GPT-6 Sol and Luna Launch at Half the Price of GPT-5.6, With Mixed Benchmark Results
OpenAI has released GPT-6 Sol and Luna into the API, ChatGPT Work, and Codex — extending the GPT-6 family beyond Astra with models positioned for scale and cost efficiency rather than raw capability.
The headline number is a 50% price reduction across both tiers. Sol drops from $4/$20 to $2/$10 per million input/output tokens. Luna falls from $0.20/$1.20 to $0.10/$0.50. Cache read discounts are unchanged at 90%.
What Changed on Benchmarks
The picture is mixed. Sol improves where it matters most for Codex workloads but regresses on complex knowledge tasks.
Coding Agent Index (Artificial Analysis): Sol gains 2 points to 57. Its Terminal-Bench 4.0 score rises from 37% to 43%, and SWE-Atlas-QnA from 54% to 58%. Luna regresses by 2 points to 41, with lower SWE-Atlas-QnA (44% vs 49%) and DeepSWE v1.1 (64% vs 66%).
Hallucination (AA-Omniscience): Sol cuts its error rate from 92% to 60% — driven partly by declining to answer more questions (attempts 83% vs 99%). Luna improves from 93% to 77% wrong. The tradeoff: Sol’s factual accuracy drops 5 points to 54% on attempted questions.
Knowledge work regressions: Both models drop in GDPval-AA v2.1 — Sol by roughly 100 Elo points, Luna by 75. Luna also falls ~45 Elo on AA-Briefcase v1.1, Artificial Analysis’s multi-week project benchmark. The regressions are driven by shorter deliverables that more frequently omit required elements; Sol uses roughly 31k output tokens per task and Luna 51k, both slightly above GPT-5.6 levels.
Cost per task: Sol runs the Artificial Analysis Intelligence Index at $1.06 per task — about 50% less than GPT-5.6 Sol at $1.99. Luna costs $0.07 per task versus $0.18 for its predecessor.
Positioning in the GPT-6 Family
GPT-6 Astra, launched earlier this month, remains the capability leader. Sol and Luna inherit “much of its strengths,” per the OpenAI announcement, and both show alignment improvements over their GPT-5.6 counterparts.
The 50% price reduction establishes Sol and Luna firmly on the Pareto cost-efficiency frontier below Astra. For teams running high-volume inference — Codex pipelines, customer-facing agents, coding assistants — the economics shift meaningfully.
Availability
Both models are live in the API today. ChatGPT Work and Codex access extends to Plus, Pro, Business, Enterprise, and Edu subscribers. Free and Go users get Luna access on the desktop app. OpenAI also loaded a banked usage limit credit into Plus, Pro, and Business accounts alongside the launch.
Key Numbers
| Model | Input | Output | Coding Agent Index | Terminal-Bench 4.0 |
|---|---|---|---|---|
| GPT-6 Sol (max) | $2/M | $10/M | 57 (+2) | 43% |
| GPT-6 Luna (max) | $0.10/M | $0.50/M | 41 (-2) | 13% |
| GPT-5.6 Sol (max) | $4/M | $20/M | 55 | 37% |
| GPT-5.6 Luna (max) | $0.20/M | $1.20/M | 43 | 12% |