Seven new world Class models at scalable price performance | Microsoft AI
Microsoft’s seven-model MAI launch makes a strong price-performance argument, but its real significance depends on whether the vendor-reported comparisons survive independent testing and production workloads. The practical promise is a portfolio of cheaper, specialized defaults rather than one universal frontier winner.
Transcript
Edmund Microsoft says MAI Cyber 2 Flash matches GPT-6 Astra's quality for less than two percent of the cost per task. Which is a very loud way of saying agents can run all day now.
Geffen Maybe. The word doing the work there is task. That number comes from one security benchmark, not some universal unit of intelligence.
Edmund Sure, but the argument's bigger than one model. Microsoft's launching seven models at once. Coding, cyber, images, audio, transcription, decision work.
Geffen And the strategy's pretty legible, right? Make the cheap specialist the default, keep the expensive model for the ugly cases.
Edmund Which is a place we keep ending up, honestly. The model menu matters less than whether a workflow can afford to keep calling one.
Geffen That's the actual pressure, yeah.
Edmund So the evidence is a stack of comparisons. MAI Code 2, ninety-eight percent of GPT-6.1 Sol's coding quality, for about forty percent of the price.
Geffen Mm.
Edmund And against Opus 5.5, ninety-six percent of the performance for FIVE percent of the price.
Geffen Which sounds insane until you look at the denominator. The supplementary numbers put Code 2 at sixty-three point two percent across four coding tests. Sol's at sixty-four point one, Opus at sixty-five point three. Close… not identical.
Edmund Close might be exactly what procurement needs, though. If the code still gets merged, are the extra few points worth paying five or ten times as much?
Geffen That's a product claim, not something the benchmark proves.
Edmund Granted. Cyber's the sharper one anyway.
Geffen So, Cyber 2 Flash. Ninety-two point three percent task success on CyberGym-E2E, Astra's at ninety-one point five. And the outside summary says it used half the reasoning budget.
Edmund Wait, half?
Geffen Half. Which, fine, I'll admit that combination got me. Similar or better success, less thinking, less money.
Edmund That changes the shape of a security agent. It can inspect more code, retry more often, leave a human a better trail, without every run becoming a budget event.
Geffen With the giant caveat that these are Microsoft's own comparisons. Their pricing, their test setup. And we're doing a read-it, reacted-to-it episode here, not a deep research dive, so a version or a number could already be behind us. Starting point, not gospel.
Edmund Which goes double for the hardware claim.
Geffen Maia 200. Supposedly one point seven times more tokens per watt, and the article doesn't say against which systems, or give a reproducible method.
Edmund So why does it matter at all?
Geffen Because it's the mechanism under the whole announcement. If you control the models, the serving infrastructure AND your own accelerator, you can optimize the full path. Price-performance here is partly a model achievement and partly a systems integration achievement.
Edmund And that's why seven launches at once reads as one move. MAI Realtime 1 starts at one cent per audio minute. Decision 1.1 is claimed fifty times faster and seventy-nine percent cheaper than GPT-6.1 Sol. Workload-specific economics, all the way down.
Geffen They're not interchangeable wins, though. The transcription one, serving ten million patient encounters a month, that's evidence of deployment. It says nothing about a different language mix or a noisy room.
Edmund Yeah.
Edmund See, the deployment detail is the part I actually care about. Dragon Copilot Transcribe reportedly halves multilingual error rates in an existing workflow. A model matters when someone's wrapped it in a product people use, not when it merely occupies a card.
Geffen And I want the failure distribution, because average accuracy can look great while one smaller language carries most of the operational pain. We're not settling that one today.
Edmund No. So, who should care? Persistent agents, high-volume coding, security analysis, live audio. Those teams have a reason to look closely.
Geffen And a team doing occasional chat won't feel a two-percent price story at all. What everyone actually wants is cost per successful outcome, after retries, latency, review, failures. Microsoft's headline points that direction, but it hasn't published enough to close the case.
Edmund My read is modest but real. If these ship with stable limits and the external tests hold up, they become the cheap first call in a bunch of workflows. Not the best model in the abstract. The one you can afford to keep on.
Geffen Same place, with an asterisk. Independent evals have to separate model quality from the serving stack and the reasoning budget. Otherwise we're comparing whole recipes and calling them models.
Edmund Episode eleven eighty-two ends with us agreeing, which is mildly suspicious. I'll leave the benchmark spreadsheet open so you can find one more missing denominator.
Geffen I was already looking for it.
Edmund Of course you were. That's enough internet for today.