OpenAI didn’t launch one mannequin on July 9, 2026. It launched a three-tier value menu, and the menu issues greater than any single benchmark rating within the launch notes. Sol, Terra, and Luna now let a purchaser route every job to a mannequin priced for the job as a substitute of defaulting each request to the flagship.
Three fashions divide the workload
GPT-5.6 reached normal availability on July 9, 2026, following a restricted preview introduced in late June. The household splits into Sol, the flagship; Terra, a balanced on a regular basis possibility; and Luna, the quickest and most cost-effective of the three. Builders can attain all three via the API, and normal short-context pricing per million tokens breaks down within the desk beneath.
| Mannequin | Enter | Cached enter | Cache write | Output |
| Sol | $5.00 | $0.50 | $6.25 | $30.00 |
| Terra | $2.50 | $0.25 | $3.125 | $15.00 |
| Luna | $1.00 | $0.10 | $1.25 | $6.00 |
Entry varies by product and plan moderately than following one easy rule. In Codex, Free and Go customers can attain Terra, whereas Plus, Professional, Enterprise, and Enterprise customers can select amongst Sol, Terra, and Luna. ChatGPT Work gives all three fashions to Plus and better plans. Customary ChatGPT conversations work in a different way: Terra and Luna should not selectable there, and eligible paid customers attain Sol via the Medium, Excessive, and Additional Excessive reasoning choices, primarily based on their plan. OpenAI’s present Assist Heart lists Sol Professional, a higher-quality variant for the toughest duties, as out there to Professional, Enterprise, and Enterprise plans, although the unique launch announcement named solely Professional and Enterprise. OpenAI folded GPT-5.6 into eligible current ChatGPT plans moderately than introducing a brand new subscription tier, and normal ChatGPT subscription costs didn’t rise alongside the mannequin launch.
Sol retains GPT-5.5’s base value
The usual short-context price for Sol, $5 enter and $30 output per million tokens, matches GPT-5.5’s headline pricing precisely. OpenAI is positioning Sol as a functionality and effectivity achieve with out elevating the sticker value of its prime mannequin.
The comparability stops on the sticker, although. GPT-5.6 provides cache-write expenses GPT-5.5 by no means billed. OpenAI studies increased scores and decrease output-token use on a number of coding and computer-use evaluations, though the comparability depends upon the benchmark, reasoning setting, and competing mannequin. An actual bill depends upon reasoning effort, device calls, context size, and the way a lot of a immediate will get cached, not on the 2 headline numbers alone.
Terra could also be an important mannequin for enterprise patrons
Terra prices half of Sol and lands on the precise normal pricing GPT-5.4 carried at launch: $2.50 enter, $15 output. For a big share of enterprise workloads, the positioning issues greater than Sol’s frontier scores, as a result of most manufacturing site visitors doesn’t want frontier reasoning.
Notion, quoted in OpenAI’s launch supplies, mentioned many brokers constructed on GPT-5.5 carry out simply as nicely on Terra at half the worth and with 16% fewer tokens. Deal with the determine as a customer-reported outcome from one firm’s workloads, not a common benchmark each crew will replicate. Terra is value a migration check on current brokers earlier than Sol turns into the default assumption for brand spanking new ones.
Luna creates a low-cost routing tier
Luna runs at one-fifth of Sol’s enter value and one-fifth of its output value: $1 and $6 per million tokens. The seemingly house for Luna is high-volume, low-complexity work: buyer help triage, content material classification, extraction, and light-weight coding help, the place throughput and value per name matter greater than peak reasoning high quality.
Luna isn’t mechanically low cost in follow. Output nonetheless prices six occasions enter at normal charges, so a verbose Luna deployment can rack up spend sooner than a concise Sol deployment dealing with fewer, shorter exchanges. OpenAI’s scores additionally present actual gaps between Luna and Sol on tougher evaluations, so routing choices must weigh job problem alongside quantity.
Caching and lengthy context reshape the invoice
GPT-5.6 introduces specific cache breakpoints and a 30-minute minimal cache life. Cached enter reads earn a 90% low cost off the uncached enter price, whereas a cache write prices 1.25 occasions the uncached enter price. Steady system prompts, device definitions, and reference materials change into low cost to reuse after the primary write; prompts altering on each name might rack up write expenses with out sufficient reuse to offset them.
