cost per task" figure reflects offline evaluation cost, not customer spend. Your bill depends on your own prompts, caching and volume. Second, while reviewing its CyberGym-E2E results, Microsoft found that 92 of the benchmark's 905 tasks are based on the same open-source projects and vulnerability reports as some of its training data, so it excluded those 92 tasks for every model and reports results on the remaining 813. That is the right call, and it is also a reminder that the headline 92.3% is a filtered number.

There is a catch on access. Cyber 2 Flash is coming exclusively to Microsoft's security products, and it will automatically replace the existing model in the MDASH profile for eligible customers starting in mid-December. You cannot simply call it from Foundry the way you can the image models. One customer datapoint Microsoft cites: Insight Enterprises found MAI Cyber models 20% more token-efficient than GPT Profile in its own evaluations.

MAI Image 2.7: number one by four Elo points

MAI Image 2.7 is the launch's clearest ranking win. Microsoft says it takes the number one spot on Artificial Analysis's image-editing board at 1,215 Elo across 5,478 votes, ahead of GPT Image 2.6 Sunburst (max) by four Elo points. On text-to-image generation it ranks fourth, which Microsoft frames as making it the second-best lab overall. Pricing is per image rather than per token in Microsoft's announcement: about $35 per 1,000 images for Image 2.7 and about $10 per 1,000 for Flash, which works out to roughly $0.035 and $0.01 per image. In pounds and Canadian dollars, that is about £26.46 and C$50.02 per 1,000 images for the full model, and £7.56 and C$14.29 for Flash.

The models are in public preview on Microsoft Foundry, and Microsoft has also launched Canvas, a workspace in MAI Labs powered by the Flash model for creating, editing and refining images. MAI Image is now the default model family behind Bing Image Creator, where Microsoft says usage has grown at a compounded weekly rate of about 15% since June. Figma Weave, Minute (Sports Illustrated), Wirestock and Manus are named as early users.

Keep the ranking in proportion. Four Elo points is a narrow lead, and leaderboard snapshots move: when Microsoft first previewed Image 2.7 on October 27, the same board had it second at 1,193 Elo, behind GPT Image 2.5 Sunburst. The November 17 figure is a new reading against a newer rival, GPT Image 2.6 Sunburst, not the same measurement repeated. A four-point lead on a board where the top models sit within a few points of each other is a win, not a rout.

MAI Realtime 1: a cent a minute, with a reasoning surcharge

MAI Realtime 1 is Microsoft's first live speech-to-speech model, and its pricing is the simplest in the launch: base speech-to-speech usage starts at $0.01 per audio minute, or roughly £0.008 and C$0.014 per minute. The fine print matters here. Reasoning is billed separately at $5 per million output tokens (about £3.78 and C$7.15), so a voice agent that has to think hard before it speaks costs more than the headline minute rate suggests. Microsoft says the model outperforms GPT-Realtime 2.1 on quality, speed and price across public benchmarks and its own Dynamics 365 evaluations, citing Big Bench Audio for speech reasoning and the Tau-Voice benchmark for agentic task completion.

It is built to be a drop-in option for existing voice stacks: Microsoft says it is compatible with the OpenAI Realtime API, so developers can test it without rewriting their integration. Microsoft is also bringing it to Dynamics 365 Contact Center for customer service and to Dragon Copilot, where it will support AI-simulated patient conversations for nurse training. You can try it in MAI Labs or in a limited-access Foundry preview, and preview customers will not be billed until public preview begins. That also means there is no production billing to plan around yet.

Decision 1.1, Dragon Copilot Transcribe, and the rest

Decision 1.1 is Microsoft's routing model: it sits inside applications and decides which questions need an expensive model and which can be answered cheaply. Microsoft says it is 50 times faster and 79 times cheaper than GPT-6.1 Sol, and that on Artificial Analysis's Decision Index it sits ahead of Jev, Google and OpenAI. It also says the Decision family is used by 4,000 customers.

That index claim needs a careful read. The independent Artificial Analysis Decision Index v2.1 numbers available today are for the previous version, Microsoft-Decision-1, which scores 62 against Jev 1.14's 66, with both models priced at $0.042 per million input tokens and free output. Decision 1.1's claimed lead is Microsoft's own statement, not yet an independently published score, and Microsoft has not published a separate per-token price for 1.1. Treat the 79-times-cheaper figure the same way: it is Microsoft's comparison against GPT-6.1 Sol, not a rate you can verify on a price page.

The launch also includes Dragon Copilot Transcribe, built on MAI Transcribe 2, which Microsoft says now powers all doctor-patient conversation transcription in Dragon Copilot: 10 million patient encounters a month, in the hands of a third of US physicians, with multilingual error rates halved. And there is hardware in the announcement: Maia 200, Microsoft's second-generation AI accelerator, is now in production running its largest models at 1.7 times more tokens per watt. Finally, Microsoft is offering MAI Licensed Weights, letting enterprises take a frontier MAI model, customise it with proprietary data and run it in their own Azure environment. Microsoft AI chief Mustafa Suleyman told Axios the approach would "feel like open source" while keeping deployment inside Azure.

What Microsoft did not say

First, the timing. OpenAI announced GPT-6.2 Astra and GPT-6.2 Sol on the same day as Microsoft's launch. Microsoft's coding comparisons are against GPT-6.1 Sol, the previous generation. GPT-6.2 Sol keeps the same $2 input and $10 output per million token pricing as 6.1 Sol, according to model tracker ArtificialWatch, and the new Astra is priced 20% below the previous version, according to Axios's reporting. Microsoft's percentages may still hold against the new models, but they were not measured against them.

