ModelBrief · MAI Code 2 · Topic Page
Last updated Nov 20, 2026
USA · LLM · GA · Released Nov 17, 2026

MAI Code 2: The Topic Page

Microsoft's in-house coding flagship, released at Ignite 2026 on November 17, 2026, and live in GitHub Copilot the same day alongside its lightweight sibling MAI Code 2 Flash. Microsoft's own claim: a four-benchmark average reaching 98% of GPT-6.1 Sol at about 40% of the cost (Microsoft-chosen benchmarks, 50% promotional-price basis), and 96% of Claude Opus 5.5 at just 5% of the cost (same basis). But the positioning needs stating first: every benchmark figure is Microsoft's own, with no independent score anywhere; the price is GitHub's published promotional rate ($0.50/$2.50 per 1M tokens, ending May 17, 2027, doubling after), and the 40% and 5% figures are relative costs computed at that promo price; the model card is published (about 80B active, ~2.8T total, 1.05M context, knowledge cutoff before August 2026 data). This page collects the official specs, the ways to try it, a pricing breakdown and 6 in-depth reviews — with unverified claims clearly marked.

98%Microsoft-reported four-benchmark average vs GPT-6.1 Sol (Microsoft-chosen benchmarks, 50% promo-price basis)
96%vs Claude Opus 5.5, at just ~5% of the cost (same basis)
Nov 17, 2026Ignite launch, live in GitHub Copilot the same day
$0.50 / $2.50Promo price per 1M tokens (input/output), to May 17, 2027, then $1.00/$5.00
1.05M contextModel card published: about 80B active, ~2.8T total, knowledge cutoff before August 2026 data
Seven MAI modelsThe MAI family launched the same day, all on the Maia 200 chip

01Key facts

What is confirmed: Microsoft's first in-house coding model positioned as a Copilot workhorse — "near-flagship scores at a fraction of the price" is its pitch, and also the claim most in need of checking.

Released
2026-11-17 (Tuesday, Ignite 2026, San Francisco, Nov 17–20); Microsoft AI launched seven MAI models the same day, with MAI Code 2 and MAI Code 2 Flash live in GitHub Copilot the same day
Developer
Microsoft AI (Microsoft), USA
Positioning
Everyday enterprise coding flagship: production code, deep debugging, multi-file changes, repo-wide reasoning, UI work; Microsoft says both models are "explicitly not yet tuned" for fundamental science research or very-long-horizon agentic work
Sibling
MAI Code 2 Flash: the lightweight tier for code explanations, targeted edits, routine fixes and rapid iteration; comparable to GPT-6 Luna and Claude Haiku 5.5 (Microsoft's own runs), at 65% and 20% of the cost (50% promo-price basis); covered inside this topic, not as its own topic
Parameters / architecture
Decoder-only transformer with sparse MoE layers and an integrated vision encoder; MAI Code 2 about 80B active (77B active + 3B embedded, ~2.8T total), Flash about 12B active (~398B total); base MAI-Base-2 (model card, Nov 17, 2026)
Context window
1,050,000 tokens (model card); max input 922,000, max output 128,000; the 1M window works only in VS Code and Copilot CLI (GitHub docs)
Knowledge cutoff
Data before August 2026 (model card); it may not know newer libraries, APIs or vulnerabilities
Benchmarks
Microsoft-reported four-benchmark average: 98% of GPT-6.1 Sol at about 40% of the cost (Microsoft-chosen benchmarks, 50% promo-price basis); 96% of Claude Opus 5.5 at about 5% of the cost (same basis). No independent score exists: not listed by Artificial Analysis, LiveBench or LMArena
Blind test
A 258-person front-end blind test: developers preferred MAI Code 2's output over GPT-6 Sol, now stated as 64% in both Microsoft posts (better on both aesthetics and functionality); 67% and 68% were the pre-correction values on the pages as saved around 04:00 JST on Nov 18, now kept only as a historical note (see sections 12 and 15)
Price
$0.50 / $2.50 per 1M tokens (GitHub Copilot pricing page): $0.50 input, $0.025 cached input, $2.50 output, promotional to May 17, 2027, then $1.00/$0.05/$5.00; Flash $0.05/$0.005/$0.25, then $0.10/$0.01/$0.50; usage-based inside GitHub Copilot (1 AI credit = $0.01)
Access
GitHub Copilot (VS Code, Visual Studio, Copilot CLI, cloud agent, GitHub.com, mobile, JetBrains, Xcode, Eclipse); MAI Code 2 Flash also in limited-access preview in Microsoft Foundry; MAI Code 2 has joined Copilot Fusion Standard
Plans
MAI Code 2: Copilot Pro, Pro+, Max, Business, Enterprise; MAI Code 2 Flash: those plans plus Free and Student
Weights / open source
Closed, no public weights
Chip
All seven MAI models run on Microsoft's second-generation Maia 200 chip, now in production

02Quick comparison

Microsoft's pricing story in one table: 2–4% less performance at a fraction of the cost — but both the performance and the cost come from Microsoft's own arithmetic; see the note below the table.

