The Sequence Radar - Issue 927: Last Week in AI: Model Madness: The Frontier Has a Refresh Button
TheSequence Jesus Rodriguez
OpenAI, Anthropic, Meta, and Google all shipped new AI models this week. The odd part: the real race is now speed, cost, and the software around the model.
Based on reporting by TheSequence, Jesus Rodriguez — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
This was a packed week even by AI’s usual standards. OpenAI, Anthropic, Meta, and Google all pushed out new models or updates, and the headline wasn’t just raw capability. It was how quickly the frontier is now being refreshed, repackaged, and tied to surrounding tools that make the models actually usable.
OpenAI’s GPT-6 Astra comes with a broad pitch: computer use, software engineering, and scientific work. The company says it hit 98% on FrontierMath Tier 4, while also improving computer interaction. That combination matters. A model that can reason is useful; a model that can move through software and complete complicated workflows with less supervision is something builders can plug into real work. OpenAI is also updating Codex harnesses to speed up computer use, which is a reminder that the model itself is only half the product now.
Anthropic took a slightly different path with Claude Fable 5.1 and Mythos 5.1. Both are built on the same underlying model, but they ship with different safeguards and access rules. Fable is generally available. Mythos stays in trusted access programs. Anthropic also says cheaper cache reads should cut costs for typical workloads by around 25%, with bigger savings possible for more agent-heavy work. So capability and economics arrived together, as they increasingly do.
Meta’s Muse Spark 1.3 is aimed at the dull but essential parts of agent work: keeping requirements straight over long tasks, dealing with conflicting information, revising plans, and asking for help when needed. Meta says it trained the model across multiple agent harnesses to improve generalization between environments. Google, meanwhile, showed just how fast this cycle has become with Gemini 3.8 Flash, its third Flash release in six weeks. It keeps 3.7 Flash’s speed and introductory pricing, while claiming better coding and reasoning. There’s also a restricted Flash Cyber variant for vulnerability discovery and patching.
Put together, the week points to a market that’s shifting from model bragging rights to execution quality. The question is no longer just who has the smartest demo. It’s who can ship a model, wrap it in the right harness, keep costs sane, and survive the next refresh without forcing everyone to rebuild their evals from scratch.
My take — AI-written commentary, not fact-checked reporting
The industry has finally admitted that a model without good harnesses is just an expensive parlor trick. The smart money is on whoever makes upgrades boring: safer, cheaper, reversible, and fast enough that teams stop treating every release like a fire drill. That won’t fit on a keynote slide, which is probably why it’s the real competition.
Read more about this at: TheSequence