Lengthy-context requests set off the next pricing tier as soon as a immediate exceeds 272,000 enter tokens. OpenAI expenses twice the usual enter price and 1.5 occasions the usual output price for the complete request, not solely the portion above the brink. Sol subsequently rises to $10 per million enter tokens and $45 per million output tokens, Terra rises to $5 and $22.50, and Luna rises to $2 and $9. Cache-write costs double together with the enter price.
A simplified instance illustrates the unfold. For a million uncached enter tokens and 250,000 output tokens at short-context normal charges, Sol prices roughly $12.50, Terra roughly $6.25, and Luna roughly $2.50. Actual payments will differ as soon as instruments, caching, reasoning effort, and regional processing enter the calculation.
Batch, Flex, and Precedence create one other pricing layer
Processing class multiplies the pricing grid once more. Batch and Flex processing cost half the usual short-context price throughout the GPT-5.6 household: Sol falls to $2.50 per million enter tokens and $15 per million output tokens, Terra to $1.25 and $7.50, and Luna to $0.50 and $3. Batch and Flex go well with workloads resembling classification, extraction, and enrichment tolerant of slower or asynchronous execution.
Precedence processing doubles the usual short-context price: Sol rises to $10 per million enter tokens and $60 per million output tokens, Terra to $5 and $30, and Luna to $2 and $12. OpenAI’s pricing web page exhibits solely short-context Precedence charges, so patrons mustn’t assume the identical multiplier extends past 272,000 enter tokens till OpenAI paperwork a long-context Precedence tier. Precedence might justify its price in latency-sensitive, customer-facing functions, however a crew budgeting round normal pricing and deploying Precedence by default can shortly double its model-token spend.
Regional processing introduces one other price layer. OpenAI expenses a ten% uplift for eligible data-residency fashions launched on or after March 5, 2026, although regional storage availability doesn’t mechanically imply mannequin inference occurs inside the chosen area.
Mannequin tier is just one routing choice. Sol additionally gives max, which spends extra compute on deeper reasoning than its excessive setting, and extremely, which coordinates 4 brokers by default. Extremely can end tough work sooner, however parallel execution can increase combination token use. Sol’s reasoning mode capabilities as one other FinOps management, not a easy high quality swap.
Programmatic Instrument Calling provides a second lever. OpenAI says GPT-5.6 can write and run light-weight packages coordinating instruments and filtering intermediate outcomes, as a substitute of routing each device response again via the mannequin. In tool-heavy workflows, OpenAI studies fewer tokens and fewer mannequin spherical journeys because of this. Price management now occurs at a number of layers, not one.
Efficiency per greenback wants workload testing
Vendor benchmarks describe circumstances OpenAI controls, not a purchaser’s precise site visitors. OpenAI’s comparability desk studies 52.7% for Sol on Brokers’ Final Examination and an index rating of 80 on the Synthetic Evaluation Coding Agent Index, with Terra and Luna trailing by broad margins on tougher duties and narrower margins on simpler ones. A separate part of OpenAI’s launch web page cites 53.6 for Sol on the identical examination below completely different settings, a reminder: reasoning configuration modifications the rating as a lot because the mannequin does. OpenAI generated the figures via its analysis harness, they usually deserve attribution as such moderately than therapy as impartial proof.
A procurement choice wants a corporation’s inner eval set: job success price, human correction time, complete tokens consumed, latency, retry price, and value per accomplished job moderately than price per token. A mannequin costing extra per token can nonetheless price much less per accomplished job if it wants fewer retries, and a multi-agent or Extremely configuration ending sooner can nonetheless burn extra tokens in combination. Unbiased, cross-enterprise proof for uniform price financial savings, latency good points, or production-performance enhancements doesn’t but exist outdoors OpenAI’s benchmark tables and the named buyer quotes cited right here.
The routing choice is now a FinOps choice
GPT-5.6’s three-tier construction pushes mannequin choice out of engineering-only territory and into funds planning. Procurement coverage now must specify which mannequin handles which knowledge class, which reasoning degree will get authorised by default, when Precedence processing is allowed, and what occurs when a smaller mannequin’s confidence drops. Organizations benchmarking full workflows in opposition to inner duties, moderately than a single leaderboard rating, will make the higher name on the place Sol earns its value and the place Terra or Luna already do the job.
