Anthropic AI Updates: September 18, 2026

1. Projects Became a Coordinator Over Parallel Cloud Threads

Anthropic rebuilt Claude Projects around a coordinator that scopes a request, splits it across threads, and assembles the results, replacing the folder-and-handoff model. Each thread is a Claude Code cloud session working on its own branch and its own copy of the repo, so connected repositories let threads open pull requests, run tests, and resolve merge conflicts through normal git workflows. Threads can spawn their own subagents and workflows for deeper tasks. Source

2. The Redesign Adds a Shared Project Memory and Rolls Out by Tier

All threads in a project draw on a common memory that accumulates decisions and relationships over time, such as a release date moving or who to consult before touching a particular service, which cuts how much context has to be restated per prompt. The beta opened to select Claude Pro and Max subscribers using cloud sessions in Claude Code, with broader Pro and Max access planned within a week and Team and Enterprise after that. Existing projects on Pro and Max keep working unchanged during the transition, and there is a waitlist for Pro and Max users without access. Source

3. The Life Sciences Verification Program Trades Verification for Looser Biology Safeguards

Anthropic launched a program giving verified life science organizations access to Mythos, Opus, and Sonnet with safeguards tuned to be more permissive for biology work than the default configuration. Standard Use grants team-wide access renewable annually for drug discovery, clinical development, manufacturing, and regulatory work; a High-risk Use add-on removes all life sciences blocking safeguards for a specific project and renews every six months, though it is currently limited for Mythos pending coordination with the U.S. government. Verification covers research credentials, security standards, and ethical oversight, cyber classifiers stay active, and enrolled organizations accept 30-day data retention for offline monitoring. Xaira Therapeutics, Edison Scientific, and Manifold Bio are named in the announcement. Source

4. Claude Sped Up 30 Open-Source Biomolecular Models by Roughly 4x in Four Weeks

Anthropic published results from Claude Science in which Claude optimized more than 30 open-source biomolecular models in just under four weeks, averaging about 4x faster with minimal precision loss and about 1.6x with bit-identical outputs. Claude also wrote FlashPairformer kernels that beat field standards by 2.7x to 3.2x on key operations and added a memory-efficient mode that predicts molecular systems above 10,000 tokens on a single GPU node, tested on structures including human mitochondrial complex I and ribosomes. On protein design, it reached comparable performance using roughly two orders of magnitude fewer GPU hours, with total GPU and token costs around $150, scored with ipSAE and DockQ. Source

5. Anthropic Proposed Three Metrics for How Fast Labs Are Moving, and Published Its Own Numbers

Anthropic argued that capability benchmarks miss how models are built and proposed tracking AI-led R&D automation, oversight of AI agents, and compute allocation instead. Against its own August 2026 snapshot, Claude led 26% of Anthropic’s AI research work, up from under 1% in February 2026; roughly 30,000 agents operated on the platform with 0.002% of actions blocked by online monitors; and about 6% of AI research compute went to safety work in the measured week. Source

6. Claude Code 2.1.275 Synced claude.ai Skills to the Terminal and Added a Send-Now Key

Claude Code 2.1.275 started syncing the skills and plugins enabled on a claude.ai account into terminal sessions signed in with it, with syncClaudeAiSkills: false and syncClaudeAiPlugins: false as opt-outs. It also added a send-now key (ctrl+enter, or ctrl+x ctrl+s) that interrupts the current turn and flushes all queued messages at once, gateway sign-in that names the signed-in account for confirmation before the credential is saved, a startup warning when a configured otelHeadersHelper fails so silently unexported telemetry is noticed, and /plugin install <plugin> --marketplace <source>. A follow-up 2.1.276 shipped hours later to fix a 2.1.275 regression that failed every request against a proxy or gateway ANTHROPIC_BASE_URL. Source