The two biggest labs spent the day pointing their frontier models at specific professions rather than making them cleverer in general. OpenAI shipped Astra for Law, GPT-6 Astra wrapped in a purpose-built legal research index, and made ChatGPT for Word generally available alongside dozens of partner plugins. Anthropic, on the same day, released public code that used Claude to speed up more than thirty biomolecular models and opened a Life Sciences Verification Program that swaps blanket refusals for verified access and after-the-fact monitoring. Around those, agents kept creeping onto the desktop, with Meta's Muse arriving on Mac nine days after launch and Perplexity replacing its model dropdown with a simple effort slider. The through-line is that the competition has moved from who has the best model to who has the best domain index, the best governance, and the surface a professional already works in.
Astra for Law puts GPT-6 Astra behind a purpose-built legal research stack
This is the first time OpenAI has shipped a frontier model configured for a single profession, and the shape of it matters more than the legal angle. Astra for Law is GPT-6 Astra plus a dedicated legal search index covering over 230 million URLs of US case law, statutes, regulations and court rules, with custom instructions for legal analysis layered on top. OpenAI reports it passed the overall correctness check on 54.0 per cent of questions from Vals AI's Legal Research Bench validation set against 38.7 per cent for plain Astra with web search, a useful reminder that retrieval quality now moves the needle more than raw model size.
It arrives first through a Trusted Access programme for selected firms in ChatGPT and Codex, appearing in the picker as GPT-6 Astra Law, with API access coming soon. Following Anthropic's Claude for Financial Advisors this week, if you are building vertical AI products in any regulated field, treat this as the template OpenAI will reuse: frontier model, domain index, domain instructions, governance controls.
ChatGPT for Word goes generally available, alongside 26 partner plugins
Buried inside the same announcement is the change most people will actually touch. ChatGPT for Word is now generally available, so proofreading, suggested edits and formatting checks happen inside the document rather than in a separate chat window. OpenAI also launched 26 partner-built plugins and nine community plugins carrying 47 custom skills, connecting ChatGPT to tools like iManage, Intapp, Relativity and Clio.
The strategic read is that OpenAI is positioning ChatGPT as the place other vendors plug into rather than the thing that replaces them. For anyone weighing whether to build a thin wrapper product, the plugin surface is now the more sensible place to compete.
Claude optimised 30+ open-source biomolecular models and the code is now public
The headline number is a roughly 4x average speed-up across more than 30 open-source models, achieved in just under four weeks, with all the optimised code released publicly. What makes this worth your attention even if you have never touched a protein structure is the supervision model: Claude was overseen by two staff with biomolecular modelling experience and no background in inference optimisation or kernel engineering. That is a concrete data point on AI closing a specialist skills gap rather than just producing more text faster.
Anthropic also added a low-memory mode that lets systems larger than 10,000 tokens be predicted accurately on a single NVIDIA GPU node, which quietly lowers the hardware bar for a whole research field. Paired with it is a protein design competition co-sponsored with Adaptyv Bio, backed by up to 1 million dollars in Claude credits and wet lab validation for over 5,000 designs.
Life Sciences Verification Program moves safeguards from blocking to monitoring
This is an access change dressed as a safety announcement, and it is the more interesting of the two. Verified life science teams can now apply for grants that give them Mythos, Opus and Sonnet models with classifiers tuned to be more permissive for biology work that the general Fable models refuse. The mechanism is the part worth noting: Anthropic is shifting from real-time request blocking to offline monitoring across patterns of behaviour, with 30-day data retention on flagged activity, compartmentalised from training.
Standard Use grants cover most research and development workflows and run for a year, while High-risk Use grants strip all life science safeguards for a single named project and renew every six months. If you have ever hit a refusal on legitimate technical work, this is the first clear signal that frontier labs will solve over-refusal with verified identity and after-the-fact auditing rather than looser filters for everyone.
Muse arrives on Mac, nine days after launch
Meta's personal AI agent landed on 8 September and already has a desktop app, which tells you how hard Meta is pushing to make Muse a habit before rivals settle into the same slot. Muse for Mac works across apps, files, calendar, notes and messages on the machine itself, with explicit permission, rather than being confined to a browser tab. That moves it into the same contested territory as ChatGPT's desktop presence and Claude's computer use, with the difference that Meta is aiming squarely at everyday personal admin rather than developer or knowledge work.
Availability remains US-only across iOS, Android, web and WhatsApp. If you are assessing agent products, the Mac release is the point where Muse stops being a phone assistant and starts being something that can touch your actual working files.
Computer gets a four-step effort slider that bundles model choice with reasoning depth
Perplexity has replaced the usual model dropdown with a single slider running Light, Standard, High and Ultra, where each position maps to both an orchestrator model and a reasoning level. This is a quietly significant interface decision. Most agent products still ask users to pick a model name, which is a question almost nobody outside the industry can answer sensibly, and instead hides the thing they actually care about, which is how much time and credit a task deserves.
Perplexity says model selection behind each preset is informed by billions of queries, and the manual controls remain for people who want to name a specific model themselves. It is live on web now, with mobile and desktop to follow. Expect this pattern to spread, because effort is a unit users understand and model names are not.
Industry themes
The vertical turn is now explicit. OpenAI shipped a profession-specific configuration of its frontier model rather than a better general model, and Anthropic opened a verified access programme for a single scientific field on the same day. The competitive frontier is moving from who has the best model to who has the best domain index, domain instructions and domain governance wrapped around it.
The second pattern is that safeguards are being rebuilt around identity rather than content. Anthropic swapped real-time refusals for verified access plus offline pattern monitoring, which is the most practical answer yet to over-refusal on legitimate technical work. Expect verified identity and after-the-fact auditing to replace blunt content filters wherever the work is legitimate but sensitive.
Third, agents are quietly colonising the desktop, with Muse reaching Mac nine days after launch and Perplexity reframing model choice as an effort dial, both of which suggest the competition is now about habit and interface rather than benchmark scores. Every product item in this digest appeared first on the company's own X account, in several cases hours before any press coverage, and Meta's Muse for Mac release was announced only through its own accounts rather than a company blog.