OpenAI lowered the API worth of its two lower-cost GPT-5.6 fashions on July 30, 2026, slicing the most affordable tier by 80% and the mid-tier by 20% whereas leaving its flagship untouched. The change is logged within the firm’s personal API changelog and is already dwell on the printed fee card.
Per million enter and output tokens, the usual charges now learn:
- 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 speed its predecessor GPT-5.5 nonetheless carries
All three tiers reached normal availability on July 9, 2026 on the greater costs, which places the repricing three weeks into the household’s industrial life.
The cuts run by means of each service tier on the sheet, not simply the headline fee. Batch and Flex processing, each half the usual worth, now put Luna at 10 cents enter and 60 cents output. Cached enter reads, discounted 90%, drop to 2 cents per million tokens on Luna and 20 cents on Terra. Lengthy-context requests, which invoice at double the enter fee and 1.5 instances output, land at 40 cents and $1.80 for Luna. Patrons going by means of Amazon (AMZN ) Bedrock are billed by AWS, and OpenAI notes these charges can differ from its personal.
The cheaper tiers are the place high-volume manufacturing visitors lives: classification, extraction, request routing, first-pass drafting, and the lengthy agent loops the place one person instruction can set off dozens of mannequin calls earlier than it returns a solution. A five-fold lower on the tier absorbing that quantity modifications the arithmetic on which workloads are price automating in any respect.
The place the cheaper tiers sit in opposition to Claude
At 20 cents in and $1.20 out, Luna undercuts Anthropic’s least expensive printed mannequin, Haiku 4.5, by an element of 5 on enter and roughly 4 on output, in response to Anthropic’s pricing web page. Terra’s new fee sits beneath the $3 and $15 that Claude Sonnet 5 is scheduled to cost as soon as its introductory fee of $2 and $10 lapses on August 31, 2026. On the prime of each lineups, Sol nonetheless prices extra on output than Opus 5, which Anthropic costs at $5 and $25.
Precedence processing turns into Quick mode
The identical changelog entry retires Precedence Processing and replaces it with Quick mode. For Sol, OpenAI says Quick mode runs as much as 2.5 instances customary velocity at twice the worth, and the change is backward suitable: requests already tagged for precedence path to Quick mode and not using a code change. Quick-mode charges are $10 and $60 for Sol, $4 and $24 for Terra, and 40 cents and $2.40 for Luna.
Anthropic sells the identical product below the identical identify and the identical phrases — quick mode for Opus 5, as much as 2.5 instances sooner at twice customary pricing. The 2 fee playing cards now converge on region-pinned inference as nicely. OpenAI fees a ten% uplift on fashions launched on or after March 5, 2026 when a buyer requires information residency; Anthropic payments US-only inference at 1.1 instances its customary fee.
What made the cheaper tiers cheaper
OpenAI printed its accounting a day earlier than the lower. In a July 29, 2026 engineering submit, 5 members of its technical workers described optimizations throughout inference and the agent harness behind Codex and ChatGPT Work. Sol, operating inside Codex, rewrote the corporate’s manufacturing GPU kernels; mixed with broader kernel work, OpenAI says that lower end-to-end serving prices by 20%. Sol additionally redesigned its personal speculative-decoding draft mannequin throughout a whole lot of experiments, which the corporate credit with elevating token-generation effectivity by greater than 15%.
These are OpenAI’s personal figures for its personal stack. The submit closed on a dedication to go “under-the-hood enhancements again to our customers and prospects within the type of extra extensively accessible, cost-efficient intelligence.” The speed card adopted a day later.
The harness work factors on the similar price heart the worth lower does. OpenAI caps software output at 10,000 tokens by default and retains model-visible historical past append-only, so an agent loop resending its directions and gear definitions at each step hits the immediate cache as an alternative of paying full enter charges. On a job that takes 30 mannequin requests, that’s 30 probabilities to keep away from recomputing the identical prefix.
The cuts land whereas consumers are auditing inference spend and widening which groups contact the fashions: OpenAI’s personal analysis discovered workers utilizing ChatGPT nicely past their job titles. Amazon’s engineering group moved to cap AI spending after price overruns the identical day, and OpenAI shipped laborious spend limits for API organizations and initiatives on July 22, 2026, letting directors set a month-to-month ceiling that begins returning errors as soon as tracked spend reaches it.
For a workforce already routing bulk work to Luna, the identical calls now price a fifth of what they did, with a ceiling they will set within the dashboard. With Sol’s worth unchanged, the dwell determination stays the place OpenAI has put it for the reason that household shipped: which tier every request requires.
