THE TWO RESULTS
● Anthropic: Claude models designed protein binders hitting 14 of 15 targets, 22 to 35 percent success against a 10 to 15 percent industry baseline. Validated by Adaptyv Bio and Twist Bioscience.
● Reconstruction benchmark: frontier models recover a paper's core ideas from its bibliography alone at 3 to 15 percent. A four-model tournament pipeline reaches 42 percent.
● Both are correct. They measure execution and origination, and those are not the same skill.
What Anthropic actually showed
Anthropic published results from work verified by two external labs rather than scored on a benchmark. Claude models — Mythos Preview and the Opus line — designed protein binders that succeeded against 14 of 15 tested targets, at a 22 to 35 percent success rate where the typical industry rate is 10 to 15 percent.
A second result: Opus 5 processed raw NMR and LC-MS instrument data in 23 and 19 minutes, returning purity figures within 0.1 percent of the lab's own reading.
WHY WET-LAB VALIDATION CHANGES THE WEIGHT
A benchmark score can be contaminated by training data. A protein either binds or it does not, and a third-party lab either measures it or it does not. That is a different class of evidence, and it is the reason this result deserves more attention than most capability announcements.
Worth noting alongside it: Anthropic says life-science tasks remain restricted in its most capable models. Publishing a strong result while keeping the underlying capability gated is an unusual posture, and it follows the company raising its own catastrophic misalignment rating from very low to low last week.
What the Reconstruction benchmark showed
Published this month, Reconstruction tests something benchmarks usually cannot: whether a model can originate a research idea rather than recall one.
It hands the model a paper's bibliography and nothing else — no full text, no author information, no post-publication signal. Everything a model might have memorised is stripped out. The task is to recover the paper's core ideas from the same starting position the original researcher occupied.
| Setup |
Score |
| Frontier model, solo |
3 to 15 percent |
| Multi-agent, four-model Swiss tournament |
42 percent |
The authors' reading is that once retrieval is removed as a crutch, current models hit a genuine ceiling on hypothesis generation. The multi-agent jump from 15 to 42 percent is real and large — but it took four models in a tournament to get there, which is an expensive way to reach a number a researcher would not consider impressive.
The distinction that reconciles them
These results are not in tension once you name what each measures.
Anthropic measured execution. The target was specified. The success criterion was defined. The task was to generate candidates against a known objective and have a lab check them. Models are extremely good at this, and getting better fast.
Reconstruction measured origination. Nobody specified the target. The task was to work out what question was worth asking. Models are currently poor at this, and adding compute helps less than you would hope.
That maps almost exactly onto what the coding-agent data showed this week: agents tripled pull request throughput in Linear's telemetry while total development time rose, because the bottleneck was never producing candidate work. It was deciding what work was worth doing and checking whether the output was right.
THE SAME SHAPE, THREE DOMAINS
Software: agents write more code, humans still set direction and review.
Science: models generate better candidates, humans still choose the target.
The capability curve is steep on execution and flat on judgement.
What this means if you use these tools
| If you are... |
The useful read |
| In a research role |
Hand models well-specified generation tasks. Do not expect them to pick the question |
| Evaluating AI research claims |
Ask whether the target was specified in advance. It usually was |
| Considering multi-agent setups |
15 to 42 percent is a real gain. Price four models running a tournament before assuming it is worth it |
| Reading capability headlines |
Wet-lab or third-party verification outranks any benchmark number |
| Worried about AI replacing researchers |
The part being automated is candidate generation. The part that is not is knowing what to look for |
The honest caveats
Anthropic's results come from Anthropic, validated by partner labs it engaged. That is considerably better than self-reported benchmarks and still not independent replication by a disinterested group.
Reconstruction is one benchmark, newly published, and new benchmarks sometimes measure their own design more than the thing they name. The 3 to 15 percent figure will firm up or move as others run it.
And both describe August 2026. The execution curve has moved fast enough this year that a twelve-month-old ceiling claim would already look silly.
FAQ
Did AI design working proteins?
Claude models designed binders that succeeded against 14 of 15 targets at 22 to 35 percent success, verified by Adaptyv Bio and Twist Bioscience, against a typical industry rate of 10 to 15 percent. Real, narrow, and externally checked.
Why do models score only 3 to 15 percent on Reconstruction?
Because the benchmark removes full paper text, author information and post-publication signals, so retrieval cannot help. What remains is generating the hypothesis, which current models do poorly.
Does the 42 percent multi-agent score change the conclusion?
It softens it. Four models in a Swiss tournament nearly tripled the solo score, which suggests the ceiling is partly architectural rather than fundamental. It is also an expensive setup for a result a researcher would not call strong.
Are these two findings contradictory?
No. One tests execution against a specified target, the other tests originating the target. Models are far stronger at the first.
Can I use Claude for life-science work?
Anthropic states that life-science tasks remain restricted in its most capable models. The published results came from internal and partner-lab work rather than general availability.
What should I take from this as a general user?
Specify the target yourself and let the model generate against it. That is where the capability actually is, and it is the same lesson the coding-agent data delivered this week from a completely different direction.