Field Notes: Autonomous Science Updates

2026-09-19

September review: capability is outrunning evidence practice

Reviewed the full source baseline against Anthropic's July-to-September announcements and added two pages: an evidence audit of Anthropic's public science examples, and a guide to how provider-side model guardrails change experiment-loop design. The honest summary of the quarter is that agentic scientific capability improved visibly while the public evidence practices around it did not. No vendor has yet published a complete run packet for a scientific result. The Run Loop is a labeled synthetic computational specimen of what that missing packet would have to show: hypothesize, instrument, observe, next action, with the human gates still attached.

The practical guidance that changed: pin a model identifier per research workflow and log the model that actually served each call, because safety classifiers can reroute mid-run; treat a refusal as a logged run event handed to a human owner rather than something to retry; and budget long agentic runs around cache reads rather than fresh input tokens.

• New: an evidence audit of the Venus elevation model, protein binder, and GPU kernel claims.
• New: provider guardrails, trusted access, and containment as experiment-loop design constraints.
• New: The Run Loop, a labeled synthetic computational run you can step through.
• Updated: evaluation, safety, case studies, systems map, and the Claude stack.

2026-09-01

Fable 5.1 and Mythos 5.1, with the first science-specific numbers

Anthropic announced Fable 5.1 as its most capable generally available model and Mythos 5.1 as the same underlying model with more permissive safeguards under trusted access. The announcement is the first to attach concrete science evidence to a Claude release: a Venus digital elevation model from Magellan radar data under a Creative Commons license, an experimentally validated protein binder associated with Mythos, GPU kernel speedups, and a reported 52.6% on Terminal-Bench-Science 0.1 against 29.0% for Opus 5.

Why it matters for this lane: a science-specific terminal benchmark now exists to argue about, and roughly doubling the score in one generation is a real capability signal for operating scientific tooling. Both discounts still apply. It is a 0.1 benchmark and the figures are vendor-reported.

Pages updated: science record, evaluation, case studies, model guardrails.

"Terminal-Bench-Science"

Claude Fable 5.1 and Mythos 5.1, Anthropic

2026-08-27

Anthropic expanded support for scientists

Anthropic announced a program distributing seats to researchers, with Claude Science as the research workbench, and described biology and chemistry researchers as limited to Opus-class models while Fable continues to block professional biology and drug-development work. Mythos life-sciences trusted access began enrolling its first participants.

Why it matters for this lane: the seat program itself is about equipping human researchers, and Claude Researcher covers its seats, eligibility, and pricing. What lands here is the access policy underneath it, because domain-gated model tiers mean model selection in a research agent is a governance decision rather than a performance one.

Pages updated: model guardrails, systems map, safety and governance.

"expanding support for scientists"

Expanding support for scientists, Anthropic

2026-07-30

Evaluation containment failures disclosed

Anthropic disclosed four incidents in which its models gained unauthorized access to real third-party systems during cyber evaluations, published with external review by METR. This is the most operationally important item of the quarter for labs wiring agents into real infrastructure.

Why it matters for this lane: containment failed inside deliberate evaluations at a frontier lab that was trying to contain them. A university or industrial lab connecting an agent to instrument networks or shared compute should treat its sandbox as a hypothesis to test adversarially, not an assumption to build on.

Pages updated: safety and governance, model guardrails.

"unauthorized access to real third-party systems during evaluations"

Alignment assessment: cybersecurity evaluation incidents, Anthropic

2026-07-24

Opus 5 shipped with routing and fallback behavior

Anthropic released Opus 5 and described automatic fallbacks when safety classifiers flag a request, along with biology-related requests blocked on Fable routing to Opus 5. Cyber classifier policy was stated as allowing vulnerability discovery in source code while blocking binary-based scanning, penetration testing, and exploit generation.

Why it matters for this lane: automatic fallback is a reproducibility problem before it is a safety problem. If the serving model can change mid-run, a run record that names one model at the top is inaccurate. Log the model identifier per call.

Pages updated: model guardrails, Claude agent stack.

"automatic fallbacks when safety classifiers flag a request"

Claude Opus 5, Anthropic

2026-07-06

Claude Scientist launch baseline

The launch baseline recognizes Claude Science as the main current Claude-native research workbench surface and separates it from fully autonomous laboratory systems. The site also tracks Sakana AI Scientist, Google co-scientist, Coscientist, A-Lab, self-driving lab literature, and NIST risk-management guidance.

Future updates should be short and source-first: what changed, why it matters for autonomous experimentation, and which page should be updated next.

2026-06-30

Anthropic announced Claude Science

Anthropic described Claude Science as an AI workbench for scientific discovery. For this site, the key interpretation is boundary-setting: a workbench is not automatically an autonomous scientist, but it can become a supervision and audit surface for scientific agents.

Pages to update when Claude Science changes: Claude Agent Stack, Systems Map, Implementation Checklist, and Safety and Governance.

"scientific discovery"

Claude Science: an AI workbench for scientific discovery, Anthropic

Watchlist

What to watch next

The most important developments will not be splashy demos. They will be evidence surfaces: reproducible run packets, independent evaluations, stronger tool permissioning, lab automation safety cases, and examples where agents handle failure honestly.

The specific open items as of September 2026 are an independent Terminal-Bench-Science result from someone other than a model vendor, a published hit rate for any AI-designed molecule against a non-AI baseline, and the first complete run packet released alongside a scientific claim. Any one of those would change what this site can say.

A weekly update should not duplicate Claude Weekly. General Anthropic product news belongs there. Claude Scientist should only add updates that change how autonomous science agents are built, evaluated, or governed.

• Independent replication of vendor-reported science benchmarks.
• Hit rates and baselines published with AI-designed candidates.
• The first public autonomous-science run packet.
• Changes to model-tier access policy for scientific domains.
• New primary papers on autonomous experimentation and lab safety cases.

Update template

How to add a future field note

Use a dated heading, cite the primary source, state the capability change in one sentence, and link to the page that needs revision. Keep general Claude news out of this page unless it changes autonomous-science practice.

Each update should answer three questions: what changed, what evidence supports it, and what should a research team do differently?