OpenAI didn’t launch one mannequin on July 9, 2026. It launched a three-tier value menu, and the menu issues greater than any single benchmark rating within the launch notes. Sol, Terra, and Luna now let a purchaser route every job to a mannequin priced for the job as a substitute of defaulting each request to the flagship.
Three fashions divide the workload
GPT-5.6 reached normal availability on July 9, 2026, following a restricted preview introduced in late June. The household splits into Sol, the flagship; Terra, a balanced on a regular basis possibility; and Luna, the quickest and most cost-effective of the three. Builders can attain all three via the API, and normal short-context pricing per million tokens breaks down within the desk beneath.
| Mannequin | Enter | Cached enter | Cache write | Output |
| Sol | $5.00 | $0.50 | $6.25 | $30.00 |
| Terra | $2.50 | $0.25 | $3.125 | $15.00 |
| Luna | $1.00 | $0.10 | $1.25 | $6.00 |
Entry varies by product and plan moderately than following one easy rule. In Codex, Free and Go customers can attain Terra, whereas Plus, Professional, Enterprise, and Enterprise customers can select amongst Sol, Terra, and Luna. ChatGPT Work gives all three fashions to Plus and better plans. Customary ChatGPT conversations work in a different way: Terra and Luna should not selectable there, and eligible paid customers attain Sol via the Medium, Excessive, and Additional Excessive reasoning choices, primarily based on their plan. OpenAI’s present Assist Heart lists Sol Professional, a higher-quality variant for the toughest duties, as out there to Professional, Enterprise, and Enterprise plans, although the unique launch announcement named solely Professional and Enterprise. OpenAI folded GPT-5.6 into eligible current ChatGPT plans moderately than introducing a brand new subscription tier, and normal ChatGPT subscription costs didn’t rise alongside the mannequin launch.
Sol retains GPT-5.5’s base value
The usual short-context price for Sol, $5 enter and $30 output per million tokens, matches GPT-5.5’s headline pricing precisely. OpenAI is positioning Sol as a functionality and effectivity achieve with out elevating the sticker value of its prime mannequin.
The comparability stops on the sticker, although. GPT-5.6 provides cache-write expenses GPT-5.5 by no means billed. OpenAI studies increased scores and decrease output-token use on a number of coding and computer-use evaluations, though the comparability depends upon the benchmark, reasoning setting, and competing mannequin. An actual bill depends upon reasoning effort, device calls, context size, and the way a lot of a immediate will get cached, not on the 2 headline numbers alone.
Terra could also be an important mannequin for enterprise patrons
Terra prices half of Sol and lands on the precise normal pricing GPT-5.4 carried at launch: $2.50 enter, $15 output. For a big share of enterprise workloads, the positioning issues greater than Sol’s frontier scores, as a result of most manufacturing site visitors doesn’t want frontier reasoning.
Notion, quoted in OpenAI’s launch supplies, mentioned many brokers constructed on GPT-5.5 carry out simply as nicely on Terra at half the worth and with 16% fewer tokens. Deal with the determine as a customer-reported outcome from one firm’s workloads, not a common benchmark each crew will replicate. Terra is value a migration check on current brokers earlier than Sol turns into the default assumption for brand spanking new ones.
Luna creates a low-cost routing tier
Luna runs at one-fifth of Sol’s enter value and one-fifth of its output value: $1 and $6 per million tokens. The seemingly house for Luna is high-volume, low-complexity work: buyer help triage, content material classification, extraction, and light-weight coding help, the place throughput and value per name matter greater than peak reasoning high quality.
Luna isn’t mechanically low cost in follow. Output nonetheless prices six occasions enter at normal charges, so a verbose Luna deployment can rack up spend sooner than a concise Sol deployment dealing with fewer, shorter exchanges. OpenAI’s scores additionally present actual gaps between Luna and Sol on tougher evaluations, so routing choices must weigh job problem alongside quantity.
Caching and lengthy context reshape the invoice
GPT-5.6 introduces specific cache breakpoints and a 30-minute minimal cache life. Cached enter reads earn a 90% low cost off the uncached enter price, whereas a cache write prices 1.25 occasions the uncached enter price. Steady system prompts, device definitions, and reference materials change into low cost to reuse after the primary write; prompts altering on each name might rack up write expenses with out sufficient reuse to offset them.
