Claude Went Down for Thousands of Users, and Nobody Has Said Why
Claude suffered a widespread outage affecting users across multiple regions, with Downdetector logging more than 1,600 reports at the peak. The failure showed up as elevated error rates on requests to Claude Mythos 5.1, Claude Fable 5.1 and Claude Opus 5 — which is to say it was not confined to one model tier or one surface.
The complaint mix tells you where the pain landed. Roughly 60% of reports pointed at Claude Chat, followed by the mobile app at about 24% and Claude Code at about 8%. Services have since returned to normal across Claude.ai, Claude Code, the API and Console, and status trackers show everything operational.
What has not been published is a cause. Anthropic did not disclose a confirmed technical explanation for this incident, which is its prerogative but also a gap that matters more than it used to. When Claude Code sits inside CI pipelines and Managed Agents run unattended, an outage is not a chat interruption — it is a build queue backing up. If you run anything production-critical on the API, this is your periodic reminder that failover is a design decision, not an optimization.
Accenture Is the First Evaluator Anthropic Will Let Inside the Building
Anthropic named Accenture as its first embedded evaluation partner, delivering on the commitment made in Dario Amodei’s We Must Pace the Frontier essay to put independent evaluators inside the company. The work will be led by Faculty, Accenture’s specialist AI business, and covers model evaluation and red-teaming, alignment assessments and safeguard testing. Both companies expect to invest at least $1 billion each over the next five years.
The structural novelty is the access level. Where external evaluators get a model and a deadline, embedded evaluators get access comparable to an employee’s — they can watch models take shape during training, follow the decisions that govern how models are built and deployed, and talk directly to staff. From there they can verify that safety commitments are actually being kept, name blind spots, report incidents, and give the public a more informed account than a press release provides. Anthropic is explicit that this does not reduce its own accountability; the model’s safety remains the company’s responsibility.
The unsolved part is money. There is no settled system for funding independent evaluation, and Anthropic’s stated preference — pooled or government funding, as argued in its June Advanced AI Framework — does not exist yet. So for now Anthropic is funding Accenture’s work directly, which is the obvious objection to the whole arrangement and which Anthropic names itself. Mitigations: the partnership is non-exclusive on both sides, more evaluators are promised in the coming weeks, and the company says it is in dialogue with METR and other nonprofits about piloting embedded evaluation on their own funding.
Claude Is Now Leading a Quarter of the Work That Builds Claude
Anthropic disclosed a number that is easy to read past and hard to unsee: Claude now leads 26% of the company’s model research and development. “Leads” has a specific definition here — the model completes most of a given task end-to-end from a high-level prompt, still under human supervision. Broaden the frame and the figure gets larger: about 90% of Anthropic’s R&D is done in collaboration with Claude, meaning large chunks of work under close human direction.
Pair it with the second thing Anthropic published this week: a proposed set of public metrics for measuring the pace of AI development inside frontier labs. The argument is that outsiders currently cannot see what is happening inside these companies, and that some standardized, publishable measures would fix that. The 26% figure is effectively the first one on offer.
Why that pairing is the story rather than either half: the recursive-improvement question has spent a decade as a thought experiment with no instrument attached. Putting a percentage on how much of model development the model itself leads, then proposing it as a metric the whole industry reports, converts an argument into a measurement. Whether rivals adopt it is the interesting follow-up — and the answer will say more about competitive positioning than about safety.
Your claude.ai Skills Now Follow You Into the Terminal
The change most likely to alter your daily habits shipped quietly in 2.1.275: the skills and plugins enabled on your claude.ai account now sync to terminal sessions signed in with that account. If you have been maintaining two parallel setups — one in the web app, one in ~/.claude — that duplication is over. Opt out with syncClaudeAiSkills: false or syncClaudeAiPlugins: false if you would rather keep them separate.
Also new and immediately useful: a send-now key (ctrl+enter, or ctrl+x ctrl+s) that interrupts the current turn and sends every queued message at once, with sent and queued messages shown in gray until the model receives them. Anyone who has typed three follow-ups while watching Claude go down the wrong path already knows why this matters. Rounding out the stretch: /plugin install <plugin> --marketplace <source> now offers to add the marketplace first, and gateway sign-in names the account it is about to save so you confirm before the credential lands.
The billing item from 2.1.278 is still the one to check: auto mode defaults to the server-side classifier for Claude API and Enterprise users and on Bedrock, Vertex, Foundry and gateways, and that classifier does not charge for classifier overhead. A new Auto mode server row in /status tells you which side of that line your session is on.
