Anthropic says its staff's agents worked 6x their hours by August, and the week in Claude
PLUS: sources tell CNBC Anthropic's sales hit $16.8 billion last quarter, Anthropic says Claude beat the best published quantum circuits, and six outside testers now work at Anthropic.
Happy Sunday. Here's what changed in Claude from Oct 11 to 17.
This week, Anthropic said its staff's agents now log more hours than they do. It says Claude drew better quantum circuits. And outside testers got office space at Anthropic.
Plus 10 quick hits. Every claim has a link.
Anthropic says its staff's agents worked 6x their hours
By August, agents run by Anthropic's staff worked six times as many hours as the staff.
That's per Anthropic's Oct 16 report. It counts only interactive sessions that staff direct, with subagents pooled. It assumes each person works 40 hours a week.
Agent hours count runtime. Waits over 30 minutes don't count.
Anthropic says all agent work hours passed human work hours for the first time in March. That figure is based on agent session data.
Anthropic says juniors' share of technical staff fell from 21% to 14% since early 2025. Their share of technical hires fell from 27% to 9%.
Sources tell CNBC Anthropic's sales hit $16.8 billion last quarter
Anthropic told investors its third-quarter revenue was $16.8 billion, CNBC reports. It cites people familiar with the matter.
That's up from $11.6 billion in the second quarter. The first quarter was $4.7 billion, the people said. They added that Anthropic said it was profitable "on an adjusted basis" in the third quarter.
CNBC cites the Financial Times for one more figure. Anthropic told investors it made $1.70 in revenue per dollar spent on computing power in the third quarter. CNBC says the IPO is reportedly slated for November.
Claude Code's Code Review is out of preview, and faster
Code Review is now generally available on Claude Team and Enterprise plans.
Most reviews now finish about 3x faster than in the preview, per Anthropic's Oct 15 post. That's "~20 minutes down to 6-10 minutes."
Admins pick Quick, Standard or Deep reviews for each repo. Standard is the default. Once it's on, reviews run on new pull requests with no setup for developers.
Eligible Enterprise orgs can claim at least $10,000 in Code Review credits by Jan 15, 2027. The credits expire 30 days after you claim them. Terms apply.
Anthropic says copycats now get 10 to 100x less from Claude
Anthropic says some companies copy Claude by training their models on its answers. It says they now get far less of its reasoning.
That's per Anthropic's Oct 12 post. Anthropic estimates how much reasoning these "distillers" pull on its platform. It says the amount is now "consistently 10 to 100 times lower" than before.
Anthropic says early evidence suggests the data they harvest fell by "around 90 to 99%." It has banned tens of thousands of linked accounts.
Anthropic lists trade-offs. Some developers had to update their integrations.
Its filters wrongly flag or block harmless requests in fewer than 0.1% of cases. It now charges for those.
Anthropic says it doesn't yet know how long the gains will last. It hasn't yet seen a highly sophisticated and persistent attempt to get around them.
Anthropic says Claude beat the best published quantum circuits
Anthropic says Claude beat the best published quantum circuits on eight tasks.
Per Anthropic's Oct 16 post, Claude Mythos 5 worked in a loop. A program checked and scored each design. Claude's designs used fewer qubits and fewer gates than the best published ones.
In some cases, the gain was "an order of magnitude or more."
On one task, Claude worked for 53 hours with minimal human input. That input was the first brief and occasional prompts to think more broadly. Anthropic says practical quantum computers could arrive "sooner than previously anticipated."
Anthropic notes that circuit design is only one part of building a useful quantum computer. An Anthropic STEM Fellow led the work, with Harvard researchers. The full paper is on arXiv.
Six outside testers now work at Anthropic
Six staff from METR, a nonprofit that tests AI, now have "employee-like access" at Anthropic.
They got company hardware, office space and internal tools, per Anthropic's Oct 15 post. That access is "roughly on par" with what Anthropic's own risk teams have.
Anthropic staff are encouraged to share info with them as with a coworker. A few set categories are excluded, like customer data.
METR will publish its findings, and Anthropic can't veto them. Anthropic can require it to black out set kinds of information.
Anthropic decides what falls into those kinds. METR can still make certain statements about the access and information it got.
The six work under enforceable agreements, walled off from the rest of METR.
Claude will show official US voting info and vote counts
Claude.ai will soon answer US voting questions with official info for your state.
Anthropic said on Oct 14 that this would start later this month. The data comes from Democracy Works. On election night and the days after, Claude will also show the latest official vote counts.
Anthropic looked at about 250,000 US Claude.ai chats from Sept 21 to 27. About 0.01% asked Claude whom or what to vote for.
Testers from two outside groups found Claude reliably refused a single harmful request. But it "could miss a malicious campaign" spread across many harmless-looking requests. Anthropic says it has added new classifiers since.
Quick hits
- Fast mode for Opus 5.5 no longer needs an access request on the Start, Build and Scale usage tiers (Oct 15). It's out of research preview. Fast mode for Opus 5 and Opus 4.8 stays in research preview.
- Fast mode gives up to 2.5x higher output speed at premium pricing, per Anthropic's docs. Opus 5.5 fast mode also came to Microsoft Foundry for Azure-hosted deployments (Oct 15).
- You can now trace the agents you build on the Claude Platform (public beta, Oct 14). OpenTelemetry traces cover Managed Agents, the Messages API and Anthropic's SDKs. They work with OTel tools that accept a standard OTLP endpoint.
- Claude Code 2.1.299 (Oct 14) compacts idle long chats in the terminal. A chat over 200K tokens left idle for about 54 minutes gets compacted before its one-hour prompt cache expires. A setting turns it off.
- Claude Code MCP allow rules now reach only the server they name (2.1.299, Oct 14). Before, a rule for a server named "a" also allowed ones named "a__b" or "a_".
- Claude Code 2.1.301 (Oct 15) handles model switches better. Say you switch to a model with a smaller context window. The old model now sums up a chat too long for the new one.
- Managed Agents scheduled deployments came to Claude Platform on AWS (Oct 14).
- The Compliance API now returns each member's saved instructions for Claude (beta, Enterprise, Oct 15). That covers claude.ai and Cowork, per the release notes.
- Anthropic opened a Singapore office (Oct 14), its fifth in Asia-Pacific. It also signed an MOU with Singapore's government. It covers AI safety, security and model evaluation, among other areas.
- Replit's CEO said onstage that Replit is no longer working with Anthropic, TechCrunch reports (Oct 15). He later said customers can still use Anthropic models, but OpenAI's ChatGPT is now the default. A person familiar with the partnership said it was still in place.
Worth your time
- The AI expectations gap (Oct 16): Anthropic surveyed its staff and US adults. By 2030, 91% of staff and 39% of adults expect AI to be able to make a Nobel-level discovery.
- How Anthropic rebuilt source control (Oct 13): Anthropic merged about 2,500 code changes a day in May. By September, it merged over 30,000. It built its own version control system and moved off git in five months.
- How the Claude Code team does code review (Oct 15): Claude Code's merged pull requests more than doubled from June to September. The share reverted fell by more than 40%.
That's the week. The next issue lands Sunday, Oct 25.
Reply and tell me what you're building. I read every reply.
— Sam