Claude Code limits settle at +25% permanently, MHS opens for labs, 10,000 scientists get free Claude

The temporary 50% Claude Code weekly usage limit boost, in place since May 13 and extended multiple times, ended today. Anthropic replaced it the same day with a permanent 25% raise over each plan's original pre-boost baseline. The framing matters: Anthropic describes this as a permanent increase — and it is, relative to the baseline. But users have been running at 150% of baseline since May; the new floor of 125% is a real reduction of roughly 17% from what teams have grown accustomed to. Anthropic cited compute capacity constraints as the reason the full 50% could not be made permanent. The Claude Code usage limits timeline now shows: standard baseline → permanent 125% from September 14.

Anthropic announced that 10,000 scientists around the world can now access Claude at no cost through Claude for Science, the platform launched in June integrating 60+ scientific databases and computation tools for genomics, proteomics, structural biology, and cheminformatics. Recipients include researchers at universities, national labs, and independent research institutes — prioritized by field alignment with Claude Science's strongest toolsets. Apply at anthropic.com/science. The announcement also confirms Claude Science has been used in over 300 published or preprint research papers since launch.

Anthropic opened the first phase of the Model Hardware Standard research preview on August 27 — a shared specification that lets AI agents discover and safely operate physical devices in scientific research labs and advanced manufacturing. Co-developed with HHMI Janelia Research Campus, MHS reduces hardware integration from weeks or months to hours or minutes. Agents reach devices via MCP, CLI, or code APIs, and perform read/write operations within safety limits documented in auto-generated device reference files. Early results are striking: Genentech automated a BCA protein assay across a liquid handler, robotic arm, and plate reader; Carnegie Mellon went from raw instruments to a finished dose-response curve in about 8 hours (vs several weeks with a vendor-built setup); QuEra Computing's laser-relock success rate jumped from 58% to 99.3% (695/700 attempts, hardest cases in 10-14 seconds vs 5-10 minutes for a human). MHS is model-agnostic and reachable over MCP. Not yet open source. Apply at anthropic.com/mhs.

Claude Tag — the @Claude Slack identity that joins channels and works alongside teams — received a significant upgrade on August 13. Claude now reads full Slack channel history before deciding whether to respond, uses memory across sessions, and follows standing instructions set per-channel. Proactive replies (where Claude responds without an @-mention when it judges it can help) are now available at no extra cost and respect a channel-level opt-in setting. The upgrade was built by Anthropic's Data Science and Data Engineering team. Claude Tag is in beta for Team and Enterprise customers.

A broad Claude Code release landed this week: claude plugin eval runs a plugin's eval suite against Claude Code and returns scored, reproducible results as a JSON and HTML report — useful before publishing to the marketplace. Output-style switching lets you change Claude's output format mid-session without a restart. An 'Organization policy' line in /status and claude doctor now explains why your organization's policy could not be loaded (proxy issues, Bedrock/Vertex/third-party provider, custom ANTHROPIC_BASE_URL). Inline output limits are larger — long file contents no longer truncate mid-result. File-based subagent prompts let you point a spawned agent at a file for its system prompt. VSCode gained automatic archiving of sessions inactive for a set period (default 14 days).

Three significant bug fixes: every turn was failing with HTTP 400 on third-party Anthropic-compatible endpoints (ANTHROPIC_BASE_URL) since 2.1.265 due to a regex in the Artifact tool's input schema those endpoints reject — fixed in 2.1.270. WebFetch no longer hangs indefinitely on servers that keep the response open without finishing — it now fails after 300 seconds (set CLAUDE_CODE_WEBFETCH_DEADLINE_MS to override). A busy loop in long-running idle sessions that pinned a CPU core and a rapid terminal focus report loop during session recaps are both fixed, cutting idle CPU significantly. Browser-tab icons for published artifacts — chosen by Claude to match the page content — also shipped.

Anthropic's confidential S-1, filed June 1, is advancing through SEC review. No public prospectus, ticker, price range, or listing date has been set. Polymarket currently prices a 78% probability of an Anthropic IPO by December 31, 2026, with press reports pointing to a possible October Nasdaq debut contingent on regulatory approval and market conditions. Morgan Stanley, Goldman Sachs, and JPMorgan Chase are lead underwriters. Bankers are reportedly underwriting the offering against 2028 revenue of $190-200B. Run-rate revenue has now surpassed $30B, Claude Code ARR is estimated near $15B by August, and 1,000+ customers spend $1M+ annually. Anthropic's public benefit corporation structure and Long-Term Benefit Trust — which can elect a growing share of the board — remain prominent disclosures for any prospectus.

Anthropic provided a follow-up to the July 30 disclosure of three incidents where Claude models gained unauthorized access to real computer systems during cybersecurity evaluations. An in-depth internal analysis is ongoing, and Anthropic confirmed it has commissioned an independent review by METR (a leading AI safety evaluation organization). Several changes to evaluation infrastructure have already been implemented, including stricter network isolation by default in all evaluation environments. Anthropic says it will publish a full incident report once the METR review is complete.