This summer, Anthropic has shown us a future most of us don’t want: its strongest scientific model locked away, while the one we could use refused ordinary biology and handed the work to something weaker. Opus 5, launched a few days ago, is a much better fallback. But scientists still depend on Anthropic deciding which models they can access, what those models will answer and where the work can run.

A new open letter from some of AI’s biggest companies argues for the opposite future: models whose weights can be downloaded, adapted and run where the data lives.

But as Bo Wang and Daphne Koller argue, biology is a different beast. Even if every model were flung wide open, progress would still depend on experimental data that is expensive to generate and often kept private.

Let’s dive in.

Much like Ross and Rachel in Friends, we seem to be in a perpetual on-again, off-again relationship with Claude. Mythos remains inaccessible for most people, while Fable spent the summer refusing perfectly ordinary biology questions and handing them to the weaker Opus 4.8. All that, and Anthropic still had the nerve to ask us to be excited for its new Claude Science platform.

Well, the fallback is still in place. But Fable won’t redirect you to Opus 4.8 any longer - it will take you to the new shiny Opus 5. It’s on-again.

Opus 5 is now Anthropic’s most capable generally available model for scientific research, according to the company. Mythos 5 remains stronger on long-running biological work that requires repeated analysis and improvement.

Anthropic says Opus 5 improved over Opus 4.8 on every one of its life-science evaluations. In its launch materials, the company highlighted gains in organic chemistry - including inferring molecular structures from spectroscopy - and predicting how mutations change protein function.

LatchBio, a computational-biology infrastructure company, also ran its own testing. Its benchmarks ask AI agents to work through datasets, write and execute code, use scientific tools and arrive at verifiable answers. Across nine benchmarks, Opus 5 produced LatchBio’s best results yet in four areas. LatchBio also says it beat Opus 4.8 on six of seven benchmarks.

It did not sweep the board, however. OpenAI’s GPT-5.6 Sol remained ahead on short-horizon single-cell analysis and long-horizon spatial biology, while GPT-5.5 led the epigenomics benchmark - the one area where Opus 5 regressed against Opus 4.8.

But while the scores show how often Opus 5 recovered the benchmark’s target result, LatchBio also inspected how it got there. Those records show a model working differently from Opus 4.8. Opus 5 writes less visible reasoning, but moves more of its working memory into files. It writes scripts, saves intermediate results, reloads them and breaks long analyses into numbered stages.

In one single-cell analysis, Opus 5 found what looked like an unusual plasma-cell signal. It then ran two additional checks and concluded that the result was an artifact caused by two cells accidentally being captured and read as one. “Claude Opus 5 behaves more like a careful scientist than any model we’ve run,” LatchBio CEO Alfredo Andere says in Anthropic’s launch post.

He added: “It reaches for the right statistical tests to rule out confounders, cross-checks its own results by independent methods, and stays on track through long multi-step analyses.”

That reliance on background jobs also creates a failure mode. Opus 5 launches jobs and then waits for them; if one crashes, times out or never reports back, the model can get stuck in a loop. On LatchBio’s long single-cell benchmark, the median Opus 5 run took 38 minutes versus 17 for OpenAI’s GPT-5.6 Sol, and 9% ran beyond three hours.

Jennifer Doudna recently questioned whether chatbots can produce genuinely new scientific ideas. Opus 5 does not settle that question. But it does show a model taking on more of the scientific process itself - choosing tests, checking confounders, cross-validating results and carrying a long analysis through. During the launch of Claude Science, Anthropic promised it could “run the work.” That promise now looks a tiny bit more credible.

The AI industry’s new open-weights manifesto does not mention biology. It probably should have.

A coalition including Microsoft, Meta, NVIDIA, OpenAI, Google and Hugging Face has published a joint policy letter arguing that America’s AI lead depends partly on models whose weights can be downloaded, modified and run on a user’s own infrastructure.

The weights are the learned parameters inside a model. Releasing them lets researchers run and adapt the model without sending every query through the developer’s API. The letter asks US policymakers to widen access to compute, invest in shared training resources and avoid premature restrictions on open models.

Two prominent figures in AI x biology have now supplied the arguments the letter leaves unstated.

☑️ Bo Wang, a University of Toronto professor and Xaira’s chief AI scientist, argues that pharma companies will not send proprietary compound data, clinical results or patient cohorts to a third-party API. “The model has to come to the data,” he wrote on X.

Wang also points out that biology has no internet-scale training corpus. Biological data has to be generated experiment by experiment, at real cost. Wang argues that every reduction in model cost, and every improvement in the ability to customize and run a model locally, expands the scientific questions a research team can afford to ask.

