Anthropic Previews the Model Hardware Standard — Agents That Drive Real Machines
Anthropic has opened the first phase of a research preview for the Model Hardware Standard (MHS), a shared specification for AI agents to safely operate physical equipment in scientific research and advanced manufacturing. It goes to a first group of research labs and manufacturers. This is the company’s first serious move out of the text box and into machinery.
What it actually does: an agent can operate multiple instruments in parallel — microscopes, liquid handlers, robotic arms — and run tasks that range from routine drug-discovery experiments to laser calibration on a quantum computer. MHS works with any device exposing a programmable interface, and it is model-agnostic: any agent harness can reach it over standard protocols, MCP included.
The problem being solved is unglamorous and enormous. Labs currently spend weeks or months wiring instruments together, because most devices do not speak to each other and every pairing needs a specialist to build a bespoke integration. Anthropic claims MHS reduces that to hours or minutes. The work began as a collaboration with HHMI Janelia Research Campus, and Anthropic says it plans to open source the standard.
10,000 Free and Discounted Claude Seats Open for Scientists
Anthropic is opening 10,000 Claude Team subscription seats to researchers. Standard seats are free. Premium seats, carrying five times the usage limits, run $15 per month with pricing fixed for one year. Eligibility runs through a verification form and requires principal-investigator status or equivalent at an academic or nonprofit research institution; once verified, a PI can add their whole lab.
Stacked on top is the AI for Science credit program, which grants up to $50,000 per project for high-impact research. And underneath both sits Claude Science, the research workbench: a generalist coordinating agent fronting more than 60 curated skills and connectors pre-wired for genomics, single-cell analysis, proteomics, structural biology and cheminformatics.
Anthropic says it intends to push well past the initial 10,000 seats in the coming months. Read this next to the MHS preview and the shape becomes obvious — software seats, credits, a domain workbench and now a hardware protocol, all pointed at the same lab bench.
Cowork Gets Its Own Browser — No Extension, No Setup
Claude Cowork now ships a browser built into the desktop app. When a task needs a website, a browser opens in the side panel next to your work and Claude navigates, reads pages, clicks, types and fills forms while you watch. No window switching, no extension install.
It is rolling out gradually this week on macOS, Windows and Linux (beta) for Pro, Max and Team plans, and on Enterprise where an owner has enabled it. Capabilities match Claude in Chrome, but it does not touch your own browser and works regardless of which browser you normally use. Nothing is shared from your personal profile unless you explicitly import logins.
The precedence rule is worth knowing before you file a bug: if you already run Claude in Chrome, that stays the default for Cowork web tasks. The built-in browser only takes over if you do not have the extension, or you are new to browser use in Cowork.
Managed Agents Ship Four Updates — Including a Hard Spend Cap
Four changes landed on Claude Managed Agents in one week, and the first one closes the single loudest complaint about running autonomous agents in production.
Session budgets. You can now set a hard cap on a session’s spend, priced at public list rates. A session that hits its budget stops issuing new model requests and pauses with a budget_reached stop reason. Raise or remove the budget and it resumes. That is the difference between an agent you can leave running and one you have to babysit.
Advisor models. A session can be given an advisor — a model at least as capable as the agent’s own — that the primary thread consults mid-turn for strategic guidance. Configure it as a {"type": "advisor"} entry in the multiagent roster. Inference geo pinning. Set inference_geo to "us" or "global" on the agent’s model object, or override per session — the data-residency answer procurement keeps asking for. GitHub-hosted skills. Mount a repository and any skills in its root .claude/skills directory are discovered automatically at session start.
Claude Is First to Ship MCP 2026-07-28 — the Stateless Rewrite
The MCP 2026-07-28 specification is final, and it is the largest change to the protocol since it launched in November 2024. Claude is the first major client to support it, rolling out across Claude products including Cowork’s connector directory of 950+ MCP servers.
Three things changed. The core moved from a bidirectional stateful protocol to request/response, which means MCP servers can finally deploy on serverless and edge infrastructure instead of holding open connections — a straight cost and reliability win for anyone running servers at scale. Apps and Tasks now ship under a versioned extensions framework, so interactive UIs and long-running work get a formal path that does not require touching the core spec. And authorization aligns with production OAuth 2.0 and OIDC, so servers plug into Entra or Okta without workarounds.
Scale marker for context: MCP has passed 400 million monthly SDK downloads, roughly 4x growth this year.
Claude Code Adds a Default Model Setting and Tightens the macOS Sandbox
The current release line sits at 2.1.247 (published August 26). Recent additions: a new default model setting, cross-session idle notifications so a long job can tell you it is waiting, and stronger macOS sandbox protections. Auto mode, Remote Control, startup speed and VS Code screen reader support all improved alongside.