ModelFour-benchmark average (relative performance)Relative costContextOpen weights
MAI Code 298 (GPT-6.1 Sol = 100)~40% (50% promo-price basis)1.05M tokensNo
GPT-6.1 Sol100 (reference)100 (reference)1M tokensNo
Claude Opus 5.5100 (reference)100 (reference)1M tokensNo
MAI Code 2 FlashComparable to GPT-6 Luna and Claude Haiku 5.5 (Microsoft's own runs, three benchmarks)65% of Luna / 20% of Haiku 5.5 (50% promo-price basis)1.05M tokensNo

Relative performance is the average of four Microsoft-chosen benchmarks (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, FrontierCode v1.1), all models at max reasoning effort; three of the four ran on Microsoft's internal MSBench platform (via the Copilot CLI), and the FrontierCode v1.1 scores and costs come from Artificial Analysis. Costs are computed at the 50% promotional price, which ends May 17, 2027. All figures are Microsoft-reported, with no independent verification.

03Charts

Relative performance vs relative cost, grouped: MAI Code 2 trails the reference models by only 2–4% on performance while costing 40% and 5% — but both sets of numbers are computed at the 50% promotional price and all are Microsoft-reported.

Relative performance vs relative cost: MAI Code 2 vs GPT-6.1 Sol at 98% performance and 40% cost; vs Claude Opus 5.5 at 96% performance and 5% cost (Microsoft-reported, 50% promo-price basis)
MAI Code 2 relative performance vs relative cost (Microsoft-reported, 50% promo-price basis): 98% performance at 40% of the cost vs GPT-6.1 Sol; 96% at 5% vs Claude Opus 5.5. Source: Microsoft AI product post (Nov 17, 2026).

04Top 3 reads

Short on time? Start with these three: the benchmark tally, the pricing breakdown and the Opus 5.5 comparison — all three take apart Microsoft's line "98% of the performance at 40% of the cost".

05Timeline

Three days after launch, the official posts, GitHub changelog and tier-1 coverage all arrived the same day; Chinese coverage is still thin and no independent score has appeared.

  • June 2026

    MAI Code 1 Flash launches, the starting point of the MAI coding line (per industry reporting); the line stayed lightweight until MAI Code 2 added a general-purpose flagship tier.

  • 2026-11-17

    Seven MAI models launch at Ignite 2026: MAI Code 2, MAI Code 2 Flash (coding), MAI Cyber 2 Flash (security), MAI Image 2.7 (image editing), MAI Realtime 1 (speech-to-speech), MAI Transcribe 2 (transcription, powering Dragon Copilot Transcribe), Decision 1.1 (task routing); all run on the second-generation Maia 200 chip, now in production. The Microsoft AI product post gives all the benchmark claims: MAI Code 2 reaches 98% of GPT-6.1 Sol at about 40% of the cost (Microsoft-chosen benchmarks, 50% promo-price basis), and 96% of Claude Opus 5.5 at about 5% of the cost (same basis).

  • 2026-11-17

    MAI Code 2 and Code 2 Flash go live in GitHub Copilot the same day: GitHub's changelog confirms both models in the model picker across VS Code, Visual Studio, Copilot CLI, cloud agent, GitHub.com, mobile, JetBrains, Xcode and Eclipse; MAI Code 2 covers Pro, Pro+, Max, Business and Enterprise, Flash adds Free and Student; both are usage-billed. GitHub also confirms MAI-Code-2 joins Copilot Fusion Standard the same day, and announces MAI-Code-1.1-Flash will retire on Dec 18.

  • 2026-11-17–18

    Axios covers the launch (via superpowerdaily), noting the pricing advantages reflect promotional discounts; edgeconsultancykw, rrtechpulse, pasqualepillitteri and scanradar publish launch-day analyses, all labelling the comparisons as Microsoft's own and the prices as promotional. English autocomplete forms the same day with comparison-shaped long tails: "mai code 2 benchmarks", "mai code 2 vs opus 5.5", "mai code 2 artificial analysis".

