← Back to all posts
News

Microsoft's MAI Code 2 Hits GitHub Copilot at 96% of Opus 5.5 for 5% of the Cost

November 18, 2026 · 05:14 UTC · News
On this page 9 sections
Microsoft's MAI Code 2 Hits GitHub Copilot at 96% of Opus 5.5 for 5% of the Cost

TL;DR

MAI Code 2 and MAI Code 2 Flash, the coding models Microsoft announced at Ignite on 17 November, are live in GitHub Copilot right now. Code 2 runs at $0.50 per million input tokens and $2.50 per million output tokens, promotional until 17 May 2027. Flash goes for $0.05 and $0.25. Microsoft's own benchmark average claims 98 percent of OpenAI's GPT-6.1 Sol at 40 percent of the cost per task, and 96 percent of Claude Opus 5.5 at 5 percent. The scope is in writing: everyday enterprise coding, explicitly not long-horizon agentic work.


At a glance
Code 2, by the invoice
$0.50/$2.50
promo input and output per mtok, code 2 in copilot
96%
of opus 5.5 coding performance, at 5% of cost per task
98%
of gpt-6.1 sol performance, at 40% of cost
64%
blind frontend preference vs gpt-6 sol, 258 reviewers
7
new mai models shipped at ignite, code 2 among them
2x
both list prices when the promo ends 17 may 2027

What a dollar buys now

The prices sit in Copilot's public price table, billed as AI credits where one credit is $0.01. Code 2: $0.50 input, $0.025 cached input, $2.50 output per million tokens, until 17 May 2027, after which it goes to $1.00/$0.05/$5.00. Flash: $0.05/$0.005/$0.25, doubling to $0.10/$0.01/$0.50 the same day. Someone in Redmond named the flagship "Code 2" and the fast one "Code 2 Flash", then went to lunch.

Context from the same table: GPT-6.1 Sol and GPT-6.2 Sol run $2 in and $10 out. Claude Opus 5.5 is $4/$20, Claude Fable 5.2 and GPT-6.2 Astra are $8/$40. At the bottom, GPT-6 Luna and Claude Haiku 5.5 sit at $0.10/$0.50. Code 2 slots between the cheap tier and the Sol tier while claiming Sol-adjacent quality, and Flash undercuts the previous generation of Microsoft's own model: MAI-Code-1.1-Flash was $0.20/$1.20, so the new Flash charges a quarter of the input price and a fifth of the output price of its predecessor.

Copilot output price, $ per mtok
Code 2 undercuts the Sol tier by 4x
same price tablemai code 2, promo
opus 5.5
$20
6.1 sol
$10
mai code 2
$2.50
code 2 flash
$0.25
Prices from GitHub's Copilot model table, 17 November 2026. The copper dot marks Code 2's $5.00 list price after the promotion ends on 17 May 2027.

Who graded the benchmarks

The headline ratios are averages over four suites Microsoft picked to look like Copilot traffic: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, all run at max reasoning effort. Three of the four ran on MSBench, Microsoft's internal eval platform, driven by the GitHub Copilot CLI without subagents. FrontierCode scores and costs come from Artificial Analysis, the one external grader in the set, and Microsoft's side of that comparison uses the promotional price.

On blind frontend evaluations, 258 annotators preferred Code 2's interfaces, games, data visualizations, 3D scenes and SVGs over GPT-6 Sol's 64 percent of the time, judging aesthetics and function together. Flash is claimed to match GPT-6 Luna at 65 percent of its cost and Claude Haiku 5.5 at 20 percent, averaged over three of the same suites. Microsoft also says Code 2 delivers "changes code maintainers would merge", which is the metric that actually shows up in your review queue.

The scoping is unusually blunt for a launch post. In Microsoft's words, the models are "explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work." The New Stack read the launch as a specialist bet against generalist frontier models, and noted that a 5 percent cost ratio says as much about Anthropic's pricing as Microsoft's.

Every surface Copilot has

Code 2 is available on Copilot Pro, Pro+, Max, Business and Enterprise; Flash also reaches Copilot Free and Student. The model picker carries both across VS Code, Visual Studio, the Copilot CLI, github.com, the cloud agent, the Copilot app, JetBrains, Xcode, Eclipse and the mobile apps. If you administer a Business or Enterprise plan, note the default: new models enable automatically unless a policy turned the global default off.

Outside Copilot, Code 2 Flash is in limited-access preview in Microsoft Foundry, with OpenRouter, Vercel, Hermes, Open Code and Kilo listed as coming. Microsoft also ships a derivative, MAI Work Flash, inside its own products: Excel and PowerPoint run it at parity with GPT-6 Luna on their internal scenario benchmarks, Word runs it close to GPT-6.1 Sol using about 42 percent fewer input tokens per call, Teams beats GPT-6 Luna with up to 28 percent fewer tokens, and in FabricIQ it beats GPT-5.6-sol, GPT-6-astra and Opus-5.5 at DAX query generation.

Rollout
Where Code 2 landed on day one
Copilot
Both models, all pickers
flash reaches free and student plans
Foundry
Flash, limited preview
code 2 flash only, apply for access
Third parties
OpenRouter, Vercel, more
hermes, open code and kilo listed as soon
Inside MSFT
Work Flash in M365
excel, word, teams and fabriciq already run it
Availability per the GitHub Copilot changelog and Microsoft's announcement, 17 November 2026.

