Anthropic says Sonnet 5.5 costs up to 30% less for most work, and the week in Claude
PLUS: 6 tech bosses sign a White House safety pledge, Claude Code gets mods, and an hour of Claude errors
Hi, it's Sam. Here's your week in Claude, from Sep 27 to Oct 3.
This week, Anthropic launched Sonnet 5.5. Six tech bosses signed a safety pledge at the White House. And Claude Code got mods.
Plus 13 quick hits. Every claim has a link.
Anthropic launches Sonnet 5.5 and says it runs 30%+ faster
Anthropic's new Sonnet costs the same per token as the last one. Anthropic says it typically needs fewer tokens for the same work.
Sonnet 5.5 launched on Sep 28. Anthropic calls it "a clear upgrade over Claude Sonnet 5." It says the model runs "30%+ faster" and "costs up to 30% less for most work."
The price stays at $2 per million input tokens and $10 per million output. Anthropic says it "typically needs far fewer tokens to do the same work."
Per Anthropic, it scores 70.6% on Terminal-Bench 4.0, a coding test. Sonnet 5 scored 10.3%. Anthropic says Opus 5.5 "remains clearly stronger at complex, open-ended work requiring sustained judgment."
It's the first Sonnet to launch with cyber safeguards like those on Anthropic's top models. Per Anthropic, higher-risk cyber tasks "will visibly fall back to Sonnet 5." It says routine software development is unaffected.
Six tech bosses sign a White House safety pledge
Anthropic's Dario Amodei and five other tech bosses signed a voluntary AI safety pledge. President Trump announced it on Sep 29, after a White House lunch.
Per Al Jazeera, the other signers are from OpenAI, Google, Meta, xAI and Nvidia. They are Greg Brockman, Sundar Pichai, Mark Zuckerberg, Elon Musk and Jensen Huang. The pledge is called the Joint Commitment on Frontier Responsibilities.
The companies pledged controls to monitor AI models, and internal teams to oversee them. They also pledged to work with outside auditors. And they'll meet regularly to set safety standards.
Al Jazeera says the accord doesn't name who would do the outside checks. The accord says it "may make sense" to turn the steps into law later. Per Al Jazeera, critics pointed to "its nonbinding nature and limited details."
Per TechCrunch, the signed page spells "United States" as "Unites States." The typo sits right under the president's signature.
Claude Code now has mods that change how it works
You can now change how Claude Code behaves and looks with a few lines of code.
On Oct 1, Anthropic added mods to Claude Code. Mods are small TypeScript functions. They can rewrite a prompt, add new UI or replace a built-in feature.
A mod can block, rewrite or retry a tool call. It can approve or deny a permission request. It can also hide secrets in tool output before Claude reads it.
You can write a mod yourself or ask Claude Code to write one. Mods ship inside plugins. They work in the Claude Code CLI and desktop app.
Anthropic says mods "aren't sandboxed." They run with the same access to your machine as Claude Code. Anthropic says to install them only from sources you trust.
On Team and Enterprise plans, and on machines with managed settings, a built-in mod called sec-default loads first. It stops mods that users install from risky moves, like overriding your permission deny rules.
Admins can load their own mods first instead. Anthropic says to then add sec-default to the list to keep its limits.
Anthropic: robots can do most physical tasks but cost more than people
Anthropic's economists say robots can already do most physical job tasks. But they say robots are cost-competitive for only 0.3% of job tasks.
The Sep 30 study scores how well today's robots can do US job tasks. It says robots can do three-quarters of physical tasks, "but mostly in limited settings." Those tasks make up 34% of working hours.
Per the study, driving and warehouse jobs are highly exposed. Nursing and general repair jobs are not. It says that if robot prices keep falling as they have, the cost-competitive share will take 40 years to reach 10%.
Claude had an hour of errors on Sep 29
Claude.ai, Claude Code, Claude Cowork and the Claude API had elevated errors on Sep 29.
