Briefly
- Microsoft says MAI Code 2 delivers 96% of Claude Opus 5.5’s coding efficiency at 5% of its value, although the comparability consists of promotional pricing.
- Its new cybersecurity mannequin scored 92.3% in opposition to GPT-6 Astra’s 91.5% on a vulnerability-finding take a look at, in keeping with Microsoft’s evaluations.
- MAI Picture 2.7 topped an impartial image-editing leaderboard, whereas our fast comparisons favored its realism, creativity and general output high quality.
Microsoft unveiled a brand new lineup of specialised AI fashions at its Ignite convention Tuesday, making a direct run at Anthropic and OpenAI in coding, cybersecurity, and picture technology.
As a substitute of attempting to construct a single chatbot that beats everybody at all the things, it’s pitching fashions educated to do explicit jobs at considerably decrease value.
The largest declare entails MAI Code 2, which Microsoft says delivers 96% of Claude Opus 5.5’s efficiency on widespread enterprise coding duties for simply 5% of the fee. It additionally claims 98% of GPT-6.1 Sol’s efficiency at 40% of the worth.
OpenAI additionally released GPT-6.2 on Tuesday, so Microsoft’s Sol comparability already issues an older technology.
These figures common outcomes from 4 evaluations centered on sensible software program improvement, together with fixing code and finishing multi-file adjustments. Microsoft ran three of the checks by itself analysis platform, used outdoors outcomes for the fourth, and calculated MAI Code 2’s prices utilizing a 50% promotional low cost.
A lighter Code 2 Flash mannequin targets routine fixes and fast edits, with Microsoft claiming efficiency corresponding to Claude Haiku 5.5 at 20% of the worth, so extraordinarily low-cost with out sacrificing high quality. Each coding fashions are already in GitHub Copilot, placing Microsoft’s in-house expertise straight right into a product that has relied closely on outdoors AI suppliers.
Safety is one other goal. Microsoft says MAI Cyber 2 Flash succeeded on 92.3% of CyberGym-E2E’s evaluated vulnerability-discovery duties, in contrast with 91.5% for GPT-6 Astra, at lower than 2% of the fee per activity. CyberGym-E2E scores the proportion of software-security investigations an AI completes efficiently; Microsoft excluded 92 of 905 duties after figuring out overlap between their supply materials and its coaching knowledge.
The corporate has not but printed a direct comparability in opposition to OpenAI’s restricted GPT-6 Cyber mannequin. Its safety system will start changing an older MAI mannequin for eligible MDASH prospects in mid-December, constructing on the vulnerability-scanning method Decrypt previously covered.
In AI, David may also beat Goliath
Can a small mannequin beat a giant one? Positively—in particular circumstances.
A general-purpose mannequin has to be taught to put in writing poetry, clarify physics, debug software program and reply questions on virtually something. A mannequin tuned for coding or cyber protection can spend extra of its coaching and computational finances on the patterns that matter for that job, slightly than carrying the identical breadth of information into each request.
That does not make smaller fashions universally smarter. It will possibly make them sooner and extra economical when the duty is properly outlined, significantly in a enterprise the place the identical work runs 1000’s of instances a day.
A safety scanner analyzing each new code change doesn’t want an costly frontier mannequin to elucidate the historical past of the Roman Empire. If a mannequin can dedicate its full energy to generate Python code, it can make extra sense to pay for its API calls as an alternative of utilizing an excellent highly effective generalistic mannequin.
Specialization may also enhance the entire system across the mannequin: the instruments it calls, the checks that confirm its solutions and the info used to refine it. OpenAI has taken a associated method with its Mini and Nano fashions, letting cheaper staff deal with routine duties whereas bigger techniques sort out the exceptions.
Nonetheless, there are trade-offs. Microsoft’s personal documentation says Code 2 will not be but optimized for elementary scientific analysis or very long-running autonomous assignments. And Anthropic competes within the cheaper finish of the market with Haiku, not simply its premium Opus fashions, so Microsoft is getting into a race its rivals already perceive.
MAI Picture is the extra convincing demonstration
Microsoft’s MAI Image 2.7 could provide the clearest proof that its technique is working. In our fast side-by-side checks, its photographs beat GPT Picture and Google’s Nano Banana on realism, creativity, and general high quality.

These have been simply first impressions primarily based on a couple of generations, however they broadly observe its sturdy exhibiting in impartial rankings.
On Artificial Analysis’s image-editing test, which asks individuals to decide on between anonymously generated edits and turns these preferences into an Elo score, MAI Picture 2.7 presently leads with 1,215 factors in opposition to GPT Picture 2.6 Sunburst’s 1,211.
The four-point hole is sufficiently small that their reported uncertainty ranges overlap; the separate text-to-image technology rating, scored utilizing the identical preference-based technique, places Microsoft in fourth place.

The pricing is more durable to disregard. Microsoft quotes about $35 per 1,000 photographs for MAI Picture 2.7 and $10 for its Flash model, versus roughly $211 for GPT Picture 2.6 Sunburst at most high quality in Synthetic Evaluation’s comparability. Google’s Nano Banana 2.1 is about $34 per 1,000 there, whereas Nano Banana Professional is about $134; precise payments rely on the supplier, decision and settings.
Our earlier testing of MAI Picture 2 discovered wonderful realism however irritating limits on modifying and output codecs. The newer launch provides interactive modifying via Microsoft’s Canvas workspace, clear backgrounds and broader entry via Foundry, though picture high quality and a usable product are nonetheless two various things. That mentioned the mannequin is kind of censored too.

Microsoft can be providing chosen enterprise prospects the power to customise and run its fashions’ licensed weights inside Azure, lowering reliance on outdoors distributors with out making the expertise open supply. For the following take a look at of its specialist technique, Code 2 is already obtainable in GitHub Copilot, and the Cyber 2 Flash rollout is scheduled to start out in mid-December.
Day by day Debrief E-newsletter
Begin daily with the highest information tales proper now, plus unique options, a podcast, movies and extra.
Be part of 72,000+ readers
You might also like
More from Web3
Cathie Wood’s ARK Sells $85 Million in Nvidia, Buys $40 Million in Crypto Stocks
In short ARK bought 543,092 Nvidia shares, valued at about $84.8 million utilizing Monday's closing worth. The agency purchased roughly $39.9 …
Ethereum Devs Weigh a Scaling Upgrade That Could Break 15,000 Apps
In short Ethereum’s proposed Glamsterdam improve might break as much as 15,000 sensible contracts, EF developer Marius van der Wijden …
Standard Life Chair Wants Tokenized Gilts to Open UK Government Bonds to Retail Savers
Briefly Normal Life chairman Sir Nicholas Lyons stated tokenized gilts may make it simpler for extraordinary savers to purchase UK …





