Anthropic Acquires Biotech AI Startup Coefficient Bio for $400 Million
Anthropic has acquired Coefficient Bio, a stealth AI biotech startup, in an all-stock deal worth approximately $400 million. The startup, founded just eight months ago by Samuel Stanton and Nathan Frey — both veterans of Genentech’s internal computational drug discovery unit Prescient Design — had been building a platform that uses AI for drug R&D planning, clinical regulatory strategy, and novel drug opportunity identification. The team, fewer than 10 people, will join Anthropic’s healthcare and life sciences group.
The acquisition follows Anthropic’s October 2025 launch of Claude for Life Sciences and signals an aggressive move into pharmaceutical and biotech markets. The founders’ deep experience in protein modelling and biomolecular representation is expected to add specialized molecular biology capabilities to Claude that go well beyond general scientific reasoning. Anthropic has not disclosed a specific product roadmap for the combined team.
Claude Mythos Enters Limited Early Access — Anthropic Flags “Unprecedented Cybersecurity Risks”
After a high-profile accidental leak of internal source code exposed the existence of a model called “Claude Mythos,” Anthropic confirmed the model is now in limited early access with a select group of enterprise partners. Described internally as “by far the most powerful AI model we’ve ever developed” and a “step change in capabilities,” Mythos is notable for its autonomous multi-step reasoning — it plans and executes sequences of actions across systems without waiting for human input at each stage.
The draft blog post that was inadvertently made publicly searchable stated the company believes Mythos poses “unprecedented cybersecurity risks,” a rare and unusually direct safety disclosure from an AI lab about its own unreleased model. Anthropic is working with early access customers specifically to evaluate and mitigate those risks before broader rollout. Wider API access for enterprise partners, then Claude.ai availability, is expected later in 2026.
Anthropic Announces Major Infrastructure Expansion With Google and Broadcom
Anthropic has extended its partnership with Google and Broadcom to secure several gigawatts of computing capacity to power the next generation of Claude models. The expansion is planned to ramp from 2027 and is designed to meet rapidly increasing demand for AI services at scale. Anthropic described it as one of the most significant infrastructure commitments in the AI sector to date.
The announcement comes as Anthropic manages growing compute pressure from power users running agentic workloads — the same dynamic that forced the recent OpenClaw subscription change. Locking in gigawatt-scale infrastructure well in advance signals the company is planning for substantially higher throughput requirements as autonomous agent usage grows.
OpenClaw Subscription Ban Sparks Developer Backlash and Fork Activity
Anthropic’s April 4 announcement that Claude subscriptions can no longer power third-party harnesses like OpenClaw has landed hard in the developer community. A Hacker News thread tracking the change drew hundreds of comments, with many devs reporting disrupted workflows and debating whether to migrate to direct API billing or switch to competing models entirely. OpenClaw’s open-source fork, which began receiving community contributions after founder Peter Steinberger joined OpenAI in February, saw a significant spike in GitHub activity.
Anthropic’s head of Claude Code Boris Cherny said subscriptions “weren’t built for the usage patterns of these third-party tools,” framing the change as a cost and capacity management decision rather than a competitive one. Users who want to continue with third-party harnesses now need to pay via the API or Anthropic’s new “extra usage” pay-as-you-go billing. The move exposes a structural tension: flat-rate subscriptions and unlimited agent loops are fundamentally incompatible.
Model Retirements and API Deprecation Deadlines: What to Know
Anthropic has confirmed two upcoming API deprecation deadlines developers need to plan for. Claude Haiku 3 will be retired on April 19, 2026 — migrate to Haiku 4.5 now if you haven’t already. The 1M token context window beta for Claude Sonnet 4.5 and Claude Sonnet 4 ends April 30, 2026; after that date, the context-1m-2025-08-07 beta header will have no effect. Sonnet 4.6 and Opus 4.6 support long contexts natively without a beta header.
Separately, the new inference_geo parameter is now available on the API, letting teams lock model inference to US-only processing at 1.1x standard pricing — useful for compliance and data residency requirements. The Models API also now returns max_input_tokens, max_tokens, and a capabilities object for each model, making it easier to build routing logic that adapts to model constraints automatically.
Claude.ai Outage on April 6 — Login Errors Hit Claude.ai and Claude Code
From 15:00 to 16:30 UTC on April 6, users experienced elevated login errors across Claude.ai and Claude Code. Anthropic’s status page confirmed the incident was resolved within 90 minutes. The outage coincided with the continuing OpenClaw migration wave, which may have contributed to elevated traffic patterns. No data loss or security incidents were reported.
Anthropic Partners With Australia on AI Safety and Economic Impact Tracking
Anthropic announced a partnership with the Australian government focused on AI safety research and tracking the economic impacts of AI deployment. The collaboration includes joint research on AI risk evaluation and an economic monitoring framework designed to measure how AI adoption affects labor markets, productivity, and industry structures. Australia joins the UK as a government with a formal Anthropic partnership — a growing list that stands in contrast to Anthropic’s ongoing legal conflict with the US Pentagon.
Anthropic Is Playing a Very Long Game — And It Shows
Look at today’s news as a single picture. Anthropic paid $400M for a sub-10-person biotech team. It locked up gigawatt-scale compute through 2027+. It’s testing a model it says poses unprecedented cybersecurity risks — and it’s being honest about that publicly. It’s winning government AI partnerships in the UK and Australia while fighting the US military in court over safety guardrails it refuses to remove.
This is not a company acting like it needs to win the next quarter. The Coefficient Bio deal, the infrastructure partnership, the Mythos risk disclosures — these are all bets on being the dominant AI platform for high-stakes, specialized applications in five to ten years. The OpenClaw subscription change is annoying for power users today, but it’s also consistent: manage costs, manage capacity, build the infrastructure for sustainable scale. The tensions are real, but the strategic coherence is there if you zoom out.