Steady Sonnet 5 Pricing, Safer Fable 5 Biology Access, and a Claude-Assisted Math Result

Claude Code ships three consecutive stability patches

Between versions 2.1.225 and 2.1.228, Claude Code addressed a cluster of session-integrity bugs. Version 2.1.225 added gateway spend-limit visibility to the usage-warning message (naming the cap, its reset time, and the operator's note), extended the workspace-trust prompt to `claude agents` running in untrusted directories, and fixed a transient 401 error that was overwriting a long-lived OAuth token with a short-lived one, breaking headless sessions until restart. Version 2.1.227 fixed a feature-flag evaluation bug that could wrongly prompt Max-plan users to enable usage credits for Fable when a session started with an expired login token, a GitHub Actions regression that made every Bash command fail under claude-code-action with allowed_non_write_users on GitHub-hosted runners, and a /tui bug that resurrected conversations rewound past their first message.

Version 2.1.228 closed out the run with fixes for a rare internal layout error that could freeze a session's redraw entirely while the underlying process kept running, a Windows-specific failure to locate git/Git Bash when Claude Code was launched from a parent folder of the git installation, and a /tui issue that silently reverted a session to an earlier model after a mid-session /model change. Taken together, the sequence reads as routine hardening of session state and terminal rendering rather than new capability work.

Sources: Claude Code 2.1.225 release notes — Claude Code · Claude Code 2.1.227 release notes — Claude Code · Claude Code 2.1.228 release notes — Claude Code

Compliance API extended to local Cowork and Claude Code sessions

Anthropic's API release notes for August 11 describe a beta Compliance API expansion for Claude Enterprise organizations: a new GET /v1/compliance/apps/sessions/local endpoint lists sessions that an organization's users have run locally, with a companion per-session endpoint returning the session's transcript. This gives Enterprise administrators visibility into agentic work that executes on end-user machines, rather than only tasks run through Anthropic-hosted infrastructure, closing a gap between what was already auditable for cloud-run activity and what happens when Cowork or Claude Code run locally.

Sources: Anthropic API release notes — August 11, 2026 — Anthropic

Anthropic details a Claude-assisted argument on Riemann zeta zeros

Anthropic published a set of posts describing how Claude helped surface what it calls the 'two-thirds argument': a case in which two agent runs, combined with a review of existing mathematical literature, led to an argument that more than two-thirds of the zeros of the Riemann zeta function lie on the critical line. Anthropic frames this as part of a broader look at 'Claude's mathematical capabilities' rather than as a resolution of the full Riemann Hypothesis, which remains an open problem.

The framing Anthropic gives emphasizes process as much as outcome: agentic literature search combined with iterative agent runs, presented as a case study in using Claude as a research collaborator on a specific, narrow technical question within a famously hard open problem, rather than as a claim of a general mathematical breakthrough.

Sources: How the two-thirds argument was found: two agent runs and their literature — Anthropic · Learning more about Claude's mathematical capabilities — Anthropic · More Than Two Thirds of the Zeros of the Riemann Zeta Function Lie on the Critical Line — Anthropic · 67% of the zeroes are on the line — Anthropic

Sonnet 5 pricing frozen, Fable 5 biology fallbacks reduced

Anthropic confirmed that Claude Sonnet 5's introductory API pricing of $2/$10 per million tokens is now the standing price: the increase to $3/$15 per MTok that had been scheduled for September 1, 2026 will not take effect. For teams budgeting around Sonnet 5, this removes a previously announced cost step-up.

Separately, Anthropic said it has reworked Fable 5's biology safety classifier to cut false-positive 'fallbacks' — moments when the system switches a query to the less capable Opus 5 — by about 85% across its product surfaces in its own testing, which it expects to mean fewer interruptions on everyday questions such as interpreting lab results or learning biology, and more support for healthcare professionals on clinical tasks. Anthropic is explicit that the change narrows false positives rather than loosening the underlying safeguard: Fable 5 still falls back to Opus 5 for topics it classifies as dual-use — virology, toxicology, and molecular design — because, per its own capability assessments, the model's biology performance could give a malicious actor 'significant uplift' toward a biological weapon that is not otherwise available elsewhere, even as that same capability is what makes the model valuable to a legitimate biology researcher.

Sources: Anthropic API release notes — August 10, 2026 — Anthropic · Improving Fable 5's biology safeguards — Anthropic