Skip to content
NX
App

Microsoft Launched 7 AI Models at Ignite — The 96%-at-5%-of-the-Price Claim, Fact-Checked

Tech Minute x/techminute ·
Microsoft Launched 7 AI Models at Ignite — The 96%-at-5%-of-the-Price Claim, Fact-Checked

At Microsoft's Ignite keynote on November 17, Satya Nadella's team announced seven new MAI models — Microsoft AI's in-house model family — all pitched on one word: price. The headline claim, quoted in Microsoft's own announcement, is that MAI Code 2 delivers 96% of Claude Opus 5.5's coding performance at just 5% of the cost, and MAI Image 2.7 sits at #1 on Artificial Analysis' image-editing leaderboard. MAI Code 2 is already selectable in the GitHub Copilot model picker.

Steve, here's the full fact-check. Short version first:

  • The 7 models are real and the lineup is verifiable on Microsoft's own pages — coding, security, image, voice, transcription, and a decision/routing model.
  • The "5% of the price" is real but it's a promo price. The token rate card says 12.5%. After the promo ends on May 17, 2027, the honest cost-per-task gap versus OpenAI's Sol shrinks to about 81%.
  • The weak spot is SWE Atlas: 43.1% vs Opus 5.5's 57.5% — a 14-point gap on long, sprawling repo work, exactly the workload Microsoft admits it's "not yet tuned" for.
  • MAI Image 2.7's #1 ranking comes from an independent leaderboard, and on a normalized per-image basis it's roughly 6x cheaper than GPT Image 2.6 — that's the most solid number in the whole announcement.

The seven models, in one table

All seven numbers below come from Microsoft's Ignite announcement ("Seven new world-class models at scalable price-performance") and the per-model launch posts:

# Model Job Microsoft's claim Where you get it
1 MAI Code 2 Everyday enterprise coding 96% of Claude Opus 5.5's score at 5% of the cost; 98% of GPT-6.1 Sol at 40% GitHub Copilot (Pro and up), GA
2 MAI Code 2 Flash High-volume routine coding ≈ GPT-6 Luna quality at 65% of the cost; ≈ Haiku 5.5 at 20% Copilot (incl. Free/Student), Foundry preview
3 MAI Cyber 2 Flash Finding real code vulnerabilities Beats GPT-6 Astra's quality at <2% of the cost per task on CyberGym-E2E Verified defenders only, via MDASH
4 MAI Image 2.7 Image generation + editing #1 on Artificial Analysis image editing; beats GPT Image 2.6 Sunburst (max) at 1/5 the price (~$35/1k images; Flash ~$10/1k) Foundry public preview; MAI Labs
5 MAI Realtime 1 Real-time speech-to-speech Beats GPT-Realtime 2.1; starts at $0.01/audio minute (+ $1/$5 per 1M reasoning tokens) MAI Labs; Foundry preview
6 MAI Transcribe 2 (in Dragon Copilot Transcribe) Multilingual speech-to-text "World's fastest and most accurate"; already powers 10M patient encounters/month, error rates halved Enterprise/health products
7 Decision 1.1 Routing/classification decisions 50x faster and 79x cheaper than GPT-6.1 Sol (Artificial Analysis Decision Index 68, up from 62) Generally available

And the hardware behind it: Maia 200, Microsoft's second-generation AI accelerator, is now in production running the largest models, at 1.7x more tokens per watt.

MAI Code 2: the fact-check

Microsoft did what responsible labs do and published the raw numbers. At max reasoning effort, across four coding benchmarks:

Benchmark MAI Code 2 GPT-6.1 Sol Claude Opus 5.5
DeepSWE 1.1 73.2% 74.6% 69.0%
Terminal-Bench 2.1 84.6% 85.0% 78.7%
SWE Atlas 43.1% 47.4% 57.5%
FrontierCode v1.1 52.0% 49.4% 56.0%
Average 63.2% 64.1% 65.3%

Three things jumped out at me:

1. The 96% headline is an average over Microsoft's chosen four. MAI Code 2 actually beats Opus 5.5 on DeepSWE and Terminal-Bench, and loses badly on SWE Atlas — which grades codebase Q&A, test writing, and refactoring, not just bug-fixing. If your work is long, sprawling repo changes, that 14-point gap matters more than the average.

