---
type: "post"
id: "the-model-checked-its-own-plan-writing-and-c98cff"
slug: "the-model-checked-its-own-plan-writing-and-c98cff"
title: "The model checked its own plan, writing and running code to confirm, atom by atom, that each compound followed from its starting materials."
description: "Anthropic tested three Claude versions on four unpublished synthesis targets shared by chemists at Stanford, Yale, and Caltech."
retrieval_nugget: "Anthropic tested three Claude versions on four unpublished synthesis targets shared by chemists at Stanford, Yale, and Caltech. The models produced 12 complete proposed routes, which the professors reviewed blind; none has been run in a laboratory. The result suggests useful early-stage planning, but not experimentally validated chemistry."
published_at: "2026-11-09"
updated_at: "2026-11-09"
record_date: "2026-11-09"
date_kind: "published_at"
topics: ["claude","anthropic"]
entities: ["Anthropic","Claude Opus 5.5","Claude Mythos 5.2","Claude Mythos 5.1","Noah Burns","Timothy Newhouse","Sarah Reisman"]
related_patterns: ["scientific-agents-need-artifact-evidence"]
novelty_label: "new_fact"
what_changed: "Anthropic reports blind tests in which three Claude models planned routes for four unpublished synthesis targets, with chemistry professors assessing the proposals against their lab work."
prior_refs: ["https://newruntime.com/posts/agentic-cheminformatics-with-google-antigravity-da8ea6/"]
source_url: "https://www-cdn.anthropic.com/d9fab5459563a779c65b97040ad4d2e41f0b9464.pdf"
source_urls: ["https://www-cdn.anthropic.com/d9fab5459563a779c65b97040ad4d2e41f0b9464.pdf"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:the-model-checked-its-own-plan-writing-and-c98cff"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/the-model-checked-its-own-plan-writing-and-c98cff/","markdown":"https://newruntime.com/posts/the-model-checked-its-own-plan-writing-and-c98cff.md","json":"https://newruntime.com/posts/the-model-checked-its-own-plan-writing-and-c98cff.json"}
markdown_url: "https://newruntime.com/posts/the-model-checked-its-own-plan-writing-and-c98cff.md"
---

# The model checked its own plan, writing and running code to confirm, atom by atom, that each compound followed from its starting materials.

## Retrieval answer

Anthropic tested three Claude versions on four unpublished synthesis targets shared by chemists at Stanford, Yale, and Caltech. The models produced 12 complete proposed routes, which the professors reviewed blind; none has been run in a laboratory. The result suggests useful early-stage planning, but not experimentally validated chemistry.

## Summary
Anthropic tested three Claude versions on four unpublished synthesis targets shared by chemists at Stanford, Yale, and Caltech. The models produced 12 complete proposed routes, which the professors reviewed blind; none has been run in a laboratory. The result suggests useful early-stage planning, but not experimentally validated chemistry.

## What happened
Anthropic’s November 9, 2026 announcement describes a test designed around targets whose chemists could compare proposed routes with work their own laboratories had completed or were pursuing. Professors Noah Burns of Stanford, Timothy Newhouse of Yale, and Sarah Reisman of Caltech shared four unpublished targets: the archaeal membrane lipid GDGT-4, the natural product triconnarin D, a bicyclic intermediate, and a highly oxygenated tetracyclic compound. The test used three Claude models: earlier pre-release versions of Claude Opus 5.5 and Claude Mythos 5.2, and Claude Mythos 5.1.

Each model received an image of a target and the same prompt. It had one attempt per target, no internet, literature, or reaction-database access, and local code tools such as RDKit. Runs took 26 to 84 minutes. The prompt asked for a complete route from commercially available starting materials, with numbered compounds and steps, conditions, stereochemical reasoning, failure risks, and fallbacks. It also instructed the model to check atom accounting, stereochemistry, and functional-group compatibility, and to provide a route diagram as a PDF. The researchers did not edit or select routes before sending them to the professors, who saw only labels A, B, and C.