  • 2026-11-18

    Blind-test figure unified at 64%: ai-oriens' Nov 18 report confirms both Microsoft's product post and its Ignite post now read 64% (258-person blind test, better on aesthetics and functionality); pages saved around 04:00 JST on Nov 18 showed 67% in the product post and 68% in the keynote post, both since corrected to 64%. The 67%/68% figures now stand only as a historical note.

  • 2026-11-20

    This site's topic page goes live (this page). As of publication: Artificial Analysis, LiveBench and LMArena have not listed MAI Code 2; Microsoft's model card and GitHub's pricing page are published (about 80B active, 1.05M context, $0.50/$2.50 promo to May 17, 2027); no systematic tier-1 Chinese tech-media coverage yet.

06Official specs

As published by Microsoft (Microsoft AI product post, model card and GitHub changelog, Nov 17, 2026). Undisclosed items are written as "not disclosed", not guessed.

Status
Released (GA, 2026-11-17, live in GitHub Copilot)
Model ID
MAI-Code-2 / MAI-Code-2-Flash (GitHub Copilot identifiers)
Positioning
Everyday enterprise coding: production code, deep debugging, multi-file changes, repo-wide reasoning, UI work; also tuned for dashboards and visualizations, per Microsoft
Architecture / parameters
Decoder-only transformer with sparse MoE layers and an integrated vision encoder; MAI Code 2 about 80B active (~2.8T total), Flash about 12B active (~398B total); base MAI-Base-2; trained from scratch by Microsoft (model card)
Context window
1,050,000 tokens (max input 922,000, max output 128,000); the 1M window works only in VS Code and Copilot CLI
Knowledge cutoff
Data before August 2026 (model card)
Input / output
Inputs: text and images (screenshots, UI mockups, charts and diagrams); outputs: text (model card). Optimized for English code production only at release
Out of scope
Microsoft states plainly: both models are explicitly not yet tuned for fundamental science research or very-long-horizon agentic work
Access
GitHub Copilot (VS Code, Visual Studio, Copilot CLI, cloud agent, GitHub.com, mobile, JetBrains, Xcode, Eclipse); MAI Code 2 Flash also in limited-access preview in Microsoft Foundry; MAI Code 2 has joined Copilot Fusion Standard
Plans
MAI Code 2: Pro, Pro+, Max, Business, Enterprise; MAI Code 2 Flash: those plans plus Free and Student; both models only for eligible Copilot subscribers on annual plans (GitHub changelog wording)
Billing
Usage-based inside GitHub Copilot; MAI Code 2 $0.50/$0.025/$2.50 per 1M tokens (promo to May 17, 2027, then $1.00/$0.05/$5.00), Flash $0.05/$0.005/$0.25 (then $0.10/$0.01/$0.50); 1 AI credit = $0.01
Weights / licence
Closed, no public weights
Benchmarks
Microsoft-reported: a four-benchmark average of 98% of GPT-6.1 Sol at about 40% of the cost (Microsoft-chosen benchmarks, 50% promo-price basis), and 96% of Claude Opus 5.5 at about 5% of the cost (same basis); no independent score
Chip
All seven MAI models run on Microsoft's second-generation Maia 200 chip, now in production

07How to try it: four routes

MAI Code 2 has exactly one official route: GitHub Copilot. No standalone API, no weights, no free tier — free users get Flash only, not MAI Code 2 itself.

1VS Code / Visual Studio (main route)

Pick MAI Code 2 in the Copilot Chat model picker (Pro, Pro+, Max, Business, Enterprise) or MAI Code 2 Flash (which also covers Free and Student). Suited to the full everyday coding flow: production code, deep debugging, multi-file changes, repo-wide reasoning, UI work; Flash suits code explanations, targeted edits, routine fixes and rapid iteration. See our Copilot guide.

2Copilot CLI and cloud agent

Call it from the terminal with the Copilot CLI, or use the cloud agent on GitHub.com and mobile: suited to agent-style coding, with multi-file changes and repo-wide reasoning as its headline use; note Microsoft says it is not tuned for very-long-horizon agentic work, so break long tasks into stages and check each one.

3Copilot Fusion Standard

GitHub's Nov 17 changelog confirms MAI-Code-2 has joined Copilot Fusion Standard. Fusion routes tasks across several models, so picking Standard may land you on MAI Code 2; to pin it down, pick the model explicitly in the model picker.

