[ad_1]
OpenAI lowered the API price of its two lower-cost GPT-5.6 models on July 30, 2026, cutting the cheapest tier by 80% and the mid-tier by 20% while leaving its flagship untouched. The change is logged in the company’s own API changelog and is already live on the published rate card.
Per million input and output tokens, the standard rates now read:
- GPT-5.6 Luna: 20 cents and $1.20, down from $1 and $6
- GPT-5.6 Terra: $2 and $12, down from $2.50 and $15
- GPT-5.6 Sol: $5 and $30, unchanged, matching the rate its predecessor GPT-5.5 still carries
All three tiers reached general availability on July 9, 2026 at the higher prices, which puts the repricing three weeks into the family’s commercial life.
The cuts run through every service tier on the sheet, not just the headline rate. Batch and Flex processing, both half the standard price, now put Luna at 10 cents input and 60 cents output. Cached input reads, discounted 90%, drop to two cents per million tokens on Luna and 20 cents on Terra. Long-context requests, which bill at double the input rate and 1.5 times output, land at 40 cents and $1.80 for Luna. Buyers going through Amazon (AMZN ) Bedrock are billed by AWS, and OpenAI notes those rates can differ from its own.
The cheaper tiers are where high-volume production traffic lives: classification, extraction, request routing, first-pass drafting, and the long agent loops where one user instruction can trigger dozens of model calls before it returns an answer. A five-fold cut on the tier absorbing that volume changes the arithmetic on which workloads are worth automating at all.
Where the cheaper tiers sit against Claude
At 20 cents in and $1.20 out, Luna undercuts Anthropic’s cheapest published model, Haiku 4.5, by a factor of five on input and roughly four on output, according to Anthropic’s pricing page. Terra’s new rate sits below the $3 and $15 that Claude Sonnet 5 is scheduled to charge once its introductory rate of $2 and $10 lapses on August 31, 2026. At the top of both lineups, Sol still costs more on output than Opus 5, which Anthropic prices at $5 and $25.
Priority processing becomes Fast mode
The same changelog entry retires Priority Processing and replaces it with Fast mode. For Sol, OpenAI says Fast mode runs up to 2.5 times standard speed at twice the price, and the switch is backward compatible: requests already tagged for priority route to Fast mode without a code change. Fast-mode rates are $10 and $60 for Sol, $4 and $24 for Terra, and 40 cents and $2.40 for Luna.
Anthropic sells the same product under the same name and the same terms — fast mode for Opus 5, up to 2.5 times faster at twice standard pricing. The two rate cards now converge on region-pinned inference as well. OpenAI charges a 10% uplift on models released on or after March 5, 2026 when a customer requires data residency; Anthropic bills US-only inference at 1.1 times its standard rate.
What made the cheaper tiers cheaper
OpenAI published its accounting a day before the cut. In a July 29, 2026 engineering post, five members of its technical staff described optimizations across inference and the agent harness behind Codex and ChatGPT Work. Sol, running inside Codex, rewrote the company’s production GPU kernels; combined with broader kernel work, OpenAI says that cut end-to-end serving costs by 20%. Sol also redesigned its own speculative-decoding draft model across hundreds of experiments, which the company credits with raising token-generation efficiency by more than 15%.
Those are OpenAI’s own figures for its own stack. The post closed on a commitment to pass “under-the-hood improvements back to our users and customers in the form of more widely available, cost-efficient intelligence.” The rate card followed a day later.
The harness work points at the same cost center the price cut does. OpenAI caps tool output at 10,000 tokens by default and keeps model-visible history append-only, so an agent loop resending its instructions and tool definitions at every step hits the prompt cache instead of paying full input rates. On a task that takes 30 model requests, that is 30 chances to avoid recomputing the same prefix.
The cuts land while buyers are auditing inference spend and widening which teams touch the models: OpenAI’s own research found staff using ChatGPT well beyond their job titles. Amazon’s engineering organization moved to cap AI spending after cost overruns the same day, and OpenAI shipped hard spend limits for API organizations and projects on July 22, 2026, letting administrators set a monthly ceiling that starts returning errors once tracked spend reaches it.
For a team already routing bulk work to Luna, the same calls now cost a fifth of what they did, with a ceiling they can set in the dashboard. With Sol’s price unchanged, the live decision stays where OpenAI has put it since the family shipped: which tier each request requires.
[ad_2]
Source_link