Second, the discounts. Axios notes that Microsoft's comparisons use selected benchmarks and that some of its pricing advantages reflect promotional discounts. Microsoft's own pages do not label any of the launch prices as promotional, so a buyer should ask which rates are permanent and which expire.

Third, the verification gap. Every benchmark in the launch is Microsoft-run, on Microsoft-chosen tests, and no independent evaluation of Code 2, Cyber 2 Flash or Realtime 1 exists yet. That is normal on launch day, and it is also the reason to treat the percentages as a starting point for your own tests rather than a conclusion. This is Microsoft's second major MAI push this autumn: in October it launched MAI-DxO, a medical model that it said beat doctors on diagnostic cases, a claim that also needed careful reading.

How the prices stack up

ModelInput per 1M tokensOutput per 1M tokensWhat the number is
MAI Cyber 2 Flash$0.10$0.50Microsoft list price; cached input $0.01
GPT-6 Astra$10$50OpenAI rate card
GPT-6.1 Sol$2$10OpenAI rate card
Claude Opus 5.5$4$20Anthropic pricing
GPT-6 Luna$0.10$0.50OpenAI rate card
Claude Haiku 5.5$0.10$0.50Anthropic pricing, up to 100K tokens
Microsoft-Decision-1$0.042FreeFoundry list price; AA Decision Index v2.1: 62
Jev 1.14$0.042FreeTypeSafe list price; AA Decision Index v2.1: 66
OpenAI Decisions API$0.10FreeOpenAI rate card; AA Decision Index v2.1: 60

The table shows why Cyber 2 Flash is the launch's most concrete deal: its $0.10 input and $0.50 output rates are identical to GPT-6 Luna's, and one-hundredth of GPT-6 Astra's, for a model Microsoft says matches Astra on its own security benchmark. The decision-model rows show a different pattern. Microsoft, Jev and OpenAI all charge per input token with free output, and the independent index puts Jev ahead of Microsoft's previous version, which is exactly the gap Decision 1.1 is supposed to close.

For the models without per-token prices, the honest comparison is still unwritten. Code 2's value depends entirely on the rate card Microsoft has not published. The rates that are published, Realtime 1 at $0.01 a minute and Image 2.7 Flash at about $0.01 an image, are the two numbers a small team can budget against today.

Who should try them, and who should wait

Try them now if you already live in Microsoft's stack. MAI Code 2 is already in GitHub Copilot, so Copilot users can judge it on their own code without signing anything new. MDASH customers get Cyber 2 Flash automatically from mid-December. Foundry users can test the image models in public preview and request access to the Realtime 1 preview.

Wait if you need a contracted price before you build. No team should migrate a coding workload to Code 2 on the strength of "40% of the price" alone, because there is no published rate to put in a contract. The same applies to Decision 1.1: until its own price and an independent index score appear, the safe move is to benchmark it against Jev and OpenAI's Decisions API on your own traffic, as our decision-API comparison shows how to do. And if voice is your use case, compare Realtime 1's per-minute rate with the per-minute pricing in our Mistral Voxtral 3 review before committing. A cent a minute is cheap only if the reasoning surcharge stays small on your calls.

Three mistakes to avoid. Do not quote "40% of the price" as a price. Do not read Cyber 2 Flash's cost-per-task figure as your bill. Microsoft says it is an offline evaluation cost. And do not treat a four-Elo image-editing lead as a wide gap; on these boards, four points is a photo finish.

Frequently asked questions

How much does MAI Code 2 cost?

Microsoft has not published a per-token list price for MAI Code 2 or Code 2 Flash. It says Code 2 delivers 98% of GPT-6.1 Sol's coding quality at 40% of the price and 96% of Claude Opus 5.5's at 5% of the cost, but those are relative claims from Microsoft's own benchmarks. Its Foundry pricing page lists no Code 2 row as of November 17, 2026.

Where can I try MAI Code 2?

MAI Code 2 is available in GitHub Copilot now. Code 2 Flash is also in Copilot, and in a limited-access preview on Microsoft Foundry. Neither model has a public per-token API price yet.

Is MAI Cyber 2 Flash available to everyone?

No. It is coming exclusively to Microsoft's security products, and it will automatically replace the existing model in the MDASH profile for eligible customers starting in mid-December 2026. Its published price is $0.10 per million input tokens, $0.01 cached, and $0.50 per million output tokens.

Is MAI Image 2.7 really the best image editor?

On Artificial Analysis's image-editing board it ranks first at 1,215 Elo, four points ahead of GPT Image 2.6 Sunburst (max), based on 5,478 votes. That is a real lead but a narrow one, and on text-to-image generation the same model ranks fourth. The full model costs about $35 per 1,000 images in public preview on Foundry; Flash costs about $10.

What is Decision 1.1?

Decision 1.1 is Microsoft's routing model: it answers typed questions with probabilities so applications can send only hard questions to expensive models. Microsoft says it is 50 times faster and 79 times cheaper than GPT-6.1 Sol and leads on Artificial Analysis's Decision Index. The independent index numbers available today are for the previous version, Decision-1, which scores 62 against Jev's 66, and no separate price for 1.1 has been published.

What is MAI Licensed Weights?

It is Microsoft's offer to let enterprises take a frontier MAI model, customise it with their own data and run it in their own Azure environment, delivered through a programme called Frontier Company. Suleyman told Axios it would "feel like open source", with the condition that deployment stays inside Azure.

Microsoft's bet is clear: specialist models, tuned for one job, can match frontier generalists at a fraction of the cost. On the numbers published so far, Cyber 2 Flash and Realtime 1 make that case in hard currency, while Code 2 makes it in percentages that still need a price tag. The figures to watch next are the ones Microsoft has not published yet: Code 2's list price and independent benchmarks for all seven models. When either lands, this article will be updated.

Sources

14 min read
58 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more