4Microsoft Foundry (Flash only, limited preview)

MAI Code 2 Flash is also in limited-access preview in Microsoft Foundry; MAI Code 2 itself is not in Foundry. Sources disagree on Foundry access (the Microsoft keynote post says both coding models are there; other reporting calls Flash a private preview), so go by what the Foundry console actually shows.

VS Code / Visual Studio: Pick MAI Code 2 in the model picker (Pro / Pro+ / Max / Business / Enterprise) Copilot CLI: Call from the terminal, usage-based billing Cloud agent: GitHub.com, mobile, JetBrains, Xcode, Eclipse Flash: Also on Free and Student; limited preview in Foundry MAI Code 2 itself: No free tier, no standalone API, no public weights

Also: MAI-Code-1.1-Flash retires across all Copilot experiences on Dec 18, 2026; GitHub recommends switching to the latest MAI Code Flash model. Business and Enterprise admins must allow MAI models in the Copilot model policy (enabled by default for Business and Enterprise).

Try-it warning: Free users cannot use MAI Code 2 itself, only Flash; no public weights, so there is no self-hosting route; both models bill at a 50% promotional price, so the bill changes when the promotion ends — don't treat the promo price as the long-term price; Microsoft says both models are not tuned for fundamental science research or very-long-horizon agentic work, so don't rely on them for those tasks; on launch day GitHub Copilot is the only route — don't look for its API elsewhere.

08Pricing

GitHub's pricing page publishes the rate card — $0.50/$2.50 per 1M tokens, promotional until May 17, 2027, then $1.00/$5.00. What follows is the rate card plus an analysis of the "relative cost" figures themselves: how they were computed, what they hide, and what may change when the promotion ends. Prices checked Nov 20, 2026; the checkout page price is authoritative.

ComparisonRelative performance (Microsoft-reported)Relative cost (Microsoft-reported)Basis
MAI Code 2 vs GPT-6.1 Sol98%~40%Four-benchmark average; cost at 50% promo price
MAI Code 2 vs Claude Opus 5.596%~5%Four-benchmark average; cost at 50% promo price
MAI Code 2 Flash vs GPT-6 LunaComparable65%Three benchmarks; cost at 50% promo price
MAI Code 2 Flash vs Claude Haiku 5.5Comparable20%Three benchmarks; cost at 50% promo price

A rate card exists — but it is a promo price

GitHub's Copilot pricing page already publishes per-1M-token prices, and it is the actual checkout reference: MAI Code 2 at $0.50 input, $0.025 cached input and $2.50 output, promotional until May 17, 2027, then $1.00/$0.05/$5.00; Flash at $0.05/$0.005/$0.25, then $0.10/$0.01/$0.50. Three things are confirmed: both models bill usage-based inside Copilot; the cost figures in the benchmark comparison are computed at a 50% promotional price; and the promotion ends on May 17, 2027 (GitHub pricing page).

What happens when the promotion ends

Working backwards from Microsoft's own footnote: if the promo price is half the list price, MAI Code 2's relative cost vs GPT-6.1 Sol rises from about 40% to about 80% after the promotion, and vs Claude Opus 5.5 from about 5% to about 10% — still cheaper, but the "40%" figure most often quoted doubles overnight. That is the premise that must travel with every citation of the relative costs, and Axios' launch-day reporting makes the same point: the pricing advantages reflect promotional discounts.

AT&T's "50x" is one customer, not a rate card

Microsoft says a custom AT&T version of MAI Code 2 Flash outperformed GPT-6.1 Sol on standard 5G network analysis while costing 50x less. That is a single-customer anecdote, not a public rate card, and it cannot be converted into general pricing.

Key point: Every use of "98% of the performance at about 40% of the cost" must carry its basis — a Microsoft-chosen four-benchmark average at a 50% promotional price; the same applies to "96% at 5%". The promo price runs to May 17, 2027, then doubles; for budgeting, stress-test against the post-promo relative costs (about 80% / about 10%).

Official pages: Microsoft AI — MAI Code 2 product post · Microsoft AI — MAI Code 2 model card · GitHub Changelog — MAI Code 2 in GitHub Copilot

09First hands-on reviews

All from the first wave, Nov 17–20: there is no independent measurement yet — every score comes from Microsoft, and third parties can only check the wording, not reproduce the scores. Every entry carries a verdict badge; "Read on this site" opens our rewritten deep-dive article, with the original source at the end.