Lengthy-context requests set off the next pricing tier as soon as a immediate exceeds 272,000 enter tokens. OpenAI expenses twice the usual enter price and 1.5 occasions the usual output price for the complete request, not solely the portion above the brink. Sol subsequently rises to $10 per million enter tokens and $45 per million output tokens, Terra rises to $5 and $22.50, and Luna rises to $2 and $9. Cache-write costs double together with the enter price.
A simplified instance illustrates the unfold. For a million uncached enter tokens and 250,000 output tokens at short-context normal charges, Sol prices roughly $12.50, Terra roughly $6.25, and Luna roughly $2.50. Actual payments will differ as soon as instruments, caching, reasoning effort, and regional processing enter the calculation.
Batch, Flex, and Precedence create one other pricing layer
Processing class multiplies the pricing grid once more. Batch and Flex processing cost half the usual short-context price throughout the GPT-5.6 household: Sol falls to $2.50 per million enter tokens and $15 per million output tokens, Terra to $1.25 and $7.50, and Luna to $0.50 and $3. Batch and Flex go well with workloads resembling classification, extraction, and enrichment tolerant of slower or asynchronous execution.
Precedence processing doubles the usual short-context price: Sol rises to $10 per million enter tokens and $60 per million output tokens, Terra to $5 and $30, and Luna to $2 and $12. OpenAI’s pricing web page exhibits solely short-context Precedence charges, so patrons mustn’t assume the identical multiplier extends past 272,000 enter tokens till OpenAI paperwork a long-context Precedence tier. Precedence might justify its price in latency-sensitive, customer-facing functions, however a crew budgeting round normal pricing and deploying Precedence by default can shortly double its model-token spend.
Regional processing introduces one other price layer. OpenAI expenses a ten% uplift for eligible data-residency fashions launched on or after March 5, 2026, although regional storage availability doesn’t mechanically imply mannequin inference occurs inside the chosen area.
Mannequin tier is just one routing choice. Sol additionally gives max, which spends extra compute on deeper reasoning than its excessive setting, and extremely, which coordinates 4 brokers by default. Extremely can end tough work sooner, however parallel execution can increase combination token use. Sol’s reasoning mode capabilities as one other FinOps management, not a easy high quality swap.
Programmatic Instrument Calling provides a second lever. OpenAI says GPT-5.6 can write and run light-weight packages coordinating instruments and filtering intermediate outcomes, as a substitute of routing each device response again via the mannequin. In tool-heavy workflows, OpenAI studies fewer tokens and fewer mannequin spherical journeys because of this. Price management now occurs at a number of layers, not one.
Efficiency per greenback wants workload testing
Vendor benchmarks describe circumstances OpenAI controls, not a purchaser’s precise site visitors. OpenAI’s comparability desk studies 52.7% for Sol on Brokers’ Final Examination and an index rating of 80 on the Synthetic Evaluation Coding Agent Index, with Terra and Luna trailing by broad margins on tougher duties and narrower margins on simpler ones. A separate part of OpenAI’s launch web page cites 53.6 for Sol on the identical examination below completely different settings, a reminder: reasoning configuration modifications the rating as a lot because the mannequin does. OpenAI generated the figures via its analysis harness, they usually deserve attribution as such moderately than therapy as impartial proof.
A procurement choice wants a corporation’s inner eval set: job success price, human correction time, complete tokens consumed, latency, retry price, and value per accomplished job moderately than price per token. A mannequin costing extra per token can nonetheless price much less per accomplished job if it wants fewer retries, and a multi-agent or Extremely configuration ending sooner can nonetheless burn extra tokens in combination. Unbiased, cross-enterprise proof for uniform price financial savings, latency good points, or production-performance enhancements doesn’t but exist outdoors OpenAI’s benchmark tables and the named buyer quotes cited right here.
The routing choice is now a FinOps choice
GPT-5.6’s three-tier construction pushes mannequin choice out of engineering-only territory and into funds planning. Procurement coverage now must specify which mannequin handles which knowledge class, which reasoning degree will get authorised by default, when Precedence processing is allowed, and what occurs when a smaller mannequin’s confidence drops. Organizations benchmarking full workflows in opposition to inner duties, moderately than a single leaderboard rating, will make the higher name on the place Sol earns its value and the place Terra or Luna already do the job.
