An AI Research Agent Scores 18.6% at the One Job Research Agents Exist to Do. It's Also the Leaderboard's Best Model. Here's Everything Else That Shipped.

18.59%. That's how often the top-ranked model on a new AI-for-science benchmark correctly combines evidence from multiple sources into one answer, the single task a research agent exists to perform. That same system holds the highest overall score of any model tested: 49.39%. We go deep on the gap between finding information and combining it, and why human experts hit a version of the same wall, in today's feature. Here's everything else that landed this week.

A synthesis cliff, quantified

SciExplore(arXiv 2607.20926) tests AI research agents on 103 tasks across four increasingly difficult stages: database navigation, ambiguous literature retrieval, missing-reference completion, and cross-source synthesis. OpenAI Deep Research posts the top overall average, 49.39%, but scores just 18.59% at row-level accuracy on the synthesis stage alone — the step that requires combining facts from multiple sources into one coherent, structured answer. Human experts show a related pattern on the benchmark's navigation task, losing 41.7 points of accuracy as reasoning hops rise from 2 to 6. Full breakdown, including the per-task table, in today's feature.

A 4B open model beats a 35B model from its own family at deep research

AREX(arXiv 2607.21461, Lu et al., 24 authors, submitted July 23) is a recursively self-improving deep-research agent that alternates an inner loop gathering evidence with an outer loop that audits its own provisional answer against the task's constraints. The dense 4B variant, AREX-Turbo, scores 70.7 on BrowseComp against Qwen3.5-35B's 61.0 — a 9.7-point win for the model with roughly a ninth of the parameters — and a similar 11.4-point margin on WideSearch-en. The larger 122B-A10B mixture-of-experts variant, AREX-Base, beats DeepSeek-V4-Pro on WideSearch-en, DeepSearchQA, and HLE, though DeepSeek-V4-Pro still edges it on BrowseComp, 83.4 to 82.5. A capable 4B research agent is small enough to be a realistic candidate for the kind of local, on-device deep-research workflow this blog keeps tracking.

The US government just backed AI-for-science with $293.76M now, $5B pledged overall

The Department of Energy's Genesis Mission selected 278 projects on July 22 — what DOE calls the largest response to any of its funding calls ever — spanning all 50 states and 342 institutions: 16 DOE/NNSA national labs, 142 universities, 157 companies, and 13 nonprofits. The cohort is backed by a $293.76 million DOE grants program, inside a broader $5 billion-plus federal commitment across 15 agencies plus over $500 million in private consortium funding. The largest single award, $60 million over three years, targets AI-accelerated nuclear facility delivery. DOE describes the program as giving funded teams access to shared "AI agent frameworks" through a Genesis Mission Platform — a concrete signal that agentic research tooling, not just bigger models, is now an explicit line item in US federal science funding.

Claude Opus 5 posts real gains on life-sciences evals

Anthropic's Claude Opus 5 launched July 24 at the same $5/$25 per-million-token pricing as Opus 4.8. Anthropic reports Opus 5 beats Opus 4.8 on every one of its internal life-sciences evaluations, with a 10.2 percentage-point gain on an organic-chemistry task (inferring molecular structure from spectroscopy data) and a 7.7 percentage-point gain on protein variant-effect prediction. These are vendor-reported, internal-benchmark numbers rather than independently audited ones, worth tracking rather than taking as a final word.

What this means for reproducible, local-first science

Three different stories this week, one shared thread: the field is getting better at the parts that are easy to measure — overall leaderboard averages, federal funding totals, per-token pricing — while the part that's hard to measure, whether an agent's output can actually be trusted once multiple sources have to agree with each other, is the one still sitting at 18.59%.