Microsoft AI product post: the source of every benchmark claim

Verdict: Vendor-reported, no independent verification yet

Source: Microsoft AI, Nov 17, 2026 Original →

The post "MAI Code 2 and Code 2 Flash: Everyday coding at scalable prices" is the source of all the numbers: a four-benchmark average of 98% of GPT-6.1 Sol at about 40% of the cost (Microsoft-chosen benchmarks, 50% promo-price basis), and 96% of Claude Opus 5.5 at about 5% of the cost (same basis); Flash comparable to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of the cost. The footnotes also disclose the methodology: all models at max reasoning effort, three tests on Microsoft's internal MSBench platform (via the Copilot CLI, no subagents), with the FrontierCode v1.1 scores and costs from Artificial Analysis; MAI Code 2's cost is shown at its 50% promotional price. The same post states plainly that both models are not tuned for fundamental science research or very-long-horizon agentic work.

Read on this site → (original ModelBrief rewrite)

Microsoft AI Ignite post: the seven-model family and Maia 200

Verdict: Vendor-reported; its blind-test figure once disagreed with the product post, now unified at 64%

Source: Microsoft AI, Nov 17, 2026 Original →

The Ignite post "Seven new world-class models at scalable price-performance" presents the full seven-model family and the Maia 200 chip in production, and repeats the 98% / 40% and 96% / 5% claims. Notably, its blind-test preference figure once differed from the product post's — pages saved around 04:00 JST on Nov 18 showed 67% in the product post and 68% in the keynote post; both have since been corrected to 64%, matching the product post's main text (see sections 12 and 15).

Read on this site → (original ModelBrief rewrite)

GitHub Changelog: the only authority on plans and access

Verdict: Availability confirmed; pricing details still missing

Source: GitHub Changelog, Nov 17, 2026 Original →

The changelog confirms: MAI Code 2 covers Copilot Pro, Pro+, Max, Business and Enterprise, Flash adds Free and Student; both enter the model picker in VS Code, Visual Studio, Copilot CLI, cloud agent, GitHub.com, mobile, JetBrains, Xcode and Eclipse; both are usage-billed and limited to eligible subscribers on annual plans. A second changelog entry the same day confirms MAI-Code-2 joins Copilot Fusion Standard, and announces MAI-Code-1.1-Flash retires on Dec 18. The changelog gives no price figures at all.

Read on this site → (original ModelBrief rewrite)

Axios: the pricing advantage reflects promotional discounts

Verdict: Tier-1 launch-day coverage, with the right caveat

Source: Axios (via superpowerdaily), Nov 17, 2026 Summary →

Axios' launch-day report repeats Microsoft's pricing claims while stating plainly that the pricing advantages reflect promotional discounts. It is the only confirmed tier-1 English media report so far; no systematic tier-1 Chinese tech-media coverage exists yet.

Read on this site → (original ModelBrief rewrite)

Third-party launch-day analyses: all checking the wording, none reproducing the scores

Verdict: Wording checks agree; independent reproduction is zero

Source: edgeconsultancykw, rrtechpulse, pasqualepillitteri, scanradar, Nov 17–18, 2026

Four third-party launch-day analyses share one pattern: they repeat Microsoft's numbers while labelling the comparisons as Microsoft's own and the prices as promotional; none attempts an independent benchmark reproduction. scanradar additionally notes Microsoft says OpenRouter, Vercel and other channels are "coming soon" (wording incomplete, unconfirmed whether live), and that sources disagree on Flash's Foundry access status.

Chinese coverage: still thin three days in

Verdict: Chinese search demand has not formed

Source: This site's research, Nov 18, 2026 (Google Autocomplete, Chinese queries)

The Chinese query "mai code 2" produced no usable autocomplete (only French promo-code noise), and no systematic tier-1 Chinese media coverage has been verified. English demand is the opposite: within 24 hours of launch, comparison-shaped long tails formed — "mai code 2 benchmarks", "mai code 2 vs opus 5.5", "mai code 2 artificial analysis". This topic's Chinese traffic path is the GitHub Copilot user base, not Chinese search.

10Articles: our deep dives

The six articles below are original ModelBrief rewrites and reorganization — restructured with added conclusions and buying advice, with facts and numbers checked against the original sources. Read them on this site; each ends with its original source.