The Compliance API Can Now See What Claude Did in Your Browser
The Compliance API’s local session endpoints now also return transcripts of Claude in Chrome sessions, carrying the product_surface value claude_in_chrome. It is in beta for Claude Enterprise organizations and works with your existing Compliance Access Key and the read:compliance_user_data scope — no new credential to provision.
This is a small endpoint change with a large procurement implication. Browser agents are the surface most likely to make a security team say no, because the agent is acting on live authenticated sessions across arbitrary sites. Retrievable transcripts turn that from an unauditable black box into something an eDiscovery or compliance workflow can actually ingest.
Also on the platform this month, if you missed them: Managed Agents permission policies gained an auto mode where the server evaluates each agent or MCP tool call and runs, denies, or pauses it for approval, with the reasoning surfaced in an evaluation field. And ant beta:sessions connect attaches your terminal to a running Managed Agents session so you can follow it live and approve tool calls, with --web serving the Console’s session viewer locally instead.
Snowflake Says the Reasoning Is Moving to Where the Data Already Lives
Snowflake and Anthropic reported continued momentum on their partnership, with enterprises deploying Claude models through Snowflake Cortex AI against governed data rather than exporting that data somewhere else first. The framing the two companies keep repeating is worth quoting plainly: move AI reasoning to where governed data already lives, rather than moving data to where the AI operates.
The underlying deal is a multi-year, $200 million agreement that puts Claude in front of more than 12,600 Snowflake customers across all three major clouds. Claude is the reasoning engine behind Snowflake Intelligence, Cortex Code (north of 7,100 users), and Cortex Agents for production autonomous workflows.
Stack it next to Salesforce in Claude, which moved toward open beta this month with 37 prebuilt sales skills under the Claudeforce partnership, and the pattern is hard to miss. Anthropic is not trying to be the system of record. It is trying to be the reasoning layer that every system of record calls — and it keeps winning those slots on governance and permissions rather than on benchmarks.
The Enterprise Lead Is Real but Thinner Than the Headlines Suggest
The number circulating in enterprise AI coverage: Anthropic overtook OpenAI in business adoption, reaching 34.4% share against OpenAI’s 32.3% as of May 2026. Alongside it, Claude Code reached $2.5 billion in annualized revenue by February, and the company crossed $30 billion annualized in April on its way past $65 billion by the end of July.
Read the spread before you read the ranking. Two points of share is inside the noise for this kind of survey, and the measurement is four months old in a market where quarter-over-quarter movement has been enormous. “Anthropic leads enterprise” and “the two are effectively tied in enterprise” are both fair readings of the same data.
The durable signal is not the rank order, it is the composition. Claude Code alone became a multi-billion-dollar line inside twelve months, and the enterprise wins are landing through governed integrations — Snowflake, Salesforce, Microsoft 365 — rather than direct seat sales. That revenue is stickier than a chat subscription and much harder for a competitor to take back with a price cut. It is also, conveniently, the exact story a prospectus wants to tell.
Two Kinds of Trust, and Only One of Them Got an Audit This Week
Today put two very different trust problems on the same page, and it is worth noticing that Anthropic has built elaborate machinery for one of them and almost none for the other.
The Accenture partnership is a serious institutional answer to the question how do we know the lab is doing what it says about safety? Employee-level access, alignment assessments, red-teaming, incident reporting, a billion dollars a side, a stated intent to bring in METR and other nonprofits next. You can argue that Anthropic paying the evaluator undercuts the independence — Anthropic makes that argument itself and says the fix is pooled or government funding that does not exist yet — but the direction is right and nobody else has shipped anything comparable.
Now the other question: how do we know the thing will be up? More than 1,600 outage reports, three model tiers throwing errors, and no published cause. Those are not comparable failures in severity, and it would be silly to pretend a few hours of degraded service ranks with alignment risk. But they are comparable in kind, because both are cases where users have to take the company’s word for something they cannot verify. One got a billion-dollar external verification apparatus this week. The other got a status page turning green.
That gap matters more as the product changes shape. When Claude was a chat window, an outage was an inconvenience you waited out. Now it is a CI dependency, an unattended Managed Agents runtime and the reasoning layer inside Snowflake and Salesforce — the exact governed-integration position the ecosystem section describes. Enterprise buyers who accept “reasoning layer for your system of record” will eventually ask for the thing they ask every other vendor in that position for: a postmortem, an SLA and a credit. Anthropic has spent the week demonstrating it knows how to build verifiable accountability. The reliability story is the one where it still asks to be taken at its word.