Three Frontier Model Launches in Three Days

Three frontier launches in three days

The first week of September is usually quiet. This year's was not. Three labs shipped flagship models on three consecutive days, and the week ended with a live outage across several major AI services and a launch page that went up, came down, and reappeared ninety minutes later.

Once the noise settles, the interesting part is not who won. It is that the three releases are optimising for genuinely different things, and the usual single-number leaderboard cannot see the difference.

Tuesday: reasoning depth, shown through case studies rather than charts

Anthropic opened with Fable and Mythos 5.1, positioned as their strongest models for coding and knowledge work. The headline benchmark deltas were modest. The case studies were not.

• A quantitative trading firm had carried a bug for five years that crashed roughly once in a million runs, the kind of failure that survives every attempt to reproduce it. Handed the crash dump, the model followed the faulting address into a compiled vendor library it had no source for, disassembled that library, and traced the fault to a bug in the vendor's own code.

• On protein binder design, the reported hit rate for producing a molecule that actually sticks to its target moved from around 10% to around 50%.

• It trained a network on thirty-year-old NASA radar data to produce a new elevation map for part of the surface of Venus.

These are lab-reported results and should be read as such. But the shape of the first one is worth sitting with: the useful work was not writing code, it was reverse engineering a black box nobody on the team could open.

Wednesday: the price of your data, itemised

Meta followed with Muse Spark 1.3, its fourth release in five months. The benchmarks were respectable. The pricing was the actual news: roughly $1.25 per million input tokens and $4.25 out on the standard endpoint, or 10 cents in and 20 cents out on a contributor tier where you agree to let Meta train on everything you send.

Reportedly a double-digit percentage of developers are taking the cheap tier. That is a twelvefold discount in exchange for your codebase, now an explicit published line item rather than something buried in terms of service.

Training rights on your source code have a public market price. Anyone shipping client work needs an answer to that before a developer picks the cheaper endpoint on their behalf.

Thursday: computer use, and a threshold nobody wanted to cross first

Several major AI services went down simultaneously on Thursday morning, most plausibly an upstream cloud incident rather than anything more interesting. Then OpenAI announced GPT-6 Astra, pulled the announcement, restored it, and confirmed the model was not actually available yet.

The pitch is computer use: operating a desktop the way a person does, filling in forms, working spreadsheets, driving engineering tools like KiCad and Blender. On OSWorld, which drops a model into a real desktop and grades it on office work, it scored 73% at about forty minutes per task, against 65% at seventy-five minutes for the previous best. The demos leaned hard on spatial work, including a modelled house exported into a walkable Unreal 5 scene.

Self-reported scores include 100% on an exploit benchmark and 99% on ARC-AGI-3, and deserve the scepticism any first-party number does. One accompanying claim is harder to wave off: this is described as the first model to reach the critical cyber threshold in its own preparedness framework, meaning it can find and exploit previously unknown vulnerabilities without a human directing it. Pricing lands at $10 per million in and $50 out.

And yet an independent intelligence index put it at 61, level with the previous generation and a few points behind Tuesday's release. Both things are true. General reasoning barely moved while the ability to sit at a machine for forty minutes and finish a real task moved a lot.

What we think:

1. Capability beats leaderboard position. A model that scores identically on a reasoning index can be dramatically better at long-horizon tool use. We route work by what a model can actually hold onto, not by a headline number.

2. Data terms are a procurement decision. When the discount for training rights is this large, someone will take it by accident. The boundary belongs at the platform level, not with whoever configures the endpoint.

3. Autonomous exploitation cuts both ways. A model that finds zero-days unsupervised is a genuinely useful adversarial reviewer and a genuinely serious thing to hand broad credentials to. Least-privilege sandboxing stops being hygiene and becomes the control that matters.

The pattern under all three launches is the same: models are getting better at doing work, not at answering questions. That rewards teams who have already built the scaffolding around them.