OpenAI Cuts GPT-5.6 Luna 80% and Terra 20% as Sol Optimises Its Own Inference Stack
OpenAI cut the price of GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20% on July 30, turning the lower two tiers of the 5.6 family into a direct volume play. Luna now costs $0.20 per million input tokens and $1.20 per million output tokens. Terra now costs $2 input and $12 output. Sol pricing stays unchanged, but the API gets a new Fast mode that runs up to 2.5x faster than standard processing at twice the price.
The surface story is cheaper tokens. The structural story is that OpenAI is now explicitly using its frontier model to improve the economics of serving that same model family.
New GPT-5.6 Pricing
| Model | Old Input | Old Output | New Input | New Output | Change |
|---|---|---|---|---|---|
| GPT-5.6 Luna | $1.00/M | $6.00/M | $0.20/M | $1.20/M | 80% lower |
| GPT-5.6 Terra | $2.50/M | $15.00/M | $2.00/M | $12.00/M | 20% lower |
| GPT-5.6 Sol | $5.00/M | $30.00/M | $5.00/M | $30.00/M | unchanged |
| GPT-5.6 Sol Fast | n/a | n/a | 2x standard | 2x standard | up to 2.5x faster |
The change also affects ChatGPT Work and Codex quota accounting: Luna and Terra usage consumes fewer subscription credits. That matters more than the headline API price for teams already rationing agent runs inside fixed monthly plans.
Why Luna Is the Real Product Move
Luna launched as the cheap 5.6 tier at $1/$6 per million tokens. At $0.20/$1.20, it becomes something different: a subagent model for background automation, classification, routine implementation, test writing, and any workflow where a stronger model can plan while a cheaper model executes.
OpenAI is positioning Sol as the scarce frontier planner and Luna as the default worker. That is the same architecture buyers are already assembling manually with Claude Code, Cursor, Devin, and internal agent platforms: route uncertainty to the expensive model, route repeatable work to the cheaper one, then let the harness decide how much intelligence each step needs.
The new prices make that architecture easier to defend financially. Luna output is now 25x cheaper than Sol output. If a coding agent can hand even half of its implementation and verification calls to Luna without quality loss, the blended cost of a long run changes materially.
Sol as Infrastructure Engineer
OpenAI says GPT-5.6 Sol helped reduce the end-to-end cost of serving the model by 20%. The work included production-kernel rewrites, routing improvements, load-balancing changes, and workload-specific serving configuration.
The most concrete technical claim is in token generation. Sol designed and ran hundreds of experiments on its own draft model for speculative decoding, then monitored training and intervened during hardware failures and instability. OpenAI says those improvements increased token-generation efficiency by more than 15%.
That is not model training recursive self-improvement in the science-fiction sense. It is narrower and more operational: the model writes GPU kernels, tests serving changes, tunes draft-model architecture, and helps engineers ship lower-cost inference. But it is still a feedback loop where capability improves the economics of deploying capability.
Fast Mode Replaces Priority Processing
Fast mode is the other half of the announcement. GPT-5.6 Sol Fast replaces Priority Processing in the API and maps to /fast in Codex. Existing API calls tagged as priority continue to work.
The pricing is blunt: twice the standard Sol price for up to 2.5x faster response. That is not cheap, but it is rational for latency-sensitive agent loops. A human waiting on a code review, incident diagnosis, or customer-support escalation often cares more about wall-clock time than token price. A background batch job does not.
The product split is becoming clear. Sol Fast is for scarce human attention. Sol Standard is for high-stakes reasoning. Luna is for volume. Terra is awkwardly placed between them unless it maintains a visible quality edge over Luna in enterprise workloads.
The Buyer Signal
This is the first GPT-5.6 announcement where pricing, quota accounting, and inference engineering are the story rather than raw benchmark movement. That is what frontier AI looks like after the capability race compresses: buyers still want the best model, but they increasingly buy the cheapest reliable unit of completed work.
OpenAI is trying to make that unit measurable across the whole stack: model, harness, context management, caching, routing, and GPU kernels. The claim to watch is not whether Luna is cheap. It is whether Luna can reliably take enough work away from Sol to make full agentic workflows affordable at production scale.