Science Ships

Anthropic shipped Claude Science on June 30, its third flagship product alongside Claude Code and Claude Cowork, with tools for computational biology and drug development and day-one access for every paid subscriber. The same day, OpenAI published GeneBench-Pro, a 129-problem benchmark for scientific research judgment, and reported GPT-5.6 Sol topping it at 31.5%. Two labs converging on the same vertical on the same day is the shape worth reading, and it matters more than either announcement on its own. Science graduated from a use case to a product category on both sides at once.

Claude Science ships as a research agent, not a general model with a science plugin. It writes code and runs it on compute clusters, prioritizes reproducibility and traceability, and interfaces with the standard tooling in genetics, chemistry, and protein biology. Anthropic's demo case identified drug candidates for phenylketonuria autonomously. Internal use is directed at rare and neglected diseases. Anthropic's own capability estimate puts Claude Science at roughly a second-year grad student's level. This is a step past the Claude for Life Sciences plugins Anthropic shipped in October 2025: a full product with its own surface, name, and support model, not general model access with domain add-ons.

OpenAI's counter is a benchmark, not a product. GeneBench-Pro spans 129 problems across ten domains of computational biology, from population genetics to cancer genomics to clinical diagnostics. What OpenAI is trying to measure it calls "research taste": whether a model can pick the right analytical approach in ambiguous data, revise assumptions when a diagnostic pushes back, and know when a result is ready to act on. That is a different target than the ability to execute a predefined workflow. GPT-5.6 Sol scores 28.7% at the highest reasoning level and 31.5% with Pro mode. GPT-5, the frontier model when the original GeneBench came out, scored below 5%. Reviewers estimated each problem would take a human expert 20 to 40 hours. Inference cost per problem: several dollars. Open-source models fall further behind on GeneBench-Pro than on coding benchmarks, a gap OpenAI flags directly.

The pairing on the same day isn't coincidence. Anthropic's product answer to "where does the science agent live" landed at the same time as OpenAI's answer to "how do we measure whether it's any good." The vertical opened on both sides at once, with one lab building into it and the other framing the evaluation of it. A product without a benchmark is a marketing claim, and a benchmark without a product is a research paper; neither move works without the other on the same axis.

This continues the arc across the last two weeks. Ten days ago the news at both labs was workflow installs: Claude Tag into Slack, Daybreak into security stacks. Model capability was the substrate; integration was the surface. The Science product moves one step further out. Not just where the agent lives inside a general workflow, but what a productized agent looks like inside a specific domain, with a distinct name, its own tooling, and evaluation criteria native to the domain rather than borrowed from coding. The shape across two weeks is model, then install, then vertical product, with each step adding a fresh layer of packaging on the same underlying model tier.

The vertical isn't just biology. Claude Code is software engineering, Claude Cowork is knowledge work inside Slack channels, Claude Science is computational biology and drug discovery. Anthropic orders the products this way because it sees these as separable long-lived categories, not three variants of one general assistant. Whichever lab moves next will likely do so on the same axis: law, finance, medicine, materials. The next lab to elevate a science-adjacent product line is what to watch, not the next model release.

For anyone picking AI tools, the practical change is that the same underlying model shows up under different product names depending on the domain. A researcher picking up Claude Science isn't getting Claude Sonnet with a science skin; the tool set, evaluation criteria, and support surface are the product. Choosing between labs now happens per vertical: Claude Science against whatever ships from OpenAI or Google inside the same domain, judged on domain-native benchmarks like GeneBench-Pro rather than general coding scores.

Watch the slope on GeneBench-Pro, not the current score. OpenAI itself predicts the benchmark could saturate before year-end. A benchmark that saturates in six months from 28.7% isn't measuring a stable capability gap between labs; it's measuring a frontier where the leader on any given month may not be the leader the next. The vertical opened on June 30, and the leaderboard inside it starts fresh.

What to Do With This

If you're picking AI tooling as an individual or a small team working inside a domain (biology, law, finance, medicine), start looking for the vertical product from each major lab, not the flagship model. That version usually has better tooling, domain-native evaluation, and a support surface that matches how the work actually happens.

If you lead a team that uses AI across several workflows, expect a wave of vertical products from every lab in the next two quarters. Don't lock into a general-model contract that penalizes swapping across labs at the vertical level, because the best-in-vertical answer is going to keep rotating for a while.

If you decide what AI tools the whole company runs, add "which vertical products has each lab shipped, and where are we exposed" to the annual review. The general capability comparison across labs is one decision; the vertical-by-vertical comparison is a different one, and the answers usually don't line up.

Want More Than This Newsletter?

Alcreon publishes a daily AI briefing, long-form dossiers, and an analysis feed for the teams actually shipping AI in production. This newsletter is one read out of the full library.

Read the daily feed or browse the editorials.

Also on the Radar

Anthropic Ships Claude Sonnet 5, Positioned as Its Most Agentic Sonnet Yet

Anthropic released Claude Sonnet 5 on June 30, priced at $2 per million input tokens and $10 output through August 31, standard $3/$15 after. It becomes the default model on Free and Pro. Anthropic reports 50.7% on Humanity's Last Exam (with tools), up from Sonnet 4.6's 46.8%, and 84.0% on OSWorld-Verified, up from 78.5%. Sonnet 5 arrives the same day as Claude Science, and is the model tier that most of Anthropic's vertical products will run on underneath.

AI Engineer World's Fair Opens With "Coding Agents" Keynote and a "Software Factories" Track

The AI Engineer World's Fair ran from June 29 through July 2 in San Francisco, with June 30 headlined by a Coding Agents keynote and a Software Factories parallel track alongside twelve tracks per day. Language matters here: "Software Factory" is graduating from a discussion phrase into a named category for how teams of humans and coding agents get organized. Track naming at the largest AI-engineering event of the year is a working proxy for what practitioner vocabulary sticks.