OpenAI didn’t launch one model on July 9, 2026. It launched a three-tier price menu, and the menu matters more than any single benchmark score in the release notes. Sol, Terra, and Luna now let a buyer route each task to a model priced for the job instead of defaulting every request to the flagship.
Three models divide the workload
GPT-5.6 reached general availability on July 9, 2026, following a limited preview announced in late June. The family splits into Sol, the flagship; Terra, a balanced everyday option; and Luna, the fastest and cheapest of the three. Developers can reach all three through the API, and standard short-context pricing per million tokens breaks down in the table below.
| Model | Input | Cached input | 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 |
Access varies by product and plan rather than following one simple rule. In Codex, Free and Go users can reach Terra, while Plus, Pro, Business, and Enterprise users can choose among Sol, Terra, and Luna. ChatGPT Work offers all three models to Plus and higher plans. Standard ChatGPT conversations work differently: Terra and Luna are not selectable there, and eligible paid users reach Sol through the Medium, High, and Extra High reasoning options, based on their plan. OpenAI’s current Help Center lists Sol Pro, a higher-quality variant for the hardest tasks, as available to Pro, Business, and Enterprise plans, though the original launch announcement named only Pro and Enterprise. OpenAI folded GPT-5.6 into eligible existing ChatGPT plans rather than introducing a new subscription tier, and standard ChatGPT subscription prices did not rise alongside the model launch.
Sol keeps GPT-5.5’s base price
The standard short-context rate for Sol, $5 input and $30 output per million tokens, matches GPT-5.5’s headline pricing exactly. OpenAI is positioning Sol as a capability and efficiency gain without raising the sticker price of its top model.
The comparison stops at the sticker, though. GPT-5.6 adds cache-write charges GPT-5.5 never billed. OpenAI reports higher scores and lower output-token use on several coding and computer-use evaluations, although the comparison depends on the benchmark, reasoning setting, and competing model. A real invoice depends on reasoning effort, tool calls, context length, and how much of a prompt gets cached, not on the two headline numbers alone.
Terra may be the most important model for enterprise buyers
Terra costs half of Sol and lands on the exact standard pricing GPT-5.4 carried at launch: $2.50 input, $15 output. For a large share of enterprise workloads, the positioning matters more than Sol’s frontier scores, because most production traffic doesn’t need frontier reasoning.
Notion, quoted in OpenAI’s launch materials, said many agents built on GPT-5.5 perform just as well on Terra at half the price and with 16% fewer tokens. Treat the figure as a customer-reported result from one company’s workloads, not a universal benchmark every team will replicate. Terra is worth a migration test on existing agents before Sol becomes the default assumption for new ones.
Luna creates a low-cost routing tier
Luna runs at one-fifth of Sol’s input price and one-fifth of its output price: $1 and $6 per million tokens. The likely home for Luna is high-volume, low-complexity work: customer support triage, content classification, extraction, and lightweight coding assistance, where throughput and cost per call matter more than peak reasoning quality.
Luna isn’t automatically cheap in practice. Output still costs six times input at standard rates, so a verbose Luna deployment can rack up spend faster than a concise Sol deployment handling fewer, shorter exchanges. OpenAI’s scores also show real gaps between Luna and Sol on harder evaluations, so routing decisions need to weigh task difficulty alongside volume.
Caching and long context reshape the bill
GPT-5.6 introduces explicit cache breakpoints and a 30-minute minimum cache life. Cached input reads earn a 90% discount off the uncached input rate, while a cache write costs 1.25 times the uncached input rate. Stable system prompts, tool definitions, and reference material become cheap to reuse after the first write; prompts changing on every call may rack up write charges without enough reuse to offset them.
Long-context requests trigger a higher pricing tier once a prompt exceeds 272,000 input tokens. OpenAI charges twice the standard input rate and 1.5 times the standard output rate for the entire request, not only the portion above the threshold. Sol therefore rises to $10 per million input tokens and $45 per million output tokens, Terra rises to $5 and $22.50, and Luna rises to $2 and $9. Cache-write prices double along with the input rate.
A simplified example illustrates the spread. For one million uncached input tokens and 250,000 output tokens at short-context standard rates, Sol costs roughly $12.50, Terra roughly $6.25, and Luna roughly $2.50. Real bills will differ once tools, caching, reasoning effort, and regional processing enter the calculation.
Batch, Flex, and Priority create another pricing layer
Processing class multiplies the pricing grid again. Batch and Flex processing charge half the standard short-context rate across the GPT-5.6 family: Sol falls to $2.50 per million input tokens and $15 per million output tokens, Terra to $1.25 and $7.50, and Luna to $0.50 and $3. Batch and Flex suit workloads such as classification, extraction, and enrichment tolerant of slower or asynchronous execution.
Priority processing doubles the standard short-context rate: Sol rises to $10 per million input tokens and $60 per million output tokens, Terra to $5 and $30, and Luna to $2 and $12. OpenAI’s pricing page shows only short-context Priority rates, so buyers should not assume the same multiplier extends beyond 272,000 input tokens until OpenAI documents a long-context Priority tier. Priority may justify its cost in latency-sensitive, customer-facing applications, but a team budgeting around standard pricing and deploying Priority by default can quickly double its model-token spend.
Regional processing introduces another cost layer. OpenAI charges a 10% uplift for eligible data-residency models released on or after March 5, 2026, though regional storage availability does not automatically mean model inference happens inside the selected region.
Model tier is only one routing decision. Sol also offers max, which spends more compute on deeper reasoning than its high setting, and ultra, which coordinates four agents by default. Ultra can finish difficult work faster, but parallel execution can raise aggregate token use. Sol’s reasoning mode functions as another FinOps control, not a simple quality switch.
Programmatic Tool Calling adds a second lever. OpenAI says GPT-5.6 can write and run lightweight programs coordinating tools and filtering intermediate results, instead of routing every tool response back through the model. In tool-heavy workflows, OpenAI reports fewer tokens and fewer model round trips as a result. Cost control now happens at several layers, not one.
Performance per dollar needs workload testing
Vendor benchmarks describe conditions OpenAI controls, not a buyer’s actual traffic. OpenAI’s comparison table reports 52.7% for Sol on Agents’ Last Exam and an index score of 80 on the Artificial Analysis Coding Agent Index, with Terra and Luna trailing by wide margins on harder tasks and narrower margins on easier ones. A separate section of OpenAI’s launch page cites 53.6 for Sol on the same exam under different settings, a reminder: reasoning configuration changes the score as much as the model does. OpenAI generated the figures through its evaluation harness, and they deserve attribution as such rather than treatment as neutral proof.
A procurement decision needs an organization’s internal eval set: task success rate, human correction time, total tokens consumed, latency, retry rate, and cost per completed task rather than cost per token. A model costing more per token can still cost less per completed job if it needs fewer retries, and a multi-agent or Ultra configuration finishing faster can still burn more tokens in aggregate. Independent, cross-enterprise evidence for uniform cost savings, latency gains, or production-performance improvements does not yet exist outside OpenAI’s benchmark tables and the named customer quotes cited here.
The routing decision is now a FinOps decision
GPT-5.6’s three-tier structure pushes model selection out of engineering-only territory and into budget planning. Procurement policy now needs to specify which model handles which data class, which reasoning level gets approved by default, when Priority processing is allowed, and what happens when a smaller model’s confidence drops. Organizations benchmarking full workflows against internal tasks, rather than a single leaderboard score, will make the better call on where Sol earns its price and where Terra or Luna already do the job.
