OpenAI reduced the API price of GPT-5.6 Luna by 80 percent and GPT-5.6 Terra by 20 percent starting July 30, 2026. Luna now costs $0.20 per million input tokens and $1.20 per million output tokens, Terra costs $2 and $12, and the flagship Sol model is unchanged at $5 and $30. OpenAI frames the cut as passing along its own serving efficiency, stating that GPT-5.6 Sol autonomously rewrote and optimized production kernels in a human-led process, reducing end-to-end serving cost by 20 percent and increasing token-generation efficiency by more than 15 percent. ASAP works from OpenAI's two official announcements to lay out the structure behind the price list.

The new prices apply to the two lower tiers only

OpenAI's July 30, 2026 pricing change touches only the bottom two models of the three-model GPT-5.6 family. Luna, the fastest and most affordable model, fell 80 percent to $0.20 per million input tokens and $1.20 per million output tokens. Terra, the balanced model for everyday work, fell 20 percent to $2 and $12. Sol's pricing is unchanged.

The change does not stop at the API. OpenAI states that the lower Luna and Terra rates are also reflected in how usage counts against paid subscriptions in Codex and ChatGPT Work: subscription prices and quota budgets stay the same while Terra and Luna usage now consumes fewer credits. By tier, Free and Go users can access Terra, while Plus, Pro, Business, and Enterprise users can choose Terra and Luna.

The same announcement introduced a speed option. Fast mode replaces Priority Processing in the API, and for GPT-5.6 Sol it delivers up to 2.5 times the speed of Standard processing at twice the price with no change in intelligence. Existing requests tagged priority automatically use Fast mode, so the change is backward compatible.

OpenAI attributes the cut to efficiency its own model produced

The justification OpenAI offers is cost structure rather than sales strategy. Within a human-led process, the company says, GPT-5.6 Sol autonomously rewrote and optimized production kernels, designed and ran hundreds of experiments to improve token generation, and monitored training and intervened when problems arose. The kernel work cut the end-to-end cost of serving the model by 20 percent, and the experiments raised token-generation efficiency by more than 15 percent.

OpenAI is equally explicit that efficiency does not come from the model alone. Better routing keeps hardware productive, smarter context management prevents agents from repeating completed work, and stronger tools and product design reduce the number of steps a task requires. The supporting number is the ARC-AGI-3 public task set: improvements to retained reasoning and context management raised GPT-5.6 Sol's score from 13.3 percent to 38.3 percent while using six times fewer output tokens. The model did not change; the surrounding system did.

The company's price-performance claims are specific as well. OpenAI states that Luna delivers performance comparable to models that were frontier-class a year ago at roughly 6 cents on the dollar per task and at nearly nine times the speed. On professional work as measured by Agents' Last Exam, it says Luna outperforms Fable 5 at an estimated cost per task nearly 99 percent lower.

The gap between 80 percent and 20 percent is the real message

Cutting one model by 80 percent, another by 20 percent, and holding the flagship flat is a tier redesign rather than a uniform price cut. The direction the spread points is clear: compete on price in the cheap high-volume band and defend price at the frontier. Adding Fast mode at twice the price while leaving Sol's base rate untouched reinforces that reading. What OpenAI is selling at the top is not a token rate but response speed and reliability.

Where the pressure comes from is legible too. July 30, 2026, the day OpenAI cut its low-end rate roughly fivefold, sits one day before DeepSeek opened the public beta of V4-Flash on July 31, in a stretch when price competition was concentrated in exactly that high-volume band. OpenAI's own announcement names no competitor and explains the move purely as passing along efficiency gains. The two accounts are not contradictory: unit costs falling and a reason arriving to act on that now are different things, and the announcement addresses only the first.

For practitioners, the consequence of the spread is workflow decomposition rather than model selection. OpenAI offers the example itself: a coding workflow might use Sol to resolve uncertainty and define the plan, then use Luna to implement well-specified changes, write and run tests, and evaluate results. Intelligence tiers change several times inside a single workflow. Picking one model and running everything through it is now the most expensive option on this price list.

Counting the cost of a successful outcome instead of the token rate

The measure OpenAI repeats across both announcements is the cost of a successful outcome, not the token rate. Customers do not buy tokens for their own sake; they want the support issue resolved, the software shipped, the contract reviewed, or the scientific question answered, and the true cost includes time, retries, oversight, and errors. On that basis a stronger model that finishes the work correctly the first time can be cheaper overall than a low-cost model that requires repeated attempts and human intervention.

The framing favors the company, but that does not make it wrong, because it is testable. An organization that logs retry rates and human intervention time can compare the real cost of both configurations. An organization that switches models by reading the rate card alone has no basis for either accepting or refuting the claim. Whether to adopt this framing is a question of measurement infrastructure, not belief.

There is a trap attached. Cost per successful outcome only holds when the quality bar is fixed. Slide the tolerance wider while moving down to a cheaper model and cost falls automatically, rendering the comparison meaningless. OpenAI's instruction to define the outcome and quality standard first and then use evaluations to find where additional intelligence materially improves the result is precisely the sequence that avoids this. Accept the price cut without an evaluation harness and what gets cut may be quality rather than cost.

What teams should actually change this week

The immediate gain is that the break-even point for high-volume work moves down. Large-scale document analysis, customer-interaction classification, and routine implementation, all tasks whose per-item value was too low to justify automation, start to pencil out at the new rates. The output-side cut matters most here: going from $6 to $1.20 per million output tokens lands directly on summarization and generation workloads, where output dominates.

The second target is routing. Splitting a workflow across tiers is the single largest saving available on this price list, but it only works when each stage has a quality bar and an escalation condition for handing failures up to a stronger model. Without those, the cost of humans repairing cheap-model output eats the savings. Document the per-stage acceptance criteria before introducing routing, not after.

The third is procurement and budgeting. In a market where a specific model's price falls to a fifth within three weeks of the GPT-5.6 family's July 9, 2026 launch, locking an annual token budget to a fixed unit price is closer to a loss than a hedge. Build rate-change clauses into contracts, and keep prompts and evaluation sets model-independent so switching stays cheap.

What is unverified and worth watching

The first open question is durability. OpenAI attributes the cut to efficiency, but a 20 percent reduction in serving cost and a 15 percent gain in token-generation efficiency do not, on the published numbers, add up to absorbing an 80 percent price cut on Luna. Whether that gap closes through further efficiency work or represents margin spent on share will only become visible in later quarters.

The second is Luna's real-world quality. The company cites Agents' Last Exam and per-task cost comparisons, but those are metrics OpenAI selected. Any organization putting a low-cost tier into production should compare on its own evaluation set with retry rates included, and there is no guarantee the ratios will match the announcement.

The third is when the new rates actually land. OpenAI states that the pricing changes began rolling out in AWS later on the day of the announcement. Organizations consuming the API through cloud marketplaces or resellers need to confirm separately when the change appears on their own invoices.

Source: ASAP analysis based on OpenAI's official announcements "Advancing the price-performance frontier with GPT-5.6" (July 30, 2026) and "Building abundant intelligence" (July 31, 2026)