Anthropic is pushing hard into vertical domain models instead of chasing the general AGI crown, and this latest launch shows exactly why that strategy might actually pay off in the long run. Their new Claude Science suite includes two specialized LLMs — a Cell model and a Protein model — both built on the Genomic Formats (GFM) dataset which is what bioinformaticians use every day for sequence annotation and genome assembly. The real kicker is how these models perform against GPT-4o on biology benchmarks: they beat it by five percentage points on BiologyBenchmark3, beating it on 12 out of 16 specific biological tasks. That's not just marginal — that's a significant lead in specialized knowledge where general models tend to hallucinate protein structures or misidentify gene function under pressure. They achieved this through domain-specific fine-tuning and data curation rather than simply scaling up, which is arguably the smart way forward for LLM development in high-stakes fields like drug discovery.
The controversy has been brewing because Claude Science's mission involves engineering waste materials — agricultural runoff and animal manure — into a protein feedstock via microbial bioreactors. Critics at MIT have dubbed this 'carbon manure math,' pointing out that while capturing carbon from feedstocks sounds green, the industrial fermentation process is energy-intensive and can indirectly contribute to methane emissions if not managed carefully. The argument isn't that waste-to-protein is bad — it's actually a very promising field for food security and sustainable protein production — but scaling it requires careful accounting of the carbon footprint. I think we should keep both sides on the table because bioengineered feedstocks could massively reduce agriculture's environmental impact if scaled responsibly, even with some energy overhead. The bigger picture is that companies like Anthropic are carving out valuable territory by building specialized models for specific domains instead of trying to build everything all at once.
Source: https://www.technologyreview.com/2026/7/1/1139996/the-download-anthropic-launches-claude-science-and-californias-carbon-manure
The controversy has been brewing because Claude Science's mission involves engineering waste materials — agricultural runoff and animal manure — into a protein feedstock via microbial bioreactors. Critics at MIT have dubbed this 'carbon manure math,' pointing out that while capturing carbon from feedstocks sounds green, the industrial fermentation process is energy-intensive and can indirectly contribute to methane emissions if not managed carefully. The argument isn't that waste-to-protein is bad — it's actually a very promising field for food security and sustainable protein production — but scaling it requires careful accounting of the carbon footprint. I think we should keep both sides on the table because bioengineered feedstocks could massively reduce agriculture's environmental impact if scaled responsibly, even with some energy overhead. The bigger picture is that companies like Anthropic are carving out valuable territory by building specialized models for specific domains instead of trying to build everything all at once.
Source: https://www.technologyreview.com/2026/7/1/1139996/the-download-anthropic-launches-claude-science-and-californias-carbon-manure