Claude Now Has a Browser of Its Own
Claude’s built-in browser has finished its gradual rollout to Cowork in the Claude desktop app on macOS, Windows and Linux (beta). When a task needs a website, a browser opens in the Cowork side panel and Claude navigates, reads pages, clicks and fills forms. There is nothing to install and no extension to approve.
The isolation is the part worth reading twice. The built-in browser is separate from your own browser and carries its own profile, so it does not see your logins unless you explicitly import them. Availability: Pro, Max and Team plans, plus Enterprise where an owner has enabled it. Team is on by default. Enterprise launched off by default and flips on September 10, 2026 unless an owner turns it off first — put that date in the admin calendar now rather than discovering it.
If you already use Claude in Chrome, it keeps working and stays your default; otherwise Claude reaches for the built-in browser. The toggle lives under Settings → Cowork → Preferred browser. The practical difference is who owns the session: the Chrome extension acts inside your real browsing life, and the built-in browser acts in a sandbox that starts empty.
The August Risk Report Raises Misalignment From “Very Low” to “Low”
Anthropic published its August 2026 Risk Report, the recurring disclosure required by Responsible Scaling Policy v3.0. It covers the window since the February report through a July 15 cutoff, and it is unusually plain about where the company thinks it stands.
The headline change: misalignment risk in high-stakes settings moves from “very low” to “low.” Anthropic frames this as reflecting uncertainty rather than a new finding — the evidence did not get worse, the confidence got weaker. The report also discloses that human feedback vendor traffic ran without blocking biological classifiers for nearly a year. That gap has been closed and Anthropic says it found no misuse.
Two Frontier Safety Roadmap goals were marked complete: the planned moonshot R&D projects launched, and an internal review of how Safeguards could be improved through data retention policy changes finished. Reporting your own control failures in a document you are not legally obligated to publish is the useful behavior here. The disclosure is the product; the rating change is just the number attached to it.
Opus 4.1 Is Gone. Sonnet 4.5 Has Thirty Days.
Claude Opus 4.1 is retired. Requests to claude-opus-4-1-20250805 now return an error — not a fallback, not a warning. The recommended target is Claude Opus 5, which is both cheaper and stronger on most agentic benchmarks, so the migration is unusually painless as these go.
Next on the calendar: Claude Sonnet 4.5 retires September 29, 2026. That is thirty days. Same hard cutoff, same failure mode. Opus 4.5 is unaffected.
If you are still below anthropic-sdk-python v1.0, batch the two migrations. Claude Code now ships a /claude-api upgrade command that walks a Python project from anthropic 0.x to 1.x, and the 1.x reference has one gotcha worth flagging: timeouts use anthropic.Timeout, not httpx.Timeout. Sweep config stores, notebooks, CI secrets and vendor integrations, not only the services you believe are live.
Mid-Conversation System Messages Land on Fable 5, Mythos 5 and Opus 4.8
You can now inject a system message partway through a conversation on Claude Fable 5, Claude Mythos 5 and Claude Opus 4.8, available on the Claude API, Amazon Bedrock and Google Cloud.
This is a small primitive with outsized consequences for long-running agents. Until now, changing an agent’s operating instructions mid-task meant either restarting the conversation or smuggling the change into a user turn, where the model treats it with user-level authority. A real mid-conversation system message lets you tighten a policy, narrow a tool budget or hand off to a different phase of work without losing accumulated context.
The obvious use is staged workflows — research phase, then drafting phase, then review phase, each with different rules and the same history. The less obvious use is safety: revoking a capability mid-run at the right privilege level, rather than hoping a user-turn instruction outranks the original prompt.
Claude Code: Plugin Sync in Cloud Sessions, Alpine Gets Its Add-Ons
A maintenance-flavored release with a few things worth knowing. Cloud sessions now show plugins synced from claude.ai as name@synced. They respond to claude plugin enable and disable, and critically they never override a same-named plugin you installed yourself — local wins, which is the right default.
Alpine and musl builds now load the native add-ons for image paste, clipboard and audio capture. If you run Claude Code inside a slim container, the features you assumed were missing may simply have been unbuilt.
Two proxy fixes for Bedrock users: streaming behind proxies that strip the response Content-Type header now works, and Claude Code no longer hangs at startup behind an HTTPS proxy when using Bedrock with an SSO profile. If you shelved a corporate-network Bedrock setup earlier this year, it is worth another attempt.