Six more models shipped the same morning

Code 2 was one of seven models in the Ignite drop. MAI Cyber 2 Flash claims GPT-6 Astra's quality at under 2 percent of the cost per task on CyberGymE2E, with output tokens 86 percent cheaper than MAI Cyber 1.1 Flash; Microsoft says MAI Cyber models already account for almost 40 percent of scans in MDASH. MAI Image 2.7 took the number one spot on Artificial Analysis image editing, ahead of GPT Image 2.6 Sunburst (max) at a fifth of the price, and sits fourth on text-to-image. MAI Realtime 1, a speech-to-speech model, claims wins over GPT-Realtime 2.1 starting at $0.01 per audio minute. MAI Transcribe 2 already powers Dragon Copilot across 10 million patient encounters a month with multilingual error rates halved, in the hands of a third of US physicians. Decision 1.1 is claimed at 50x faster and 79x cheaper than GPT-Sol 6.1, and the Maia 200 accelerator is in production at 1.7x tokens per watt.

The Ignite 2026 model drop
Seven models, one price thesis
code 2code 2 flashcyber 2 flashimage 2.7realtime 1transcribe 2decision 1.1
Every launch in the set leads with a cost ratio against an OpenAI or Anthropic comparator. Source: Microsoft AI, 17 November 2026.

Licensed weights, the quieter announcement

Buried in the same drop: MAI Licensed Weights. Take a frontier MAI model, customize it on your proprietary data, then run it inside your own Azure environment with access controls you own. Delivery goes through a partner Microsoft calls Frontier Company. The pitch against open weights is provenance: "if you build on open weights, can you really trust where they came from." The named reference customer is AT&T, whose custom model for 5G network operation analytics Microsoft says outperformed GPT-6.1 Sol at 50x lower cost.

Read the fine print before you switch defaults

The 5 percent and 40 percent figures are cost per task, and the unit matters. Per-token price is the price per ounce of flour; cost per task is the cake. Microsoft's cake number mixes cheaper tokens with fewer tokens burned per task, measured on FrontierCode by Artificial Analysis. Per output token in Copilot, Code 2's promo price is 12.5 percent of Opus 5.5's rate, $2.50 against $20. Absurdly cheap either way, just not sticker-5-percent.

Three more footnotes travel with the claims. The ratios use promotional pricing, so they double at list after 17 May 2027. Three of the four suites ran on Microsoft's own platform with a harness the vendor picked, without subagents. And the averages cover tasks shaped like Copilot traffic, which is the exact workload Microsoft says it tuned for. Run your own evals on your own repo before you change a default; that advice predates this launch and survives it.

Why a builder should care

  • The cheap tier inside Copilot just got a credible resident. Flash at $0.05/$0.25 on plans that include credits makes high-volume loops like code explanations, targeted edits and routine fixes close to free for anyone already paying for Copilot.
  • Evaluate against list prices, not the intro rate. Any default you pick this week needs to still make sense at $1/$5 for Code 2 and $0.10/$0.50 for Flash after May 2027.
  • Match the model to the horizon. Microsoft scoped these models to everyday coding in its own launch post; if your harness runs 40-step agent marathons, the vendor has told you to test elsewhere first.
  • Org admins get these models by default. If you manage a Business or Enterprise policy and do not want Copilot routing traffic to Microsoft's models, the setting to check is the model policy in Copilot settings.
  • If you sell developer tooling on OpenAI or Anthropic APIs, the platform vendor now sets reference prices inside its own IDE distribution. Watch what Code 2 lists for when OpenRouter and Vercel carry it.

Key Takeaways

  • On 17 November 2026 at Ignite, Microsoft released MAI Code 2 and MAI Code 2 Flash, live in GitHub Copilot the same day, with Code 2 Flash also in limited-access preview in Microsoft Foundry and OpenRouter, Vercel, Hermes, Open Code and Kilo listed as coming.
  • Copilot pricing: Code 2 at $0.50 input, $0.025 cached, $2.50 output per million tokens promotional until 17 May 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 average over DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1: 98 percent of GPT-6.1 Sol at 40 percent of the cost per task, and 96 percent of Claude Opus 5.5 at 5 percent. Three suites ran on Microsoft's MSBench with the Copilot CLI and no subagents; FrontierCode numbers come from Artificial Analysis.
  • In blind evaluation, 258 annotators preferred Code 2's frontend work over GPT-6 Sol's 64 percent of the time. Flash is claimed to match GPT-6 Luna at 65 percent of its cost and Claude Haiku 5.5 at 20 percent.
  • The models are "explicitly not yet tuned" for fundamental science research or very long horizon agentic work, per Microsoft's own announcement, and the cost ratios are cost per task at promotional prices, not per-token list prices.
  • The wider Ignite drop included MAI Cyber 2 Flash, MAI Image 2.7 at number one on Artificial Analysis image editing, MAI Realtime 1 from $0.01 per audio minute, MAI Transcribe 2 in Dragon Copilot, Decision 1.1, and MAI Licensed Weights for running customized frontier models in your own Azure environment.

Sources: Microsoft AI, GitHub Copilot changelog, Copilot models and pricing, Ignite 2026 model lineup, The New Stack

AIMicrosoftMAI Code 2GitHub CopilotCoding ModelsLLM PricingIgnite 2026Microsoft MAI
CONSOLE
$