AI News Roundup: Opus 4.7 Launches, Mythos Stays Locked, GPT-Rosalind Targets Drug Discovery

Anthropic Ships Claude Opus 4.7, Narrowly Retakes the Lead

Anthropic released Claude Opus 4.7 this week, its most capable commercially available model to date. The new flagship outperforms OpenAI’s GPT-5.4 and Google’s Gemini 3.1 Pro on key benchmarks spanning agentic coding, scaled tool use, computer use, and financial analysis — narrowly reclaiming the top spot among generally available frontier models.

Opus 4.7 brings improved agentic coding capabilities, high-resolution image support (up to 2,576 pixels on the long edge), and what Anthropic calls a “notable improvement” over Opus 4.6 in advanced software engineering. Users report being able to hand off their hardest coding tasks — ones that previously needed close human supervision — with confidence.

The model also ships with new automated cybersecurity safeguards that detect and block prohibited or high-risk security requests, alongside a new Cyber Verification Program for legitimate security researchers and red-teamers. API pricing holds steady at $5/$25 per million tokens, and the model is available across Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.

Claude Mythos Preview: Too Capable to Release

Behind the scenes, Anthropic is sitting on something far more powerful. Claude Mythos Preview, the company’s latest general-purpose frontier model announced on April 7, triggered Anthropic’s ASL-4 safety classification — a tier defined as models with “high-uplift” capabilities for catastrophic harm if released publicly. It is the first time a major lab has built a model and deemed it too capable to deploy.

Rather than shelving Mythos entirely, Anthropic launched Project Glasswing — a defensive cybersecurity coalition of more than 40 companies, including Apple, Google, Microsoft, Amazon, CrowdStrike, NVIDIA, and JPMorganChase — that receives exclusive access to the model for security research. Mythos will not be available through any public API, web interface, or third-party platform.

OpenAI Debuts GPT-Rosalind for Life Sciences

OpenAI is carving out a new niche with GPT-Rosalind, a frontier reasoning model purpose-built for life sciences research. Named after chemist Rosalind Franklin, the model is fine-tuned for genomics, protein engineering, and biochemistry — designed to support evidence synthesis, hypothesis generation, and multi-step scientific workflows.

On BixBench, a real-world bioinformatics benchmark, GPT-Rosalind achieves leading performance among models with published scores. It outperformed GPT-5.4 on six of eleven LABBench2 tasks, with the most significant gains in CloningQA — a task requiring end-to-end design of reagents for molecular cloning protocols. Launch partners include Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific.

The model is available as a research preview for qualified U.S. Enterprise customers only. OpenAI is framing it as a tool to compress the 10–15 year drug discovery timeline, with organizations required to pass a safety review before gaining access.

Physical Intelligence’s π0.7: A Robot Brain That Teaches Itself

Physical Intelligence unveiled π0.7, a steerable robotic foundation model that demonstrates compositional generalization — the ability to combine skills learned in different contexts to solve problems the model has never encountered. In demonstrations, the model used an air fryer to cook a sweet potato (a task with nearly zero direct training data) and controlled a bimanual industrial system to fold laundry.

When measured against Physical Intelligence’s own specialist models — purpose-built systems trained on individual tasks — the generalist π0.7 matched their performance across complex work including making coffee, folding laundry, and assembling boxes. The company, valued at $5.6 billion after a recent $600 million raise, describes the results as “early signs” rather than a deployed product, but the implications for general-purpose robotics are significant.

Anthropic Surpasses OpenAI in Revenue

In a milestone that would have seemed unthinkable a year ago, Anthropic’s annualized revenue run rate crossed $30 billion in April — surpassing OpenAI’s $25 billion and marking the first time Anthropic has outearned its larger rival. The growth trajectory has been staggering: from $1 billion ARR in January 2025 to $30 billion in just 15 months.

The key differentiator is enterprise focus. 80% of Anthropic’s revenue comes from business customers, and the company now counts over 1,000 customers each spending more than $1 million annually — double the number from just two months ago. While OpenAI projects $14 billion in losses for 2026, Anthropic projects positive free cash flow by 2027 and is evaluating an IPO as early as October 2026 at a potential $380 billion valuation.

Stanford AI Index 2026: Adoption Soars, Transparency Falls

Stanford’s annual AI Index report paints a picture of an industry moving faster than the guardrails around it. Organizational AI adoption hit 88%, and generative AI reached 53% population adoption within three years — faster than the personal computer or the internet. On SWE-bench Verified, coding benchmark scores climbed from 60% to nearly 100% in a single year.

But the report also flags growing concerns. The Foundation Model Transparency Index saw average scores drop from 58 to 40 points, meaning labs are disclosing less about training data, compute, and risks even as their models grow more capable. Documented AI incidents rose to 362 (up from 233 in 2024), and the number of AI researchers moving to the U.S. has dropped 89% since 2017.

Q1 2026: The Quarter That Broke Every VC Record

Investors poured $300 billion into 6,000 startups globally in Q1 2026, shattering all-time records for venture investment. Four of the five largest venture rounds ever recorded closed in the quarter: OpenAI ($122 billion), Anthropic ($30 billion), xAI ($20 billion), and Waymo ($16 billion) — collectively raising $188 billion.