TasteVal: Opus 5.5 beats expert baseline at AI research experiment design, with a 2.30x compute multiplier

P-Zero Research's new TasteVal benchmark reports that the experimental research taste of frontier AI models has doubled every 3.0 months since December 2025 (95% CI: 1.7-5.0 months). The top-scoring model, Opus 5.5, exceeds the human expert baseline with a compute multiplier of 2.30x (95% CI: 1.15-4.37x), at roughly 1/30 of the experts' average cost per attempt, according to the benchmark's authors.

What TasteVal measures

TasteVal tests 8 tasks representative of frontier AI development. They span pretraining data curation, pretraining and fine-tuning language models, preference modeling, and robustness to adversarial prompts. P-Zero Research defines research taste as the ability to pick interesting problems, design experiments, and interpret results. The benchmark covers only the experimental component. Given a fixed research problem, the model iteratively designs experiments and draws conclusions from their outcomes.

The human baseline comes from 24 experts, with at least two per task. All have recently worked at organizations including OpenAI, Google DeepMind, NVIDIA and Microsoft Research.

Results

• Growth rate: Research taste doubled every 3.0 months since December 2025. The authors say growth was much slower in prior years, with the slope changing around the end of 2025 and start of 2026.
• Top model: Opus 5.5 scores a 2.30x compute multiplier against the human baseline. The lower bound of the 95% confidence interval is 1.15x.
• Cost: About 1/30 of the expert baseliners' average cost per attempt.

Context window, pricing and parameter count for Opus 5.5 were not reported in the source material.

Second benchmark: Humanity's Sixth Sense

Elorian and Scale AI released Humanity's Sixth Sense (HSS), a benchmark of how models reason about what a visual scene implies but does not show. Examples include whether a bookcase can fit more books, or whether chains bear a plough's drag weight in a photo of horses pulling it.

HSS contains 522 questions covering 288 images and 234 short video clips, totaling 17.6 hours of video. Questions span four categories: temporal and causal dynamics, physical and spatial logic, social understanding, and abstract and contextual inference.

Subject	Aggregate pass@1	Image	Video	Time	Space
Humans	93.1	94.1	91.9	97.0	94.9
GPT 6 Astra (best model)	53.6	55.6	51.1	58.2	53.6

Chinese open-weight models scored far lower. GLM-5.3-flash reached 26.8 pass@1 and Kimi K3 reached 25.5. The authors say the two most common errors were missing a detail that decides the answer (21% of failures) and misidentifying an object, person, or role (20%).

What this means

TasteVal's headline numbers carry wide uncertainty. The 2.30x multiplier has a confidence interval whose lower bound is only 1.15x, and the doubling time ranges from 1.7 to 5.0 months. Both rest on 8 tasks and a 24-person baseline. The result is a single-group benchmark and has not been independently replicated. Still, the reported inflection around the turn of 2025-2026 is the data point to watch. If it holds, the capability most often cited as a bottleneck for recursive self-improvement is improving at an accelerating rate.

Jack Clark of Import AI, who flagged both benchmarks, now considers full recursive self-improvement more likely in 2027 than 2028. That is his opinion, not a measured result.

HSS points the other way. The best model sits roughly 40 points below humans, so intuitive visual reasoning remains an open weakness. Clark argues that the best LLMs scoring as well as they do undercuts the claim that visual reasoning requires dedicated 2D and 3D world models. The gap between GPT 6 Astra (53.6) and the Chinese open-weight models (about 25-27) is also notable. On visual inference, those models trail the frontier by roughly half.