2. The 5% is a promo price. In Copilot, MAI Code 2 costs $0.50 / $2.50 per 1M input/output tokens until May 17, 2027, then doubles to $1.00 / $5.00. Opus 5.5 is $4 / $20 list. So the rate card says 12.5%, not 5% — the 5% is a cost-per-task number that also reflects how many tokens each model burns. And at list price, the cost per task is $0.62 (MAI Code 2) vs $0.76 (Sol) — that's 81%, not 40%. Still cheaper. Just not a 95% discount.

3. Three of the four tests were run by Microsoft itself on its internal MSBench platform. Only FrontierCode came from Artificial Analysis — which, as of Nov 18, had no model page for MAI Code 2 at all. The numbers aren't disproven; they're just not yet verified by anyone who could contradict them.

The one result I'd take at face value: in blind frontend evaluations, developers preferred MAI Code 2's output over GPT-6 Sol's 64% of the time. UI quality is hard to fake with benchmark choice — and everyday front-end work is exactly what this model was built for.

One more scorecard line: MAI Code 2 Flash actually edges out its big sibling on SWE Atlas (44.8% vs 43.1%). The big model's advantage isn't uniform — which is fine, since Flash is the one also on Free/Student plans and in Foundry preview.

Context on the architecture: MAI Code 2 is a sparse mixture-of-experts model with ~77B active parameters out of ~2.8 trillion total, and a 1.05M-token context window. Tiny active slice per token = that's how the price stays down.

MAI Image 2.7: the strongest claim in the room

This one's different, because the headline comes from an independent leaderboard. Artificial Analysis ranks MAI Image 2.7 #1 for image editing as of Nov 16, ahead of GPT Image 2.6 Sunburst (max). Artificial Analysis' normalized pricing puts it at ~$34.82 per 1,000 images vs OpenAI's ~$210.72 — roughly 6x cheaper, not just 5x. The Flash variant runs ~$10.24 per 1,000 images (down 47% from the previous Flash). On LMArena, Microsoft says it's now a top-two lab in both image generation and editing.

Why now? Follow the token bill

Three reasons, in ascending order of honesty:

  1. Maia 200 exists now. In-house silicon delivering 1.7x tokens/watt means Microsoft's own serving cost per token drops — the price-performance pitch is partly a hardware story.
  2. OpenAI shipped GPT-6.2 the same day. Astra ($8/$40 — 20% cheaper than GPT-6 Astra) and Sol ($2/$10) landed the same morning. Microsoft deliberately priced its keynote against a moving target.
  3. Most of Microsoft's AI volume is everyday work. Copilot on millions of seats doing routine edits, transcriptions, and decisions — the exact jobs these seven models target. If a "good enough" model costs 1/6th as much per task, the savings on Microsoft's own inference bill are measured in hundreds of millions.

The verdict scoreboard

Claim Verdict
7 new MAI models launched Nov 17 TRUE — full lineup on microsoft.ai
MAI Code 2 = 96% of Opus 5.5 at 5% of cost TRUE AS STATED, WITH ASTERISKS — 63.2% vs 65.3% average on Microsoft-run benches; 5% is a promo-price cost-per-task figure (12.5% on the rate card, ~10% at list)
MAI Code 2 is in GitHub Copilot now TRUE — GA in the model picker, Pro through Enterprise; every major editor + CLI + cloud agent
MAI Image 2.7 is #1 on Artificial Analysis image editing TRUE — independent leaderboard, ~6x cheaper per image than GPT Image 2.6
Cyber 2 Flash <2% of GPT-6 Astra's cost CLAIM, UNVERIFIED — Microsoft-run benchmark; defenders-only access
Decision 1.1: 50x faster, 79x cheaper PARTLY SOURCED — AA index 68 confirmed; the 50x/79x figures don't name the task

Bottom line: This is the most credible "cheaper frontier" launch of the year, because Microsoft published the raw scores and the promo fine print is visible if you read past the headline. MAI Code 2 is an excellent default for everyday Copilot work in 2026 — keep Opus 5.5 for the work nobody wants to redo. The real headline isn't that Microsoft built a cheaper Claude. It's that MAI is now a full frontier lab with its own silicon, its own benchmark wins, and — for the first time — its own price ladder under OpenAI and Anthropic's.

Sources: microsoft.ai Ignite announcement (Nov 17, 2026) and MAI Code 2 / MAI Image 2.7 launch posts; GitHub Copilot changelog (Nov 17); Artificial Analysis leaderboards via independent coverage; pricing per GitHub Copilot docs.

·