12Benchmarks: all Microsoft-reported, no independent yardstick yet

MAI Code 2 has no independent board reading anywhere: Artificial Analysis appears only in Microsoft's footnotes (the source of the FrontierCode v1.1 scores and costs), not as an independent listing. Everything below is Microsoft-reported — check the methodology before the scores.

Vendor-reported (Microsoft AI product post, Nov 17, 2026)

TestSourceResultLabel
Four-benchmark average vs GPT-6.1 SolMicrosoft AI product post98% of the performance at about 40% of the cost (Microsoft-chosen benchmarks, 50% promo-price basis)Vendor-reported; no independent verification
Four-benchmark average vs Claude Opus 5.5Microsoft AI product post96% of the performance at about 5% of the cost (same basis)Vendor-reported; no independent verification
Front-end blind test vs GPT-6 Sol (258 people)Microsoft AI postsPreference for MAI Code 2: 64% (67%/68% were pre-correction values from Nov 18)Vendor-reported; corrected to 64%, see below
Flash, three benchmarks vs GPT-6 LunaMicrosoft AI product postComparable performance at 65% of the cost (50% promo-price basis)Vendor-reported; the three benchmarks are unnamed
Flash, three benchmarks vs Claude Haiku 5.5Microsoft AI product postComparable performance at 20% of the cost (50% promo-price basis)Vendor-reported; the three benchmarks are unnamed
AT&T custom Flash vs GPT-6.1 SolMicrosoft AI product postWins on standard 5G network analysis at 50x lower costVendor-reported; single-customer anecdote

Methodology: what the four benchmarks are and how they ran

The four benchmarks are DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, which Microsoft says target the everyday enterprise coding tasks it sees in GitHub Copilot. All models ran at max reasoning effort; DeepSWE 1.1, Terminal Bench 2.1 and SWE-Atlas ran on Microsoft's internal MSBench platform via the Copilot CLI with no subagents, while the FrontierCode v1.1 scores and costs come from Artificial Analysis. Average pass rates (from the chart on Microsoft's official blog): MAI Code 2 at 63.2%, GPT-6.1 Sol at 64.1%, Claude Opus 5.5 at 65.3% — MAI Code 2's average is actually the lowest of the three, so "98%" is a relative figure, not an absolute lead. Average cost per task (same chart): MAI Code 2 at $0.309, GPT-6.1 Sol at $0.761, Claude Opus 5.5 at $6.306, with MAI Code 2's cost computed at the 50% promotional price.

The blind-test figure is now settled at 64%: In the 258-person front-end blind test, the share of developers preferring MAI Code 2 over GPT-6 Sol now reads 64% in both the product post and the Ignite post (better on both aesthetics and functionality). Pages saved around 04:00 JST on November 18 showed 67% in the product post and 68% in the keynote post; both have since been corrected to 64%. The 67%/68% figures survive only as a historical footnote — cite 64%.

Full tally: our benchmarks article →

13Comparisons

vs GPT-6.1 Sol (the headline comparison)

DimensionMAI Code 2GPT-6.1 Sol
Four-benchmark average (Microsoft-reported)98% of GPT-6.1 Sol (Microsoft-chosen benchmarks, 50% promo-price basis)100 (reference)
Average pass rate (Microsoft blog chart)63.2%64.1%
Average cost per task (Microsoft blog chart)$0.309 (50% promo-price basis)$0.761
Relative cost~40% (50% promo-price basis)100
Context window1.05M tokens1M tokens
AccessGitHub CopilotChatGPT Work, Codex, API, Bedrock, Copilot
WeightsClosedClosed

The full meaning of "98% of the performance at about 40% of the cost": the average of four Microsoft-chosen benchmarks trails by only 2%, and the cost at the 50% promotional price is about 40%. But on average pass rate MAI Code 2 actually trails GPT-6.1 Sol (63.2% vs 64.1%), and after the promotion the relative cost rises to about 80%. See our comparison article.

vs Claude Opus 5.5 (the boldest claim)

DimensionMAI Code 2Claude Opus 5.5
Four-benchmark average (Microsoft-reported)96% of Opus 5.5 (Microsoft-chosen benchmarks, 50% promo-price basis)100 (reference)
Average pass rate (Microsoft blog chart)63.2%65.3%
Average cost per task (Microsoft blog chart)$0.309 (50% promo-price basis)$6.306
Relative cost~5% (50% promo-price basis)100
Context window1.05M tokens1M tokens
Independent scoreNoneAA Index (see this site's Claude Fable 5.2 topic for the same-generation figures)

"96% of the performance at 5% of the cost" is likewise a Microsoft-chosen benchmark average at a 50% promotional price; on average pass rate MAI Code 2 trails by 2.1 points, and the cost advantage comes from Opus 5.5's own high pricing stacked with the 50% promotion. See our comparison article.

