OpenAI introduces GeneBench-Pro with GPT-5.6 Sol Pro at 31.5%

• OpenAI says its GeneBench-Pro announcement is meant to test agent performance on messy computational biology work, not biology trivia or fixed pipelines.
• According to Wes Roth's summary and reach_vb's breakdown, the benchmark scores whether a model can inspect noisy data, catch problems, revise assumptions, and defend a conclusion.
• deredleritt3r's score chart puts GPT-5.6 Sol Pro at 31.5%, up from 20.5% for GPT-5.5 Pro, a gain of 11 points on OpenAI's hardest reported setting.
• The attached benchmark chart in gdb's post shows GPT-5.6 Sol Pro ahead of Opus 4.8 at 16.0%, Gemini 3.5 Flash at 8.1%, and DeepSeek V4 Pro at 2.4%.
• Even the top result is far from autonomy: Wes Roth's summary says experts estimated a typical GeneBench-Pro problem would take a human 20 to 40 hours.

You can jump to OpenAI's announcement, inspect the

, and the

is the most interesting weird bit, because the strongest line is not the one that spends the most tokens.

GeneBench-Pro is framed around judgment calls that show up in real computational research. OpenAI's launch post describes agents navigating messy biological data, while reach_vb's summary spells that out as inspecting data, catching bad samples, choosing an analysis, revising assumptions, and producing a defensible conclusion.

That makes it a workflow benchmark more than a recall benchmark. Wes Roth's summary says the goal is to see whether agents can handle noisy, judgment-heavy work instead of fixed biology QA.

The headline number is 31.5% for GPT-5.6 Sol Pro. deredleritt3r's score post also lists the earlier Pro results in sequence:

• GPT-5.2 Pro: 8.5%
• GPT-5.4 Pro: 16.3%
• GPT-5.5 Pro: 20.5%
• GPT-5.6 Sol Pro: 31.5%

The same chart in gdb's post places OpenAI's top model well ahead of the non-OpenAI models shown there, with Opus 4.8 at 16.0% and Gemini 3.5 Flash at 8.1%.

The scaling plot adds a more useful detail than the headline score. The OCR from

shows GPT-5.6 Sol reaching the highest pass rate at roughly 35,000 tokens used, while GPT-5.6 Luna stretches much farther right, to about 120,000 tokens, and still tops out much lower.

That suggests extra test-time compute is not carrying weaker variants to the same ceiling. In this benchmark, model choice still matters more than simply letting an agent think longer.

OpenAI's own framing leaves the brag muted. Wes Roth's summary and gdb's post both say experts estimated a typical GeneBench-Pro problem would take a human researcher 20 to 40 hours.

A 31.5% pass rate is still notable on tasks with that kind of labor behind them. It also means the benchmark is measuring a part of agent progress where failure still dominates the distribution.