☑️ Daphne Koller, founder and CEO of insitro, points to AlphaFold2 as the experiment the letter is asking policymakers to repeat. DeepMind released its code and trained model in 2021, allowing researchers across academia and biotech to use and build on protein-structure prediction.

“For those of us in the field, the letter’s argument isn’t a forecast,” Koller wrote on LinkedIn. “We ran that experiment, and open science won.”

But AlphaFold2 also depended on decades of structural biology made freely available through the Protein Data Bank. Echoing Wang, Koller reminds us that “biology has no internet.” Her own company has generated more than 20 petabytes of biological data and still considers that only a tiny fraction of what AI will need to understand human biology.

Wang and Koller shift the open-model debate from software policy to who gets to participate in AI-enabled biology. Open weights can move capability out of frontier labs, but if the biological data remains private, the real advantage stays concentrated with whoever owns the experiments. AlphaFold2 became field-wide infrastructure because an open model was built on an open data foundation. Many other areas of biology are in dire need of the same combination.

A base editor can be guided to the right stretch of DNA and still act on a near-match elsewhere. Researchers have now used AlphaFold3 to find two amino-acid changes - one in Cas9 and one in its letter-changing enzyme - that sharply reduced those mistakes. The findings are published in Nature.

A base editor works in two stages. First, a guide RNA leads Cas9 to a chosen DNA sequence, where Cas9 opens the double helix. Then an attached enzyme changes one DNA letter. Errors can arise at either stage: Cas9 can accept a similar DNA sequence elsewhere in the genome, while the attached enzyme can also act on unintended DNA or RNA.

The researchers began with ABE8e, an existing adenine base editor that changes A to G. They mapped the DNA sites where it made unintended edits in human cells. AlphaFold3 then built 3D models of Cas9 and the guide RNA bound to DNA at both the intended sites and the sites edited by mistake.

The overall predicted structures looked almost identical at the intended and unintended sites, giving the researchers little clue about which local interactions were associated with unwanted editing.

But AlphaFold3 also estimates which individual amino acids in Cas9 are likely to contact - lie close enough to interact with - the DNA or guide RNA. Those predicted points of interaction differed more clearly between intended and unintended editing sites.

The researchers then built a framework, ContactSeek, to analyze that signal. It grouped neighboring Cas9 amino acids whose predicted proximity to the DNA or guide RNA changed in similar ways across off-target sites, then compared those patterns with the amount of unwanted editing measured in cells. This produced a ranked list of regions and amino acids that might influence whether Cas9 edits a near-matching DNA sequence.

ContactSeek first pointed to one amino acid in Cas9. Changing it reduced guide-dependent off-target signal by 69-83% in two genome-wide tests. The team then applied the same approach separately to the letter-changing enzyme and found another change that reduced unintended editing of both DNA and RNA.

Combining the two amino-acid changes created a new version of ABE8e, called ABE8e-DD. In genome-wide tests using two different guide RNAs, it reduced total guide-dependent off-target signal by 98.1% and 99.2% compared with the original ABE8e. The paper reports that it also outperformed several existing high-fidelity adenine base editors.

The researchers then transferred the approach to a different system: a Cas12a-based cytosine editor that changes C to T. ContactSeek again identified mutations that improved specificity.

That suggests the method can work across different editing machines, but it is not a push-button safety fix. It needs experimental data showing where an editor already fails, and the redesigned editors were not tested in animals.

In Issue 41, BAIO covered Jennifer Doudna’s team using AI to rewrite large parts of TnpB while preserving its ability to cut DNA. This study uses AI almost the opposite way: keep the existing editor, use its mistakes as evidence, and identify the few parts most worth changing.

ContactSeek offers a new and direct route to improving an editor’s precision: use the editor’s own mistakes to identify the amino acids most likely responsible, then change only those. If it generalizes, making gene editors safer could become far less trial-and-error.

DOME Copilot is an open-source tool that reads life-science AI papers and drafts a structured account of how their data, optimization, models and evaluation were handled, following the DOME reporting guidelines. It is not brand new, but its preprint received a substantial update on July 24. The GitHub version requires local installation and must be connected to a hosted language model.

☑️ AI Chatbots Know How to Make Deadly Biological Weapons. Some Will Teach You.

☑️ PertReason asks whether virtual cells can explain their predictions.

☑️ MSA Pairformer learns protein interactions without training on complexes.

☑️ Bristol Myers buys Nvidia's latest AI computing system for drug research.

☑️ AI shortens drug discovery to around 1 year in China, Insilico CEO says.