On July 30, 2026, OpenAI announced a set of price and speed changes for GPT-5.6. The important change is not just a model label. OpenAI is putting model cost, response speed, and agent workflow design into the same decision: use Sol for harder planning and uncertainty, then route routine implementation and verification to the lower-cost Terra or Luna tiers.
OpenAI lists API pricing at $2 per 1 million input tokens and $12 per 1 million output tokens for Terra, and $0.20 and $1.20 for Luna. The company says Luna is 80% cheaper than the previous tier and Terra is 20% cheaper, while Sol pricing is unchanged. Actual spend will still depend on input-output mix, cache behavior, tool calls, and request volume. Unit price alone is not workflow cost.
Fast mode also replaces Priority Processing. OpenAI says Fast mode can make Sol up to 2.5 times faster than Standard at twice the price, while the existing priority API parameter remains backward compatible. This turns latency into a selectable resource: interactive debugging or an approval loop may justify the premium, while overnight batch testing may not.
OpenAI uses a simple agent split to illustrate the routing model: Sol plans and resolves uncertainty, while Luna implements, tests, and evaluates. That is closer to how real agent work behaves than sending every request to the strongest model. A workflow contains several steps, and each step can have a different requirement for reasoning depth, latency, and cost. The key is to define success and escalation conditions rather than hard-code one model everywhere.
OpenAI also claims that Luna can complete tasks comparable to frontier-class models from a year earlier for about six cents and at nearly nine times the speed. In an Agents’ Last Exam comparison, it says Luna reaches nearly 99% lower cost per task than Fable 5. These are vendor-reported benchmarks or estimates, not independent validation. Teams should measure task success, retries, tool spend, data controls, and human review time; a cheaper token can be offset by more failed runs.
The company further says Sol helped rewrite and optimize kernels, reducing serving cost by 20% and improving token-generation efficiency by 15%. That is also a provider claim, useful as a direction but not a result every API customer should expect to reproduce. Models, system prompts, cache hit rates, hardware, and serving configuration all affect the outcome.
For agent teams, the practical lesson is to make model choice part of workflow policy. Define the minimum success criteria, acceptable latency, per-task budget, and actions that require approval, then route requests based on those signals. Stronger models can remain available for uncertainty while cheaper models handle repeatable work.
Lower prices do not remove deployment risk. Availability, regional rollout, and service conditions can change, and OpenAI says some AWS availability will come later. A safer migration is to replay a small sample of production traces and compare success, latency, tokens, cache use, tool errors, and human intervention before switching. The decision should come from the team’s own workload, not one announcement benchmark.