MAI Code 2 vs MAI Code 2 Flash (choosing inside the family)

DimensionMAI Code 2MAI Code 2 Flash
PositioningGeneral-purpose flagship: production code, deep debugging, multi-file changes, repo-wide reasoning, UILightweight tier: code explanations, targeted edits, routine fixes, rapid iteration
Reference models (Microsoft-reported)GPT-6.1 Sol (98%), Claude Opus 5.5 (96%)GPT-6 Luna, Claude Haiku 5.5 (both comparable)
Relative cost (Microsoft-reported, 50% promo-price basis)~40% of Sol / ~5% of Opus 5.565% of Luna / 20% of Haiku 5.5
Copilot plansPro, Pro+, Max, Business, EnterpriseThose plans plus Free and Student
FoundryNot availableLimited preview
An honest caveat: Both sets of performance figures come from Microsoft's own four- and three-benchmark averages, with the benchmarks chosen by Microsoft and three of the four run on Microsoft's internal platform; Flash's three benchmarks are not even named. The promo price makes the cost figures look cheaper — after the promotion, MAI Code 2's relative cost vs GPT-6.1 Sol is about 80% and vs Opus 5.5 about 10%, while Flash vs Luna would be about 130% (i.e. more expensive) and vs Haiku 5.5 about 40%. Read section 08's pricing breakdown before choosing.

Read the deep comparison: MAI Code 2 vs GPT-6.1 Sol →

14What's missing (as of Nov 20)

  • Any independent score: Artificial Analysis, LiveBench and LMArena have not listed MAI Code 2 ("mai code 2 artificial analysis" in autocomplete is search interest, not a listing)
  • Post-promo relative costs: Microsoft has published no figures; this site works them backwards from the footnote at about 80% (vs Sol) and about 10% (vs Opus 5.5), with the about-$0.62 per-task figure matching Microsoft's own numbers (see the pricing guide)
  • The default context size on Copilot clients other than VS Code and Copilot CLI: GitHub does not state it; the 1M window works only in VS Code and Copilot CLI
  • Speed and accuracy at a full 1M context: Microsoft has not published them; the model card gives only the upper limits
  • Per-benchmark scores for GPT-6.1 Sol and Opus 5.5: Microsoft has not published the rivals' individual scores, so the 98%/96% averages cannot be checked item by item (MAI Code 2's own scores are in the model card's main table and our benchmarks article)
  • The names of Flash's three benchmarks: Microsoft did not name them, so the "comparable to Luna and Haiku 5.5" wording cannot be checked
  • Whether the 64% blind-test figure changes again: Microsoft corrected both posts to 64% on Nov 18 (67%/68% were the earlier values — see section 12); any further change is still to be watched
  • Systematic mainland Chinese tech-media coverage (three days after launch; Chinese coverage is still thin)
  • A consistent Foundry access statement (sources disagree) and confirmation of whether OpenRouter, Vercel and other channels are live yet

15Rumor check

Most confusion about MAI Code 2 comes from seven places: the list price, the blind-test figures, the free tier, open weights, the AA listing, Flash's status and what "98%" means — clarified one by one below.

True, with caveats "MAI Code 2 has a list price of so many dollars per 1M tokens"

GitHub's Copilot pricing page has published the prices: MAI Code 2 at $0.50 input, $0.025 cached input and $2.50 output per 1M tokens, promotional until May 17, 2027, then $1.00/$0.05/$5.00; Flash at $0.05/$0.005/$0.25, then $0.10/$0.01/$0.50. Two caveats: the prices come from GitHub's docs, not Microsoft's product post; and they are 50% promotional prices, not long-term list prices — Microsoft's relative-cost claims (about 40% of GPT-6.1 Sol, about 5% of Opus 5.5) are computed at the promo price.

Resolved "64% of developers preferred MAI Code 2 in the blind test"

Microsoft's posts now both give 64% (a 258-person blind test, vs GPT-6 Sol, better on both aesthetics and functionality). The 67% and 68% figures were the values on the pages as saved around 04:00 JST on November 18 — 67% in the product post, 68% in the keynote post — and both have since been corrected to 64%. Cite 64%; treat 67%/68% as a historical footnote.

