For general readers, the key phrase is well-specified knowledge work. In plain English, GDPval is about whether a model can produce defined work outputs across a range of professional tasks. It is useful for judging GPT-5.5 as a work-oriented model, but it should not be treated as a single all-purpose scorecard.

The main reported numbers, side by side

Why the percentages should not be compared as if they are one leaderboard

Numbers such as 84.9%, 73.1% and 80.5% look easy to rank. That can be misleading.

• 84.9% on GDPval refers to specified knowledge work across many occupations.
• 73.1% on Expert-SWE refers to coding tasks in an internal evaluation, according to a report.
• 80.5% on BixBench refers to a real-world bioinformatics benchmark.

So the better question is not “Which percentage is highest?” It is “Which benchmark matches the job?” For general workplace knowledge tasks, GDPval is the better reference point. For software engineering, Expert-SWE is closer to the target. For bioinformatics, BixBench is more relevant to the domain.

What the Artificial Analysis result adds

Artificial Analysis says GPT-5.5 tops its Intelligence Index by three points. It also says OpenAI leads five of its headline evaluations and places second to Gemini 3.1 Pro Preview on three others.

That distinction matters. A No. 1 ranking in an external index does not mean a model wins every individual test. It means GPT-5.5 comes out ahead overall under that third-party index’s methodology.

Be careful with isolated headline scores

Some reports cite additional GPT-5.5 figures, including 91.7% in connection with legal AI capabilities and 82.7% in the context of agentic coding. Those numbers may be useful for their specific areas, but they are weaker as a general answer unless the benchmark design, comparison group and measurement target are just as clear.

That is why 84.9% on GDPval remains the cleanest general-purpose benchmark to quote: it is directly stated by OpenAI and tied to a defined evaluation area.

Which GPT-5.5 benchmark should you cite?

Use the benchmark that matches the context:

• General knowledge work: 84.9% on GDPval.
• Software engineering: 73.1% on Expert-SWE.
• Bioinformatics: 80.5% on BixBench, with a note that the sourcing is less direct in the available materials.
• Broad model ranking: No. 1 in the Artificial Analysis Intelligence Index, by three points.

Bottom line

The best short benchmark for GPT-5.5 is 84.9% on GDPval. It is official, easy to cite and clearly tied to well-specified knowledge work across 44 occupations. Other scores may matter more for coding, science or specialised professional tasks, but they should always be named with the benchmark they come from.

For general readers, the key phrase is well-specified knowledge work. In plain English, GDPval is about whether a model can produce defined work outputs across a range of professional tasks. It is useful for judging GPT-5.5 as a work-oriented model, but it should not be treated as a single all-purpose scorecard.

The main reported numbers, side by side

Why the percentages should not be compared as if they are one leaderboard

Numbers such as 84.9%, 73.1% and 80.5% look easy to rank. That can be misleading.

• 84.9% on GDPval refers to specified knowledge work across many occupations.
• 73.1% on Expert-SWE refers to coding tasks in an internal evaluation, according to a report.
• 80.5% on BixBench refers to a real-world bioinformatics benchmark.

So the better question is not “Which percentage is highest?” It is “Which benchmark matches the job?” For general workplace knowledge tasks, GDPval is the better reference point. For software engineering, Expert-SWE is closer to the target. For bioinformatics, BixBench is more relevant to the domain.

What the Artificial Analysis result adds

Artificial Analysis says GPT-5.5 tops its Intelligence Index by three points. It also says OpenAI leads five of its headline evaluations and places second to Gemini 3.1 Pro Preview on three others.

That distinction matters. A No. 1 ranking in an external index does not mean a model wins every individual test. It means GPT-5.5 comes out ahead overall under that third-party index’s methodology.

Be careful with isolated headline scores

Some reports cite additional GPT-5.5 figures, including 91.7% in connection with legal AI capabilities and 82.7% in the context of agentic coding. Those numbers may be useful for their specific areas, but they are weaker as a general answer unless the benchmark design, comparison group and measurement target are just as clear.

That is why 84.9% on GDPval remains the cleanest general-purpose benchmark to quote: it is directly stated by OpenAI and tied to a defined evaluation area.

Which GPT-5.5 benchmark should you cite?

Use the benchmark that matches the context:

• General knowledge work: 84.9% on GDPval.
• Software engineering: 73.1% on Expert-SWE.
• Bioinformatics: 80.5% on BixBench, with a note that the sourcing is less direct in the available materials.
• Broad model ranking: No. 1 in the Artificial Analysis Intelligence Index, by three points.

Bottom line

The best short benchmark for GPT-5.5 is 84.9% on GDPval. It is official, easy to cite and clearly tied to well-specified knowledge work across 44 occupations. Other scores may matter more for coding, science or specialised professional tasks, but they should always be named with the benchmark they come from.