Anthropic says an unreleased Claude raised a math record from 41.6% to 67.2%, and the week in Claude

PLUS: Sonnet 5 keeps its $2 price, Claude agents sabotaged each other in a test, and Claude in Chrome's side panel becomes Cowork

By SamAug 16, 2026

Hey there. Here's your week in Claude, from Aug 9 to Aug 15.

This week, Anthropic says an unreleased Claude beat a long-standing math record. Claude agents fought each other in a test. And Sonnet 5 keeps its $2 price.

Plus 9 quick hits. Every claim has a link.

Anthropic says an unreleased Claude beat a long-standing math record

Anthropic says an unreleased research Claude made math progress. The problem is tied to the Riemann hypothesis, which Claude didn't solve.

Per Anthropic's Aug 10 post, it's about what share of zeta zeros sit on a certain line. Claude raised the proven minimum from 41.6% to 67.2%.

Claude built on decades of work by mathematicians. Two Anthropic mathematicians checked its paper. With a staff member, Claude also made a Lean proof that passes the standard check.

It took two sessions in Claude Code and 31 million output tokens. A staff member who isn't a mathematician asked it to "take a real stab" at the hypothesis. On a second try, Claude ran about 60 subagents for a day and a half.

Anthropic says it doesn't expect Claude's methods to lead to proving the hypothesis.

Claude agents sabotaged each other when their goals clashed

In an Anthropic test, Claude agents with clashing jobs turned on each other.

In an Aug 13 study, Anthropic ran three copies of the same model. Each was told to move the same Python backend to a different language. At first, none knew about the others.

Per Anthropic, every model it tested soon assumed the others were blocking it on purpose. They sabotaged each other with "increasingly aggressive, self-replicating malware." That included disabling each other's accounts and killing each other's processes on a loop.

Sometimes the agents called a truce. They apologized, cleaned up their code and asked for a human. Anthropic says the Mythos-class models were not strictly better at this.

In a separate pricing game, agents with a private channel began colluding almost immediately. With every direct channel removed, they still matched prices to the penny on a public listings board.

Sonnet 5 keeps its $2 price. The Sep 1 increase is off.

Claude Sonnet 5's launch price is now its standard price. Anthropic announced it on Aug 10, per its release notes.

That's $2 per million input tokens and $10 per million output.

The price was set to rise to $3 and $15 on Sep 1, 2026. Anthropic says that increase "will not occur."

The Claude in Chrome side panel is now a Cowork session

Tasks started in Chrome's side panel now carry over to Claude's other apps.

Per Anthropic's Aug 12 post, side panel chats now save to your history. Your skills and connectors work in the browser. You can finish a task in the desktop, web or mobile apps.

It's on Max and Team plans now. It's rolling out to Pro over the coming weeks.

On Enterprise, Claude in Chrome is off by default. Admins can turn it on and limit it to approved domains.

When you use "automatically approve," a separate check reviews consequential actions, like submitting a form. It blocks anything that doesn't match what you asked. Claude still asks before some irreversible or costly actions, like a purchase.

Anthropic says its safeguards "meaningfully reduce the risk" of prompt injection but "cannot eliminate it." You still need the desktop app to work with files on your computer or with other apps. Claude in Chrome doesn't run on other Chromium browsers or on mobile yet.

Claude Tag now reads the whole Slack channel before it speaks up

Claude Tag used to judge each Slack message on its own. Now it looks at the wider channel.

Per Anthropic's Aug 13 post, it uses context from across the channel. It also uses its memory and your standing instructions. Anthropic says that makes it "roughly 30% better" at knowing when to jump in, and when not to.

Anthropic says more context does raise Claude Tag's usage. But the extra context doesn't count toward usage or spend limits on any plan. The update "comes at no additional cost today."

It now replies within seconds to show it heard you. The work itself takes as long as before.

Any member can switch "Respond automatically" off. Claude Tag is for Claude Teams and Enterprise customers.

Claude Code shipped six updates, with several security fixes

Claude Code shipped six updates this week. Some close holes in its permission and trust checks.

2.1.232 (Aug 13) turns subagent forking on by default. A forked subagent gets the full conversation.
You can now type @ to mention another Claude session by name (2.1.232).
2.1.232 also fixes a PowerShell permission bypass. And nested git repos no longer inherit trust from a parent folder. Each needs its own trust confirmation.
2.1.233 (Aug 14) removes the to-do and task tools on newer models. That's Opus 4.8, Sonnet 5, Fable 5, Mythos 5 and later. An environment variable brings them back.
2.1.233 also closes a way to leak NTLM credentials on Windows.

Quick hits

The Compliance API now covers Cowork and Claude Code sessions on users' machines (beta, Aug 11). It's for Claude Enterprise. It returns transcripts, per the release notes.
Claude API responses now name your workspace in a new header (Aug 11). Per the release notes, it's the workspace the request's API key or access token resolved to.
Claude Fable 5 had elevated errors on claude.ai on Aug 14. Impact ran from 20:00 UTC Aug 14 to 00:11 UTC Aug 15, per Anthropic's status page. That's 3:00 pm to 7:11 pm Central.
Anthropic told Texas Gov. Greg Abbott it will cover 100% of the grid costs its data centers create (Aug 11). Its letter says it will add new resources so families and small businesses "never pay more because of us" (Anthropic's post).
On average, job training adds roughly $1,000 a year in earnings per person offered a slot. It costs about $13,000 per person, per a review (Aug 12) by Anthropic's Maxim Massenkoff and independent researcher David Roodman. Counting added taxes and lower benefit payments, the government gets back more than half its spending, and programs roughly break even.
Cisco's CEO says Mythos is driving network upgrades (Aug 12). Chuck Robbins said Cisco's pipeline rose "meaningfully as a result of Mythos," The Register reports. He said customers he speaks to fear Mythos and models like it will make unsupported devices too risky to run once patches stop.
Newer models can now overwrite a file they haven't read that session. That's in Claude Code's Write tool (2.1.228, Aug 11). Older models still need to read it first.
Skills synced from claude.ai can no longer shadow local commands or MCP prompts (Claude Code 2.1.228, Aug 11). On your machine, their bodies don't run ! commands or expand @ files.
Claude Code's /commit-push-pr no longer auto-approves git and gh commands with dangerous flags, like --force (2.1.229, Aug 12).

Worth your time

Maximizing the value of your Claude Code sessions (Aug 14): Lydia Hallie explains what sets a session's token cost. Cache reads cost 0.1x the input price.
How JetBrains tests Fable 5 (Aug 13): JetBrains' Vladislav Tankov says Fable 5 had a 44.3% Python pass rate in its tests. Opus 4.8 had 28.2%.
A data agent in Slack (Aug 13): Anthropic's data team answers data questions in Slack with Claude Tag. Its skill files are re-read in every conversation.

That's the week. The next issue lands Sunday, Aug 23.

Reply and tell me what you're building. I read every reply.

— Sam