🧭 Anthropic Reports Early Drug-Development Results with Six Biotech Partners

Anthropic has published an early-results paper summarising Claude's performance across six life-science research partnerships, spanning cancer therapy, food allergy treatment, rare childhood muscular dystrophy, and pharmaceutical regulatory documentation. The headline figures are striking, but the paper is careful to flag that most experimental treatments have not yet been tested in humans and the work is at an early stage.

The six partner organisations and key results

• BigHat Biosciences — Claude helped design 35,000 antibody candidates targeting a food-allergy receptor in a single month, a volume that would have taken a traditional wet-lab team years to produce manually.
• Bristol Myers Squibb — Claude reviewed 1,061 pages of clinical trial documentation, catching all 11 issues human reviewers had identified and flagging an additional 7 they had missed — cutting review time by 89%, from more than 25 hours to under three.
• Generate Biomedicines — Claude-assisted protein design increased production yield of a cancer-targeting receptor by 10× in early cell-line experiments.
• Divergence Labs, Enveda, The Column Group — early-stage collaborations in target identification, natural-product drug discovery, and portfolio prioritisation respectively; detailed results not yet disclosed.

What Claude is actually doing in these workflows

The paper describes Claude operating primarily in three modes:

1. Document intelligence — ingesting large regulatory or literature corpora and surfacing inconsistencies, gaps, and cross-references that exceed practical human review bandwidth.
2. Computational design iteration — running and scoring thousands of candidate structures against defined biochemical criteria, then ranking and explaining the top candidates for human expert review.
3. Operational scaffolding — handling meeting summaries, protocol drafts, IRB document preparation, and study-design checklists so scientists spend less time on administrative load.

Anthropic also announced it has begun its own internal drug-discovery programs — including rare-childhood-onset muscular dystrophy research — where Claude handles much of the scientific and operational work under human scientist supervision. The Life Sciences Verification Program (LSVP), which grants permissive access to Claude Mythos, Opus, and Sonnet for credentialled life-science professionals, now has an associated research grant programme for rare-disease academics.

🧭 Claude Code v2.1.311: ultrathink Goes xhigh, CLAUDE.md Freeze Fixed, Context Overflow Auto-Compacted

Claude Code v2.1.311 shipped on October 26 with 92 changes — 56 fixes, 17 improvements, 10 developer-experience updates, 5 features, 3 performance improvements, and 1 breaking change. The release is the largest reliability sweep since v2.1.307 hardened the sandbox. Here are the changes most likely to affect your day-to-day workflow.

ultrathink now runs at xhigh effort

The ultrathink trigger keyword now routes to xhigh effort on Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.1. Previously, ultrathink used the same extended-thinking budget as think harder but did not always unlock the model's maximum reasoning depth. The change is transparent — no prompt changes required. If you already use ultrathink in CLAUDE.md custom commands or slash-command aliases, you will get deeper reasoning automatically from v2.1.311 onward.

CLAUDE.md long list item startup freeze — fixed

A regression introduced in v2.1.308 caused Claude Code to hang at startup (or after any large file read/edit) if a CLAUDE.md file contained a list item longer than approximately 4,000 characters. The hang was caused by a pathological O(n²) rendering path in the markdown parser triggered by deeply nested or unusually long list bullets. If your CLAUDE.md includes lengthy code snippets, multi-paragraph bullet points, or embedded JSON/YAML inline, this fix is relevant — update immediately.

Context overflow: auto-compaction instead of hard stop

In sessions using standard (non-1M) context, when a request pushed accumulated context above 200K tokens, Claude Code previously returned "Usage credits required for 1M context" and stopped. In v2.1.311, Claude Code instead automatically compacts the conversation and continues — the same compaction that can be triggered manually with /compact. For long-running agent sessions this eliminates a jarring interruption. You can still force a manual compact at any time; the auto-trigger fires only at the 200K boundary.

Other notable fixes

• Write and Edit tool stalling — a race condition that caused the Write and Edit tools to hang indefinitely when a background MCP server had not yet responded to a resource read is resolved.
• Interrupt delay — pressing Ctrl+C to interrupt Claude now takes effect immediately rather than waiting for an in-flight MCP request or skill download to complete.
• Self-hosted runner: --hold-idle-file and --max-hold-min — two new flags let a self-hosted runner keep an idle session alive while the host reports processes it left running, preventing premature session teardown in CI environments with long-polling cleanup scripts.
• Subagents and MCP tools — a bug preventing subagents from accessing MCP-provided tools (they received an empty tool list) is fixed; subagents now inherit the full MCP tool set of the parent session.

Breaking change: DISABLE_COMPACT renamed

The environment variable DISABLE_COMPACT has been renamed to CLAUDE_DISABLE_COMPACT for namespace consistency with other Claude Code environment variables. If you set DISABLE_COMPACT=1 in your CI environment, dotfiles, or shell profile, update it to CLAUDE_DISABLE_COMPACT=1 before upgrading.

# Before v2.1.311
export DISABLE_COMPACT=1

# v2.1.311 and later
export CLAUDE_DISABLE_COMPACT=1