Anthropic News, September 2026: Opus 5.5, Sonnet 5.5, and Everything Else

September 2026 was Anthropic’s busiest month of the year: three new frontier models, a price cut on its flagship tier, bigger subscription limits, a biology discovery made largely by Claude agents, and new programs for scientists and outside safety evaluators. Here is everything that shipped, newest first, with links to our full guides.

September 28: Claude Sonnet 5.5

Anthropic’s new workhorse model keeps Sonnet 5’s $2/$10 pricing while generating output more than 30% faster and cutting the cost of a typical task by up to 30% through fewer tokens and tool calls. On Anthropic’s launch benchmarks it scores 1,844 on GDPval-AA v2.1 against Opus 5.5’s 1,846, and 70.6% on Terminal-Bench 4.0. It’s available on all platforms, and in Claude Code via /model sonnet. Full details: Claude Sonnet 5.5 guide and Sonnet 5.5 vs Opus 5.5.

September 23: Claude Finds a New CRISPR-Like Enzyme System

Anthropic reported that Claude agents discovered a previously undescribed family of genetic systems in bacteriophages — viruses that infect bacteria. Anthropic calls them array-associated reverse transcriptases: a reverse transcriptase (an enzyme that copies RNA into DNA), a partner gene, and a repeating array of non-coding DNA next to it that resembles a CRISPR array.

• How it was found: roughly 950 Claude agents searched genomic data over 21 hours, using about 210 million tokens, coordinated through Claude Science and Claude Code.
• How it was checked: human scientists at Anthropic’s Bay Area lab expressed the protein and characterized it biochemically and structurally. Early experiments show the array produces distinct short RNAs — a hint of CRISPR-like programmability.
• What’s still unknown: the system’s actual function. The findings are out as a preprint and have not yet been peer-reviewed.

It’s one of the clearest examples yet of AI-generated hypotheses leading to a real lab result. For the tools behind it, see Claude Science.

September 22: Claude Opus 5.5 and Bigger Usage Limits

Opus 5.5 cut Opus pricing from $5/$25 to $4/$20 per million tokens and, according to Anthropic, costs about 40% less than Opus 5 on typical workloads once token efficiency is counted. It outscored Fable 5.1 on every benchmark in Anthropic’s launch table, including 66.4% vs 55.8% on Terminal-Bench 4.0, and became the default model in Claude Code on every paid plan. A fast mode runs up to 2.5× faster at $8/$40.

For subscribers, Anthropic raised five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans, and added a one-time rate-limit reset that subscribers can save and use whenever they choose. Details: Claude Opus 5.5 guide and Opus 5.5 vs Fable 5.1.

September 22: Claude in the Ebola Response

In a feature titled The Situation Report, Anthropic described how Claude is supporting the response to a Bundibugyo Ebola outbreak in the eastern Democratic Republic of Congo. As of September 19, the outbreak had reached 7,672 confirmed cases and 3,699 deaths across 63 of the country’s 167 health zones.

• WHO Africa: a Claude skill extracts case numbers from health-zone reports and flags anomalies, so a daily situation report that took all day can be produced in under an hour.
• Forecasting: WHO analysts use Claude to run several disease models side by side for logistics planning.
• Vaccines: CEPI uses Claude to organize complex data for faster expert review of vaccine proposals, with humans making the scientific calls.
• Genomics: at the INRB lab, Claude Science assembles virus genomes and family trees from sequencing data — work that normally requires a specialist programmer.

September 18: $1 Billion Each for Embedded Safety Evaluation

Anthropic and Accenture announced a partnership for “embedded evaluation”: independent evaluators, working through Accenture’s Faculty team, get employee-level access inside Anthropic to observe model development, red-team models, test safeguards, and report findings publicly. Anthropic funds the work and remains responsible for model safety. Both companies expect to invest at least $1 billion each in building this capacity over the next five years.

September 17: Life Sciences Verification Program

A new beta program lets verified academic labs, startups, pharmaceutical companies, and research institutions use Claude for biology work its safeguards normally block — drug discovery, research biology, clinical development, manufacturing, and regulatory affairs. There are two grant types: Standard Use, renewed annually, and High-risk Use, which is project-specific, renewed every six months, and removes life-sciences safeguards. Cybersecurity and other safeguards stay in place. Applicants are vetted on research credentials, security standards, and ethical oversight, and Anthropic is working with the US government to widen high-risk access to Claude Mythos 5.1. The program launched for teams and institutions, with individual plans to follow.

September 10: Threat Intelligence Report

Anthropic’s latest misuse report covers cases disrupted between December 2025 and August 2026 across seven harm categories, and introduces a new naming scheme for AI-abusing actors (“Generative Threat Groups”). Highlights include a suspected Russian state-backed group that used Claude to rebuild malware automatically when detected; a criminal group that scanned 1.8 million Android apps for hardcoded secrets; nine influence operations spanning six continents, one running 70 fake news sites that published 8,913 articles in 20 languages; and a 1,000-account operation targeting Malaysian voters. Anthropic banned the accounts, shared indicators of compromise, and worked with authorities. Its central warning: “sophisticated attacks no longer require sophisticated attackers.”

September 1: Claude Fable 5.1 and Mythos 5.1

The month opened with Claude Fable 5.1 and the invite-only Claude Mythos 5.1, at $10/$50 with 75% cheaper cache reads, plus a preview of Enterprise Frontier Safeguards, which let enterprises run safeguards on their own cloud infrastructure. Our launch coverage has the details.

What’s Next

• Claude Haiku 5.5 is due in the coming weeks, completing the 5.5 family.
• Cyber Verification Program expansion is coming, which should loosen Opus 5.5’s default routing of security work to Opus 4.8.
• Claude Haiku 4.5 has a retirement commitment of no sooner than October 15, 2026; no deprecation notice has been issued yet.

For the whole current lineup in one place, see Claude Models Explained.

Frequently Asked Questions