Per Anthropic's status page, impact ran from 9:00 to 9:59 a.m. Central (14:00 to 14:59 UTC). The first errors were mitigated at 14:36 UTC.
A second issue kept many users from signing in, starting new chats, or starting Claude Code or Cowork sessions. Anthropic said some messages sent between 14:00 and 14:59 UTC "may not have been saved."
Anthropic puts $100 million into training 10,000 AI engineers
Anthropic aims to train 10,000 "Frontier Deployed Engineers" by the end of 2027.
On Oct 2, it launched Claude Frontier Academy, backed by $100 million. The first cohorts include engineers from Accenture, Bain, Capgemini, Deloitte, McKinsey and Morgan Stanley. Commonwealth Bank of Australia and Novo Nordisk are in too.
Joining is by nomination. Engineers start with a multi-day in-person program and a graded practical. Those who pass lead a real Claude project at their own company in a 12-week residency.
Those who pass a final check earn a badge. Anthropic expects the first in early 2027.
Quick hits
- Sonnet 4.5 retires from the Claude API on Nov 30. Anthropic announced it on Sep 30. It recommends moving to Sonnet 5.5.
- Code written for Sonnet 5 "can break on Claude Sonnet 5.5 in five ways," per the release notes (Sep 28). One example: forced tool use now returns a 400 error.
- Sonnet 5.5 is now Claude Code's default Sonnet on the Anthropic API, as of 2.1.284 (Sep 28).
- Sonnet 5.5 is in GitHub Copilot (Sep 28) for Pro, Pro+, Max, Business and Enterprise users. GitHub says the rollout "will be gradual."
- The FTC opened a broad probe into AI safety, a senior FTC official told ABC News (Sep 30). It includes Anthropic and OpenAI. It looks into claims of "unfair or deceptive acts" and potential harms to consumers.
- Claude for Government is now generally available (Sep 30) to federal and state agencies. It runs in a FedRAMP High authorized environment. The Claude Code CLI and Claude for Microsoft 365 are in early access there.
- Barclays is extending Claude across the bank (Oct 1). It expects Claude Code use to reach 50% of its developers by the end of 2026. More than 16,000 staff use its Claude-powered assistant.
- NVIDIA's open-source OpenShell can serve as the runtime for Claude Managed Agents' self-hosted sandboxes, per Anthropic (Sep 28). It blocks everything unless a rule allows it. Managed Agents is in public beta.
- Admin API methods are out of beta in new Claude SDKs, like Python SDK 1.10.0 (Sep 30). They cover members, workspaces, API keys and more, per the release notes.
- The Models API now shows each model's "line," such as opus (Oct 1), per the release notes.
- Claude Code background Bash and PowerShell commands now stop after a time limit (2.1.285, Sep 29). It's 30 minutes by default, 2 hours max. Since 2.1.288 (Oct 2), it only applies in unattended sessions: -p, the Agent SDK, CI and cloud.
- Claude Code 2.1.289 (Oct 3) fixes a gap in Bash deny and ask rules. They could miss a command behind an environment variable prefix with an expanded value. This happened when the sandbox auto-allows commands.
- Claude Code 2.1.289 also fixes a plugin bug (Oct 3). A user-installed plugin could rewrite the sign-in tool descriptions of an organization-managed MCP server.
Worth your time
- How Anthropic's sales team rebuilt inbound (Sep 30): a buying agent built on Claude Managed Agents (beta) now holds thousands of chats a day. Anthropic says the leads it passes on turn into opportunities more than twice as often as old form leads.
- Claude-shaped science (Oct 1): in a guest post, Prof. Matthew Schwartz describes building BootLoops with Claude. It's an open-source toolkit for exact calculations in quantitative science.
- Anthropic's AI interview study (Sep 29) used Anthropic Interviewer to learn about people's experiences with AI. It ran Sep 29 to Oct 6 for Free, Pro and Max users on Claude and Claude Code. Accounts had to be at least two weeks old.
That's the week. The next issue lands Sunday, Oct 11.
Reply and tell me what you're building. I read every reply.
— Sam