Microsoft's MAI Code 2 undercuts GPT-6.1 Sol — until May 2027

Microsoft used Ignite this week to put a price on its own models rather than its partner's. The numbers are real, the discount is temporary, and the benchmarks were run by the seller.
Microsoft says MAI Code 2 delivers 98% of GPT-6.1 Sol's performance for 40% of the cost — a comparison computed at a 50% launch discount that ends on May 17, 2027. The coding model is one of seven Microsoft AI shipped at the conference: MAI Code 2 and Code 2 Flash, MAI Cyber 2 Flash, MAI Image 2.7, MAI Realtime 1, Dragon Copilot Transcribe and Decision 1.1, per Microsoft's Ignite announcement. All seven are aimed at the same buyer: someone running agents around the clock and watching the token bill rather than the leaderboard.
The quality claim rests on four coding benchmarks — DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1 — where average success came in at 63.2% for MAI Code 2, 64.1% for GPT-6.1 Sol and 65.3% for Claude Opus 5.5. Three of those four ran on Microsoft's internal MSBench platform through the GitHub Copilot command line, which GeekWire fairly summarized as "on benchmarks Microsoft chose." Against Opus 5.5 the pitch is starker still: 96% of the performance at 5% of the cost.
Then the money. In GitHub Copilot, MAI Code 2 costs $0.50 per million input tokens and $2.50 per million output tokens through May 17, 2027, then steps up to $1 and $5, according to GitHub's Copilot billing table as reported by InfoWorld. Code 2 Flash runs $0.05 and $0.25, doubling to $0.10 and $0.50 on the same date. Microsoft's cost-per-task figures — $0.31 for its own model, $0.76 for Sol, $6.31 for Opus 5.5 — are computed at the promotional rate. Sanchit Vir Gogia of Greyhound Research told InfoWorld the gap "narrows sharply once the promotion ends," because what Microsoft is comparing is the cost of a benchmark run, not a published per-token price.
That is the honest read of the launch: Microsoft has priced a model it can afford to sell cheaply in order to win the default slot in Copilot, and the durable advantage in that fight is distribution, not quality. Its own documentation concedes the code models are not tuned for foundational research or long-horizon agentic work — the tasks where the frontier labs still justify their prices.
The rest of the batch follows the same logic: narrow jobs, priced to undercut.
MAI Cyber 2 Flash solves 92.3% of the CyberGym-E2E vulnerability benchmark against 91.5% for GPT-6 Astra, at half the reasoning budget. It costs $0.10 per million input tokens and $0.50 per million output tokens against $10 and $50 for Astra — $0.45 per task versus $34.38. Microsoft pulled 92 of the 905 CyberGym tasks after finding they overlapped its own training data, then scored every model on the remaining 813, and it is still comparing against a GPT-6 Cyber it does not have numbers for yet. The model ships only inside Microsoft's security products; in MDASH it replaces the previous MAI Cyber for eligible customers from mid-December.
MAI Image 2.7 leads Artificial Analysis's image-editing leaderboard at 1,215 Elo, four points ahead of GPT Image 2.6 Sunburst at roughly a fifth of the price: about $35 per 1,000 images, with the Flash variant near $10. Decision 1.1 handles routing and classification at 68 points on the Decision Index v2.1, six above its predecessor, for $0.042 per million input tokens with output free. It is built on Google's Gemma 4 26B A4B — Microsoft's routing layer now runs on a competitor's open weights. MAI Realtime 1, the company's first speech-to-speech model, costs one cent per audio minute and accepts OpenAI's Realtime API format, so switching is a configuration change rather than a rewrite.
Separately, MAI Licensed Weights lets enterprises fine-tune selected models on their own data inside their own Azure tenancy. There is no launch date and no price; Microsoft's guidance is to contact an account team. A customized AT&T model for 5G network analysis reportedly beat GPT-6.1 Sol at roughly a fiftieth of the cost — a customer claim Microsoft is repeating, not a measurement anyone outside the two companies can check.
One detail worth keeping in view: GeekWire notes that Microsoft also brought OpenAI's GPT-6.2 into Foundry on the same stage. The price war is happening inside Microsoft's own catalogue, which is an odd place to be for a company whose AI business was built on reselling OpenAI — and a good one for buyers, who now get the cheap in-house model or the expensive frontier model off the same invoice. We covered the frontier end of that menu when it landed yesterday — GPT-6.2 Astra pricing: $8 in, $40 out per million tokens.
Will Wright's next game is called MyMems, and it turns a spoken memory into a playable world. Gallium Studios, the indie Wright co-founded with Carmen Sandiego creator Lauren Elliott, announced the mobile title for 2027: you tell the game a short story about your life by text or voice, and it becomes an object, a place, or a character modelled on someone you know. "We want to take your autobiography and turn it into a toy world," Wright told The Verge.
The AI is not a frontier model, which is the interesting part. Gallium licensed a "Behavior Engine" from Artificial Agency, an Edmonton startup founded by former Google DeepMind researchers, and pairs it with its own personal-behavior model built for the game; executive producer Virginia McArthur told The Verge, "We're not making slop." Input is text and voice rather than photos, despite the framing of some coverage, and the claim that memories are learned locally "only with your permission" comes from the vendor with no published architecture behind it. There is no demo, no price, no publisher and no named foundation model — but it is Wright's first game since Spore in 2008, and it is a marquee designer betting that small, purpose-built behavior models are the cheaper route to characters that feel alive.
What to watch: whether MAI Code 2's price still looks like 40% of Sol once the May 2027 step-up lands, and whether any independent harness reproduces the 98%.
If a 40% discount is what it takes to move an enterprise off the frontier model, what happens to AI pricing when every lab runs this play? Write to us via our contact page.







