OpenAI's GPT-5.6 family dropped this week, and Artificial Analysis has the benchmarks. The headline: GPT-5.6 Sol (max) scores 59 on the Artificial Analysis Intelligence Index — one point below Claude Fable 5 — at approximately one third of the cost.
"GPT-5.6 Sol costs $1.04 per task in the Artificial Analysis Intelligence Index — offering a similar level of intelligence to Claude Fable 5 at approximately one third of the cost."
That's not a minor efficiency bump. It's a meaningful shift in the cost/intelligence curve.
What actually changed
- Three-tier family: Sol, Terra, and Luna. Sol is the flagship; Terra (~50% cheaper than Sol) and Luna (~80% cheaper) trade down on intelligence but stay on the Pareto frontier ahead of GPT-5.5 at every effort level.
- Coding agent leader: GPT-5.6 Sol (max) in OpenAI's Codex harness scores 80 on the new Artificial Analysis Coding Agent Index — first across all three evaluations (DeepSWE, Terminal-Bench v2, SWE-Atlas-QnA). It's also ~40% cheaper per task than Claude Fable 5 in Claude Code for comparable coding work.
- Low token use: Sol (max) uses ~15k output tokens per Intelligence Index task vs 16k for GPT-5.5 — and fewer than Claude Opus 4.8 and Gemini 3.5 Flash at similar intelligence levels.
- Best presentation outputs: Sol (max) takes the top Presentation Elo in the AA-Briefcase benchmark — its PowerPoint and Excel outputs rated most visually polished of any model tested.
The cache-write pricing wrinkle
GPT-5.6 introduces something new for OpenAI: cache-write pricing. Sol, Terra, and Luna are priced at $5/$30, $2.5/$15, and $1/$6 per million input/output tokens respectively. Cache reads stay at 90% discount — but cache writes now cost 1.25× the base input price.
This mirrors Anthropic's model. The logic: cached tokens occupy memory whether or not they're reused, so the write cost reflects real infrastructure cost. Fair enough — but if you're building agents with long shared contexts, this will show up in your bills. Worth a line item in your cost model.
What to do
- Evaluating frontier models for production? GPT-5.6 Sol is now the clearest cost challenger to Claude Fable 5 for general intelligence tasks. Run your own evals on your actual workloads — benchmark scores are a starting point, not a decision.
- Running coding agents? Sol in the Codex harness leads the field right now. If you've been on Claude Code for agentic coding, the cost delta is worth testing.
- Cost-sensitive use cases? Luna at $1/$6 per million tokens is on the Pareto frontier — more intelligent per dollar than GPT-5.5, GLM-5.2, and Gemini 3.5 Flash.
- Building with long cached contexts? Account for the new cache-write premium in your cost projections. 1.25× on writes is not free.
The full benchmark breakdown — including per-model effort-level comparisons — is at artificialanalysis.ai.
✏️ Drafted with KewBot (AI), edited and approved by Drew.
Top comments (0)