OpenAI’s July 30, 2026 pricing update makes one operational point clear: businesses should stop treating a single model as the default for every AI task. GPT-5.6 Luna now costs 80% less and Terra 20% less, while Sol gains a higher-priced Fast mode for latency-sensitive work. For teams building agents, the opportunity is to route work by the cost of error, required quality, volume, and response-time target.
What changed in GPT-5.6 pricing
OpenAI set GPT-5.6 Terra pricing at $2 per million input tokens and $12 per million output tokens. Luna is priced at $0.20 per million input tokens and $1.20 per million output tokens. Terra and Luna remain available through the API, ChatGPT Work, and Codex.
The other change is a service-tier shift. Sol Fast mode replaces Priority Processing, offering up to 2.5 times Standard-processing speed at twice the price, with no claimed change in intelligence. Existing API requests tagged priority automatically move to Fast mode.
The business implication: build a model-routing policy
Lower unit prices do not automatically lower the cost of an agent. An expensive model used unnecessarily on repetitive steps can still dominate spend; a cheaper model used on a high-stakes decision can create rework, escalation, or customer risk. The right measure is outcome cost: model spend plus retries, human review, latency, and the cost of mistakes.
A practical pattern is to use Luna for well-specified, high-volume actions such as classification, extraction, routing, data cleanup, routine follow-ups, and test execution. Use Terra for everyday knowledge work where more context handling or judgment matters. Reserve Sol for ambiguous planning, exception handling, complex code or analysis, and decisions where failure is costly. Use Fast mode only when faster completion creates a measurable business benefit.
How to apply this to an AI agent workflow
Start by mapping every step in a workflow rather than choosing a model at the agent level. A support-resolution agent, for example, may use Luna to classify an incoming issue and retrieve candidate policies; Terra to draft a grounded response; and Sol only when the request involves an exception, a sensitive account, or an unresolved conflict.
Then establish an evaluation set from real work. Include normal cases, edge cases, and failures that would trigger escalation. For each route, track completion quality, handoff rate, latency, token cost, and the rate at which human reviewers correct the output. Move a step to a lower-cost model only after its quality clears the agreed threshold; move it back when the cost of errors rises.
Finally, make routing observable. Log the route selected, the reason, the outcome, and review signals. That record turns model choice into an adjustable operations policy instead of a one-time architecture decision.
What not to do after a price cut
Do not simply increase autonomy because a model is cheaper. Tool permissions, approval gates, retrieval quality, and audit trails remain the controls that determine whether an agent is safe to run at higher volume. Also do not judge routes on benchmark claims alone. Test them against the specific documents, systems, customers, and failure modes that define your business process.
The durable advantage is not access to the lowest token price. It is an agent design that applies the least expensive capable model to each step while preserving review for consequential work.