Also worth noting from the same window: restricted mode, cross-session messaging, usage credits for Enterprise, and better diagnostics for server-managed settings. Self-hosted environments arrived back at 2.1.224, letting Team and Enterprise plans run sessions on their own machines rather than Anthropic’s hosted infrastructure — the option to reach for when your security review will not clear hosted execution.
Thirty-Two Days Until Sonnet 4.5 Retires
Claude Sonnet 4.5 retires September 29, 2026. After that date, calls against claude-sonnet-4-5 fail outright — the same hard stop that took Sonnet 4 and Opus 4 on June 15 and Opus 4.1 on August 5. Opus 4.5 is unaffected.
If you are still on anthropic-sdk-python below v1.0, bundle the two migrations. The pinned model identifier and the removed temperature, top_p and top_k parameters live in the same call sites, so one pass covers both. Search repos, config stores, notebooks, CI secrets and vendor integrations — not just the services you believe are live.
The Public Filing Window Is Now Measured in Days
Reporting this week says Anthropic could file publicly as soon as the end of August — which is to say, now. The company filed a confidential S-1 on June 1; the public S-1 was always expected in late summer or early fall, followed by an institutional roadshow and pricing.
The numbers being floated: an October 2026 Nasdaq listing, Goldman Sachs, JPMorgan and Morgan Stanley leading, more than $60 billion raised, and investors targeting a $2 trillion valuation. That would top SpaceX’s $1.77 trillion June offering as the largest IPO ever. The private mark stands at $965 billion from May’s Series H.
Standard caveat, and it matters more as the date closes: none of this is confirmed by Anthropic. Roughly doubling a private valuation at listing is an ambition, not a price. The one thing that will settle it is a public filing, and that document is reportedly imminent.
Zoom’s Anthropic Stake Is Now Worth $3 Billion
A small story that illustrates a large one. Zoom’s holding in Anthropic has grown to roughly $3 billion in value on paper. Zoom is not an AI lab and did not set out to be an AI investor of consequence; it made a strategic investment and the mark ran away from it.
The reason to care is not Zoom. It is that a growing set of public companies now carry Anthropic exposure large enough to move their own numbers, which is exactly the dynamic that turns a private valuation into a market-wide one. When the S-1 lands, those marks stop being estimates.
Anthropic Hires a Google Veteran for the Hardware Push
Reported last week and worth revisiting in light of today’s MHS preview: Anthropic has brought on a Google veteran as part of a hardware effort. At the time it read as a compute-and-silicon story, filed next to the Nscale deal.
With the Model Hardware Standard now public, a second reading is available. “Hardware” at Anthropic may mean more than accelerators — it may mean the physical layer agents act through. A specification for driving lab instruments is not a side project you staff casually. Treat this as a thread to keep pulling rather than a conclusion.
Anthropic Is Running the MCP Playbook Again, and This Time the Stakes Are Physical
Look at what Anthropic did with MCP and then look at MHS. The pattern is identical, deliberate, and by now well tested.
Step one: find an integration problem that is genuinely painful and that nobody owns. With MCP it was connecting models to tools — every vendor building bespoke adapters, none of them portable. With MHS it is connecting agents to instruments, where a lab burns weeks or months wiring a microscope to a liquid handler because the two were never designed to speak.
Step two: publish a standard that is deliberately model-agnostic, so competitors can adopt it. This looks like a concession and it is the whole trick. A standard only becomes a standard if rivals can use it; a proprietary interface stays a product feature. Anthropic gave MCP away and MCP won — 400 million monthly SDK downloads, native clients in Cursor, VS Code, Zed and JetBrains. MHS is being previewed with an explicit open-source commitment for the same reason.
Step three: be first and best at implementing your own standard. Claude was the reference MCP client and is the first to ship 2026-07-28. If MHS takes, Claude will be the agent every instrument vendor tests against.
What is different this time is the failure mode. A misbehaving MCP server returns bad JSON. A misbehaving MHS agent moves a robotic arm. Anthropic’s framing leans hard on the word safely, and the research preview is being handed to a small vetted group rather than published open — a sequencing choice they did not make with MCP. That is the right call, and it is also the hard part. Software standards win on convenience. Physical standards have to win on liability, and the questions that follow — who is accountable when an agent-driven instrument damages a sample, injures a technician, or quietly invalidates six months of results — are not answered by a spec document.
The practical read: if you run a lab or a production line, the integration savings are real and worth evaluating in preview. But evaluate the governance alongside the protocol. Ask what the agent is permitted to actuate without a human, what the audit trail looks like, and who signs off. MCP taught the industry that Anthropic writes good specs. Nobody has yet been taught what happens when one of those specs is wired to something that moves.