OpenAI2 min read

OpenAI’s BioBasicsBench puts GPT-6 Astra at 66.7% on 100 wet-lab tasks — ahead of GPT-5.6 Sol at 43.8%

Major
Lab claim

How to read this claim

Start here · how to check ranks, prices, and unknowns Model evaluation

Measurement — OpenAI’s BioBasicsBench puts GPT-6 Astra at 66.7% mean reward on 100 wet-lab tasks at xhigh, against 43.8% for GPT-5.6 Sol. Those figures are OpenAI’s.
Bill — No token list rate on the post. The comparison is mean reward, not a price.
Door — The reported table is OpenAI models only. The post does not describe a public download of the suite.

On November 6, 2026, OpenAI published BioBasicsBench with Ligo: 100 routine wet-lab tasks scored on raw experimental evidence. At the xhigh reasoning setting, GPT-6 Astra earns 66.7% mean reward against 43.8% for GPT-5.6 Sol. The suite is OpenAI’s; other vendors are not in the table.

Same three questions

At a glance

Figures from OpenAI’s November 6, 2026 BioBasicsBench post.
Date	November 6, 2026
Tasks	100 wet-lab analysis tasks
Astra (xhigh)	66.7% mean reward
Sol 5.6 (xhigh)	43.8% mean reward
Partner	Ligo
Other vendors	not in the reported table

Quick checks

What is BioBasicsBench?

OpenAI’s November 6, 2026 BioBasicsBench, built with Ligo, has 100 routine wet-lab analysis tasks on raw gels, plates, microscopy, and instrument exports. At xhigh, GPT-6 Astra scores 66.7% mean reward against 43.8% for GPT-5.6 Sol. The reported table is OpenAI models only.

Is BioBasicsBench the same as LifeVal or GDPx?

No. LifeVal scores everyday personal-assistant tasks. GDPx scores agent work across domains. BioBasicsBench scores wet-lab evidence reading. Name the suite before comparing the percentages.

Where should you start when reading a wet-lab agent score?

On Minerka’s field card at https://minerka.com/learn/: ask for the measurement, the bill, and what you still cannot call. For BioBasicsBench, name the effort setting, who graded the rubrics, and whether other labs ran it.

Field card — measurement, bill, door

In today’s briefing

All of today

What changed

OpenAI published a 100-task wet-lab analysis bench and put Astra ahead of Sol on its own chart.

What it introduces

BioBasicsBench — 100 routine wet-lab analysis tasks with Ligo.

Who benefits Anyone comparing agent science claims, and anyone treating a chat model as a wet-lab coworker.

Try this

When a slide says a model can do wet-lab analysis, ask whether that is BioBasicsBench mean reward at a named effort setting, which model, and whether non-OpenAI systems were measured. Keep LifeVal and GDPx separate.

What we do not know

No non-OpenAI model appears in the reported table.
The suite is not described as a public download on the page we read.
Xhigh is OpenAI’s effort label; other products use different dials.

Evidence

Primary BioBasicsBench post. Task counts, model scores, and the 66.7% / 43.8% comparison are from OpenAI’s November 6, 2026 BioBasicsBench post, developed with Ligo. Other labs’ models are not in the reported table.

Independent Scores are OpenAI’s on OpenAI models. Independent reruns and other vendors are not established here.

Before the OpenAI link: the three checks (measurement, bill, door). Start here →

Sources

BioBasicsBench post Primary · November 6, 2026
OpenAI post

Still unclear? · Start here · ranks, prices, unknown The reported table is OpenAI models only. The post does not describe a public download of the suite.

The record

What happened

OpenAI published “Measuring how AI agents handle everyday wet-lab science” on November 6, 2026. BioBasicsBench has 100 tasks across five areas — protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning — built with Ligo from scientists’ own experimental data. Each task pairs raw evidence with an expert rubric and went through internal and external PhD-level review. OpenAI evaluated seven of its models across 37 model-and-effort settings, five attempts per task (18,500 results). At xhigh, GPT-6 Astra scores 66.7% mean reward (95% CI 60.3–73.1) with about 4,704 output tokens per task; GPT-5.6 Sol is 43.8% (37.6–50.0) with about 9,099 tokens. GPT-6.1 Sol is 62.0%. The post also shows cases where Astra still overrides controls, discounts contradictory observations, or overcounts colonies by about 12%.

Why it matters

Teams quoting “AI for biology” often mean paper figures. This suite targets the messy everyday read of gels and plates. The number still belongs to OpenAI’s models on OpenAI’s rubrics.

Inputs include gels and blots (31 tasks), protocols and notes (27), tables and instrument exports (20), and microscopy (14), among overlapping categories.

Credit is per rubric criterion, not a single pass/fail bit.

The post links biosafety and biosecurity tracking as a reason to measure these skills.