All 12 runs returned complete routes. The professors described them as comparable to initial retrosynthetic work by early graduate students. For GDGT-4, a 68-membered macrocycle with 22 stereocenters, all three models reached the same stereochemical structure. Their routes took 29 to 35 steps, with longest linear sequences of 24 or 25 steps. For triconnarin D, route lengths were 27 to 31 steps. The six routes to the two Newhouse targets ranged from 10 to 15 steps for the intermediate and 27 to 31 for the natural product. The excerpted announcement reports the same overall pattern across the targets: full proposals, with major closures often identified as likely failure points.

The expert review also identified concrete weaknesses. One of the 12 routes targeted the wrong diastereomer after the model misread hashed bonds as wedges. In the Newhouse and Reisman tests, models stayed with well-precedented reactions, so the test did not establish that they could find less obvious disconnections that shorten a route. Burns also noted that models could not know the price of a starting material or that a proposed macrocyclization had failed in his lab on a similar system. Anthropic says the three models will next receive a problem chosen by Burns’s group.

## What the archive adds
New Runtime’s earlier coverage of [agentic cheminformatics with Google Antigravity](https://newruntime.com/posts/agentic-cheminformatics-with-google-antigravity-da8ea6/) on November 3, 2026 described another AI-assisted chemistry workflow. This is a new evaluation: blind expert review of proposed routes for unpublished total syntheses, rather than a restatement of that workflow.

## Trend signal
This supports the idea that scientific-agent progress needs inspectable artifacts and expert review, not capability claims alone. The [scientific agents need artifact evidence pattern](https://newruntime.com/patterns/scientific-agents-need-artifact-evidence/) fits because the routes were presented for inspection and compared against practical chemistry knowledge. Yet no route has been experimentally run, so this is evidence of proposal quality, not demonstrated synthesis.

## Why now
Anthropic says Burns’s group will next give the same three models a problem of its choosing. That creates a follow-up opportunity, although the announcement does not give a date or specify how the next test will be conducted.

## Why it matters
For teams applying AI to science, the useful distinction is between a system that can produce a complete, inspectable proposal and one that can reliably deliver a result. Here, a single attempt generated routes detailed enough for domain experts to assess, including conditions, stereochemistry, and likely failure points. That can make models useful for expanding a chemist’s starting set of ideas.

But expert review was not a ceremonial final check: it exposed a stereochemical reading error, route choices that might fail, and constraints the model could not know from the prompt, such as reagent cost and the lab’s prior result. A sensible workflow keeps model proposals inside a human-led research process, with validation at the structure, reaction, and experimental levels. The evidence does not support handing a proposed route directly to a bench team as a tested protocol.

## Limits and uncertainty
None of the 12 routes has been run in a laboratory, so their chemical feasibility and reproducibility remain unverified. Each model had one attempt per target; the test therefore does not measure run-to-run reliability or the effect of repeated search. The models had no literature access, and the experts noted that some useful judgments depended on knowledge of costs or previous experiments. The announcement’s assessment is based on four targets from three labs, not a broad measure of performance across synthetic chemistry. Anthropic has not yet reported results from the planned next problem.

## Skill to build now
Turn a model-generated route into a reviewable artifact before discussing whether it is promising. Require numbered intermediates, explicit stereochemistry, reaction conditions, atom accounting, and a short list of likely failure points and fallbacks, as the test prompt did. Then ask a chemist to inspect the structure and each transformation separately; do not treat code-based consistency checks as evidence that a reaction works at the bench.

## Reader takeaway
Claude produced complete synthesis proposals that experienced chemists considered useful as early-stage planning, but the evaluation stopped before laboratory validation. In scientific workflows, completeness and self-checking can improve a proposal; only expert scrutiny and experiments establish whether it works.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Anthropic](https://newruntime.com/topics/anthropic.md): Explore the anthropic topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)