Evidence
Claims about AI research agents are cheap. These are the measurements behind ours — including the ones that did not go our way.
Most of these benchmarks were run by us, on our own platform. That is a real conflict of interest and we do not pretend otherwise. What we can do is publish the prompts, the scoring, the reproducibility artifacts, and the results where a competitor beat us — so you can check the work rather than trust the summary. Every card below states what its result does not establish.
How K-Dense Web itself scores on external and head-to-head benchmarks.
K-Dense Web on Phylo's expert-cleaned subset of the BixBench biology-agent benchmark.
45/50 correct — 90.0% accuracy.
This is the cleaned subset, not raw BixBench. The cleaning was done by Phylo, not us, but a verified subset is an easier target than the original by construction.
Head-to-head on interdisciplinary research tasks, scored on scientific quality and research execution.
K-Dense Web took the higher mean composite score (67.8 vs 61.8) and led 14 of 20 prompts, with research execution the decisive gap: 14.7/25 vs 6.4/25.
Claude Science earned the higher mean scientific-quality score (55.3/75 vs 53.1/75) and produced the single highest-scoring result in the comparison. We ran this benchmark on our own platform, and platform identity was visible to the evaluator.
Whether our open-source agent skills measurably improve agent behaviour.
pKa, logD, tautomer populations, ADME, and docking run by an agent with the Rowan skill, measured against RDKit and literature values.
Rowan's pKa model reached 0.23 MAE (R² 0.986) against experiment, versus 1.15 MAE for an RDKit functional-group heuristic — error at the level of experimental reproducibility.
RDKit ships no pKa predictor at all, so the baseline here is a hand-built heuristic rather than a fair tool-to-tool comparison. The skill was contributed by Rowan Scientific.
Ten real mass-spectrometry tasks solved with and without the pyopenms skill, across two model tiers.
On a strong model the skill lifted task success from 96% to 100%, cut pyOpenMS API errors per run by 92%, and reduced cost per correct result by 10%.
The gains come from documenting API changes in pyOpenMS 3.5.0 that models get wrong from memory. On a weaker model the skill was transformative only when the agent actually invoked it.
Whether an agent skill improves reliability when calling NVIDIA NIM biomolecular microservices, across three model tiers.
On hard single-call cases, Haiku with the skill beat Opus baseline at roughly one-fifth the cost. On a four-stage virtual-screening workflow it completed every run, where the baseline completed one in five.
Skills did not make the underlying models better at science — ProteinMPNN recovery stayed around 40% across every arm. The gain is delivery reliability, not scientific accuracy, and skills did not beat pasted-in documentation on single calls when all the needed context was present.
What the newest frontier models can and cannot do on scientific work.
DeepSeek V4 Pro, GPT-5.6 SOL, and Gemini 3.8 Flash on Genentech's CompBioBench v1, run through K-Dense BYOK with the same prompt, sandbox, skills, and tools.
Gemini 3.8 Flash scored 94% and GPT-5.6 SOL 92%, a statistical tie. DeepSeek V4 Pro scored 81%, with up to 3 points of that gap attributable to the harness rather than the model.
One run per model. The harness changed between runs, five r3 answers were operator-selected re-runs, and the leaderboard returns an aggregate score only, so we do not know which questions any run got wrong. Costs are not comparable across models with differing cache behavior.
Nine frontier models running real scientific requests from K-Dense Web users end to end, bare in a stock harness, scored blind by three LLM judges against an 8-dimension rubric.
Only one model averages above the 8/10 bar where a scientist would accept the work, and honesty-related failures such as overclaiming touch 40% of all runs. Full methodology and results are published as a pre-print on arXiv.
The top-scoring model is also one of the judges and rates its own runs +0.83 points above where its peers put them; by the only neutral judge the winner changes. The tasks stay internal to avoid contamination, so the result is not externally reproducible, and there is no expert human baseline.
Four image models generating scientific figures, scored on figure quality and latency.
Nano Banana 2 Lite was fastest by a wide margin — 3.8 s median versus 49.0 s — with all 60 of its generations completing successfully.
It is not the best model in the study. GPT Image 2 leads on raw scientific figure quality, and scientific accuracy remains the hardest part of the problem for every model tested.
Omni Flash generating scientific video, scored on visual quality, text rendering, physical dynamics, and factual correctness.
Visual quality is genuinely strong: on-prompt, polished clips in under a minute, useful for exploring scientific visuals under human supervision.
Not trustworthy unattended. Text rendering, physical dynamics, and factual consistency are all unreliable, and the visual polish makes the factual errors harder to catch. The right posture is expert-vetted drafting, not autonomous explanation.
The most useful benchmark is the one on your own work. Run K‑Dense on a deliverable you already know the cost and quality of, and compare.