GPT-5.6 benchmarks: Sol tops coding, cuts costs

Benchmarking GPT-5.6 can now be read as a cost story, not just a score story. Artificial Analysis says OpenAI’s new family lands with three tiers, Sol, Terra and Luna, and a pricing model that changes how cached tokens are billed.

Artificial Analysis benchmarked GPT-5.6 Sol, Terra and Luna across intelligence, coding speed and cost.

What changed

Artificial Analysis says OpenAI pre-released GPT-5.6 Sol, Terra and Luna for evaluation. In the Artificial Analysis Intelligence Index, Sol (max) scored 59, just one point behind Claude Fable 5 (max), while costing about one third as much per task.

Terra and Luna trail Sol on raw intelligence, but they shift the price curve. Terra (55) and Luna (51) come in at roughly 50% and 80% lower cost per task than Sol, and Luna is said to match or beat GLM-5.2 and Gemini 3.5 Flash at lower cost. Artificial Analysis says Sol and Luna stay on the Pareto frontier across reasoning levels, while Terra does not.

• Sol: 59 intelligence score, $1.04 per task
• Terra: 55 intelligence score, $0.55 per task
• Luna: 51 intelligence score, $0.21 per task
• Sol leads the Coding Agent Index at 80

On coding, Sol (max) leads the new Coding Agent Index with a score of 80 in OpenAI’s Codex harness. That index combines DeepSWE, Terminal-Bench v2 and SWE-Atlas-QnA, and the report says Sol tops all three evaluations, while also undercutting Claude Fable 5 and Claude Opus 4.8 on per-task cost.

Why it matters

The update gives buyers a cleaner tradeoff chart. Teams choosing between model quality and spend now have three GPT-5.6 price tiers, plus cache-read discounts and a new cache-write charge that better reflects memory use. For developers running agentic workloads, that can change which model is cheapest at scale.

It also shows where OpenAI is pressing hardest: coding agents and enterprise-style document work. Sol posts the highest Presentation Elo in AA-Briefcase and ranks second overall there, which points to stronger output for slides, spreadsheets and other mixed-format tasks. The main question is whether teams value Sol’s top-end coding and near-frontier intelligence enough to pay for it, or whether Luna’s lower cost is the better fit.

For now, GPT-5.6 looks less like one model launch than a pricing and performance ladder. The real choice is not whether GPT-5.6 is good, but which tier is cheapest for the job.