Not true "Free users can use MAI Code 2"

MAI Code 2 covers Copilot Pro, Pro+, Max, Business and Enterprise only; Free and Student get MAI Code 2 Flash. Don't build a topic on "mai code 2 free".

Not true "There are open weights for self-hosting"

MAI Code 2 is closed with no public weights; access is GitHub Copilot only, with no self-hosting route.

Not true "MAI Code 2 is listed on Artificial Analysis"

"mai code 2 artificial analysis" is only search interest in autocomplete; AA has not listed the model, and the FrontierCode v1.1 scores and costs come from AA data cited in Microsoft's footnotes, not from an independent AA listing.

Not true "MAI Code 2 Flash is a separate topic"

Flash is MAI Code 2's lightweight sibling, covered inside this topic, not as its own topic. Its three benchmarks are unnamed, and it targets high-volume everyday tasks, not a flagship substitute.

Partly true "MAI Code 2 has 98% of GPT-6.1 Sol's performance at 40% of the cost"

The figures come from Microsoft's official post, but two premises are inseparable from them: the performance is an average of four Microsoft-chosen benchmarks (and on average pass rate MAI Code 2 actually trails GPT-6.1 Sol, 63.2% vs 64.1%), and the cost is computed at a 50% promotional price (after the promotion the relative cost is about 80%). Without the premises, the sentence misleads.

16FAQ

When was MAI Code 2 released?

November 17, 2026 (Tuesday), at Ignite 2026, live in GitHub Copilot the same day; six other MAI models launched the same day, including MAI Code 2 Flash. See section 05 for the timeline.

How much does MAI Code 2 cost?

There is a rate card, from GitHub's Copilot pricing page: MAI Code 2 at $0.50 input, $0.025 cached input and $2.50 output per 1M tokens, promotional until May 17, 2027, then $1.00/$0.05/$5.00; Flash at $0.05/$0.005/$0.25, then $0.10/$0.01/$0.50. Microsoft's own relative-cost claims are 98% of GPT-6.1 Sol's performance at about 40% of the cost (Microsoft-chosen benchmarks, 50% promo-price basis), and 96% of Claude Opus 5.5 at about 5% of the cost (same basis); both models bill usage-based inside Copilot. See our pricing guide.

Where can I use MAI Code 2?

Only in GitHub Copilot: VS Code, Visual Studio, Copilot CLI, cloud agent, GitHub.com, mobile, JetBrains, Xcode, Eclipse. MAI Code 2 covers Copilot Pro, Pro+, Max, Business and Enterprise; Flash adds Free and Student. See our Copilot guide.

Is MAI Code 2 better than GPT-6.1 Sol?

On Microsoft's own four-benchmark average it reaches 98% of GPT-6.1 Sol (Microsoft-chosen benchmarks, 50% promo-price basis) at about 40% of the cost; but on average pass rate it actually trails slightly (63.2% vs 64.1%), and all figures are Microsoft-reported with no independent verification. See our comparison article.

Is MAI Code 2 open source?

No. MAI Code 2 is closed with no public weights; access is GitHub Copilot only, with no self-hosting route.

How large is MAI Code 2's context window?

1,050,000 tokens per the model card (max input 922,000, max output 128,000); the 1M window works only in VS Code and Copilot CLI. Parameters are about 80B active (~2.8T total) and the knowledge cutoff is data before August 2026 — both from the model card.

Can MAI Code 2 do fundamental science research?

Microsoft says not to count on it: both models are "explicitly not yet tuned" for fundamental science research or very-long-horizon agentic work; they target everyday enterprise coding.

MAI Code 2 or MAI Code 2 Flash?

Pick MAI Code 2 for complex work: production code, deep debugging, multi-file changes, repo-wide reasoning, UI work; pick Flash for high-volume everyday work — code explanations, targeted edits, routine fixes, rapid iteration — and Flash is the only one on the Free and Student plans. See our Copilot guide.

17Sources

Every fact on this page comes from the public sources below (the Microsoft AI product post and GitHub changelog published Nov 17, 2026; this site's research on Nov 18, 2026, fact-checked Nov 20, 2026).

Official

English reporting and analysis

Chinese

  • This site's research, Nov 18, 2026: Google Autocomplete produced no usable Chinese suggestions, and no systematic tier-1 Chinese tech-media coverage was found Nov 18

More indexes: All Microsoft models · All Language Models · All models