The Compliance API Can Now Read Cowork Transcripts
In beta for Claude Enterprise organizations: the Compliance API now returns transcripts of Cowork sessions started on claude.ai web or mobile.
Agentic sessions have been the blind spot in every enterprise audit story so far. Chat transcripts were retrievable; the sessions where an agent actually touched files, browsed and executed work were not. Closing that gap is what makes Cowork deployable in regulated environments — and it is worth telling your users plainly that these sessions are retrievable, because the assumption cuts both ways.
Pair this with the skill and plugin security scanning beta and the Managed Agents domain allowlists that landed this week, and the enterprise agent story is finally three things instead of one: a spend cap, a network boundary and an audit trail.
Opus 5 Overtook the Flagship, and Fable 5 Stalled Near 11%
Payments company Ramp, analyzing spend across 70,000 businesses and reported by the Financial Times, found that Claude Opus 5 has already passed Claude Fable 5 in enterprise spending. Opus 5 shipped in late July. Fable 5 has been generally available since spring.
Two months after launch, Fable 5 accounts for roughly 11% of what customers spend on Anthropic tools, and that share has plateaued. The stated reasons are unglamorous and specific: Fable runs around $10 per million tokens, roughly double GPT-5.6 Sol, and carries a 30-day data-retention requirement that some buyers cannot accept. Opus 5 sits at $5 in / $25 out and out-benches Fable on agentic terminal coding, agentic search and computer use.
The uncomfortable framing in the coverage is that Anthropic’s best model cannot hold users. The more accurate framing is that Anthropic shipped a cheaper model that beats its flagship on the work enterprises actually buy, and customers noticed. That is a product decision working as designed — it just happens to be inconvenient in the month before an IPO.
A2A Joins MCP Under One Foundation
On August 20, Google’s A2A protocol formally joined the Linux Foundation’s Agentic AI Foundation, placing it under the same neutral governance as Anthropic’s Model Context Protocol. The AAIF now counts more than 250 members, including AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI.
The two protocols solve adjacent problems — MCP connects a model to tools and data, A2A connects agents to each other — and having both under one roof makes the boundary between them a governance question rather than a competitive one. Vendor-donated standards usually calcify or fork; this is the arrangement that avoids both.
Separately, the MCP maintainers published an updated roadmap on August 22 covering server-initiated events (webhooks and channels), composition review across the Agents, Transports and Triggers working groups, and maturing the Tasks extension toward specification status. The 2026-07-28 spec already did the heavy lifting: a remote MCP server is now no different from any other HTTP workload, hostable on infrastructure you already run.
The Fable 5 Plateau Is the Best News in Anthropic’s Worst Headline
“Anthropic’s best model can’t hold users” is a great headline and a bad reading of the data.
Here is what actually happened. Anthropic released a frontier model at a frontier price. Three months later it released a cheaper model that beats the frontier model on agentic coding, agentic search and computer use — the three categories that account for most enterprise token spend. Buyers moved. Ramp measured the move. The measurement got written up as a failure of the flagship rather than a success of the follow-up.
The revenue question is whether the mix shift is cannibalization or expansion, and 11% of Anthropic spend is a share number, not a volume number. Anthropic’s annualized run rate went from $47 billion in May to more than $65 billion in July across the same window in which Fable’s share stalled. A shrinking slice of a fast-growing pie is not the same shape of problem as a shrinking pie, and nobody reporting the 11% figure has published the volume series that would settle it.
What is genuinely awkward is the 30-day data-retention requirement. Price objections get negotiated; retention objections get vetoed by someone in legal who never joins the call. If Fable’s ceiling is partly a compliance ceiling, that is a fixable problem Anthropic has chosen not to fix yet, and it is a strange thing to carry into an S-1 when the buyers you are courting are exactly the ones with a retention policy.
The real signal sits underneath both stories: capability at the top is getting commoditized faster than anyone modeled. Opus 5 delivers near-flagship performance at half the price within a single quarter. If that cadence holds, the durable business is not the frontier model — it is the harness. The browser, the Skills API, the Compliance API, the Admin API, MCP under neutral governance. Every one of this week’s releases was plumbing, not intelligence. That is not Anthropic hedging. That is Anthropic reading its own spend data.