GPT-5.6 Benchmarks: Vendor Claims vs Verified Numbers

OpenAI says GPT-5.6 sets state-of-the-art results on coding, knowledge work, cybersecurity, and science. Those are claims from OpenAI, not independent facts. Treat them as a starting point, not a verdict.

The one benchmark result OpenAI leans on hardest is Agents' Last Exam. On that test, OpenAI reports that GPT-5.6 Luna outperforms Claude Fable 5 at an estimated cost per task nearly 99% lower. That is a striking claim about price and performance together.

This page shows you what OpenAI actually claims, what is independently verified so far, which benchmarks matter for which job, and how to check the current numbers yourself from primary sources.

What OpenAI actually claims about GPT-5.6

OpenAI claims GPT-5.6 is state-of-the-art on coding, knowledge work, cybersecurity, and science. These are qualitative marketing claims tied to the family launch, not a single verified score.

The GPT-5.6 family has three tiers. Sol is the flagship for the hardest work. Terra is the balanced everyday tier. Luna is the fastest and cheapest tier.

OpenAI positions Sol as its top reasoning model, adding new max and ultra reasoning modes, where ultra can spawn subagents. The strongest benchmark story sits with Sol.

• Sol: complex coding, cybersecurity, science, deep reasoning.
• Terra: high-volume business work like support and document analysis.
• Luna: summarization, drafting, classification, and routine automation.

The Agents' Last Exam result OpenAI highlights

Agents' Last Exam is the benchmark OpenAI cites most for GPT-5.6, and it is the strongest sourced claim in the launch. OpenAI reports that Luna outperforms Claude Fable 5 on this test at an estimated cost per task nearly 99% lower.

This benchmark measures long, multi-step agentic tasks, not a single question. It rewards models that can plan and work across many steps, which is why OpenAI features it.

Two cautions matter here. The cost figure is an estimate, and the comparison is drawn by the vendor, not by a neutral third party. It is a claim worth verifying, not a settled fact.

Which benchmarks matter for which workload

The right benchmark depends on the job you are hiring the model for. A high coding score tells you little about summarization quality. Match the test to your real task.

For software work, SWE-bench measures whether a model can resolve real GitHub issues end to end. That is closer to production coding than a quiz score.

For general knowledge, MMLU and GPQA test broad reasoning and hard science questions. For head-to-head preference, LMArena ranks models by blind human votes.

• Coding: SWE-bench for real issue resolution, not toy problems.
• Knowledge work: MMLU for broad coverage, GPQA for hard graduate-level science.
• Human preference: LMArena for blind side-by-side voting.
• Agentic workflows: Agents' Last Exam for long, multi-step tasks.

Vendor-claimed vs independently verified

As of this writing, GPT-5.6's headline results are vendor-claimed, and independent verification is still catching up. That gap is normal in the days after any launch, and it matters for your decisions.

OpenAI has published qualitative state-of-the-art claims and the Agents' Last Exam comparison. Where OpenAI has not published an exact figure for a specific benchmark, we do not print one here.

We deliberately avoid quoting SWE-bench, MMLU, or GPQA scores we cannot tie to a primary source. If you see a precise GPT-5.6 score online, check who ran the test before trusting it.

• Claimed by OpenAI: state-of-the-art on coding, knowledge work, cybersecurity, science.
• Sourced by OpenAI: Luna beats Fable 5 on Agents' Last Exam at ~99% lower estimated cost per task.
• Not yet independently confirmed here: exact SWE-bench, MMLU, or GPQA numbers.

How to read any benchmark claim critically

Read every benchmark claim by asking who ran it, on which test version, and at what cost. Those three questions catch most misleading numbers.

Watch for tricks. A model tested at maximum reasoning effort may score higher but cost far more and run slower. A score without a cost or speed figure hides half the story.

Also check the benchmark version and date. Test suites change, and an old score against a new competitor is not a fair fight.

• Who ran it: the vendor, or a neutral evaluator?
• At what cost and speed: was the winning run priced like the losing run?
• Which version: same benchmark release and same date for both models?
• Contamination: could the test questions have leaked into training data?

How to check the current numbers yourself

Start with OpenAI's own posts for what the vendor claims, then cross-check against independent evaluators. Using both keeps you honest about the claimed-versus-verified line.

For the primary claims, read OpenAI's GPT-5.6 launch post and the price-performance post. These are the source of the state-of-the-art language and the Agents' Last Exam comparison.

For independent numbers, check Artificial Analysis for intelligence, speed, and cost scores, and LMArena for blind human preference rankings. Compare their figures against OpenAI's claims.

• OpenAI GPT-5.6 launch post — the family claims and tiers.
• OpenAI price-performance post — the Luna vs Fable 5 cost claim.
• Artificial Analysis — third-party intelligence, speed, and cost.
• LMArena — blind, crowd-voted head-to-head rankings.