GPT-5.6 API Pricing: Sol vs Terra vs Luna

OpenAI’s GPT-5.6 family now spans three API tiers with a 25-to-1 difference in standard token prices. That makes model selection more important than simply choosing the newest name. The practical question is: which tier meets your quality target at the lowest total cost?

This guide compares GPT-5.6 Sol, Terra, and Luna using OpenAI’s current documentation and pricing as of July 31, 2026. It also shows how long context, caching, Batch, Fast mode, and tool fees can change the bill.

GPT-5.6 API pricing at a glance

OpenAI positions Sol as the frontier tier, Terra as the intelligence-and-cost balance, and Luna as the high-volume, cost-sensitive tier. All prices below are for one million text tokens under standard, short-context processing.

The prices come from OpenAI’s live API pricing table. At every listed standard token rate, Terra costs 40% as much as Sol, while Luna costs 10% as much as Terra and 4% as much as Sol.

The three tiers share several important specifications. Each accepts text and image input, returns text, supports reasoning, has a 1,050,000-token context window, allows up to 922,000 input tokens, and can produce up to 128,000 output tokens. Their documented knowledge cutoff is February 16, 2026.

Sol vs Terra vs Luna: which model should you choose?

Choose Sol to establish the quality ceiling

GPT-5.6 Sol is the appropriate baseline for complex analysis, difficult coding and review, high-stakes document work, and agent tasks where a wrong decision is expensive. OpenAI describes it as the family’s frontier model for complex professional work.

One naming detail can affect deployments: the unsuffixed GPT-5.6 alias routes to Sol. If an application uses that alias while the team assumes it selects a cheaper default, costs may be much higher than expected. Use the explicit Sol, Terra, or Luna model identifier when predictable routing and cost attribution matter.

Sol should not automatically process every request. Its standard output tokens cost 25 times Luna’s. Use Sol to determine the best achievable result on your evaluation set, then test whether a lower tier preserves the required quality.

See OpenAI’s GPT-5.6 Sol model page for its current endpoints, tools, rates, and limits.

Start production testing with Terra

Terra is the strongest default candidate for many business applications: research synthesis, support drafting, document extraction with judgment, code assistance, and multi-step agent workflows. It retains the family’s large context window and supported tool set while charging 60% less than Sol at standard short-context rates.

That does not mean Terra is guaranteed to match Sol on your workload. Model labels and public benchmarks cannot substitute for application-specific evaluation. Test exact prompts, documents, tools, output schemas, edge cases, and failure costs before switching production traffic.

OpenAI’s GPT-5.6 Terra documentation describes it as the tier that balances intelligence and cost.

Route simple, repeatable volume to Luna

Luna is designed for tasks where throughput and unit economics dominate: classification, tagging, simple extraction, moderation pre-screening, query routing, short summaries, and other tightly specified transformations.

Its low rate can materially change the economics of a large workload. However, cheap tokens do not rescue a workflow that frequently requires human correction or a second model pass. Luna is the right choice only when it clears the same acceptance threshold used for the larger tiers.

The GPT-5.6 Luna model page also shows higher top-tier throughput limits than Sol and Terra, reinforcing its high-volume positioning.

A realistic cost example

Consider 100,000 document-processing jobs per month. Each job sends 20,000 input tokens and receives 2,000 output tokens. The calculation below assumes short-context standard pricing, no tool calls, no regional uplift, and no retries.

The model decision creates a larger saving than most prompt-level optimizations. Moving from Sol to Terra saves $9,600 per month in this scenario. Moving from Terra to Luna saves another $5,760—but only if Luna passes the workload’s quality tests.

Caching can reduce costs further when the input contains a repeated prefix. If 15,000 of the 20,000 input tokens are served as cached reads, the same 100,000 standard jobs would cost approximately $9,250 on Sol, $3,700 on Terra, or $370 on Luna. Those figures assume the cached-read rate applies and exclude cache-write charges.

For background jobs that do not need immediate results, OpenAI’s Batch rates are half the corresponding standard rates in the current pricing table. That makes Batch a straightforward optimization for offline enrichment, nightly classification, bulk extraction, and evaluation runs.

Four pricing details that can surprise you

1. Long context changes the rate for the whole request

The headline prices are short-context rates. When a prompt exceeds 272,000 input tokens, OpenAI charges twice the input rate and 1.5 times the output rate for the full request—not just the tokens above the threshold.

Do not attach an entire repository or document archive by default. Retrieve the relevant sections, compact old conversation state, and monitor input size before crossing the threshold.

2. Cache reads are cheap, but cache writes are not free

Cached input is billed at one-tenth the uncached input rate for all three GPT-5.6 tiers. However, cache writes cost 1.25 times the uncached input rate. Caching pays when a stable prefix is reused enough times; frequently changing prefixes may create writes without enough discounted reads.

OpenAI’s GPT-5.6 usage guide recommends tracking cached tokens and cache-write tokens rather than assuming every cache hit produces net savings.

3. Batch, standard, and Fast mode solve different problems

Batch and Flex currently use half-price GPT-5.6 token rates. Fast mode charges more for prioritized latency: twice the standard GPT-5.6 rates in the current table. OpenAI renamed Priority processing to Fast mode on July 30, 2026, while continuing to accept both service-tier names.

Use Batch for asynchronous work, standard for normal interactive traffic, and Fast only where measured latency has enough business value to justify the premium.

4. Token prices are not the whole bill

Web search, file search, containers, and other tools can add per-call, storage, or session fees. Eligible regional-processing endpoints also carry a 10% uplift for models released on or after March 5, 2026. Retries, verbose outputs, and multi-agent fan-out can multiply token use as well.

For tool-heavy applications, combine model usage with tracing rather than estimating from request count alone. Our OpenTelemetry AI agent tracing guide explains the observability layer, while the Responses API shell tools guide covers the broader tool workflow.

A safer migration process

OpenAI’s model-selection guidance recommends optimizing for accuracy first, then cost and latency. A practical rollout follows six steps:

1. Define a passing score. Specify task success, schema validity, factual accuracy, acceptable latency, and the cost of an error.
2. Build a representative evaluation set. Include normal requests, difficult edge cases, long inputs, tool failures, and examples that previously required human correction.
3. Establish the Sol baseline. Run the strongest tier first so you know whether the workflow is viable and what quality ceiling is available.
4. Test Terra and Luna on identical inputs. Compare pass rate, human-review rate, tokens, latency, tool calls, and total cost—not style preference alone.
5. Test reasoning settings separately. OpenAI advises starting with the reasoning effort used on GPT-5.4 or GPT-5.5, then comparing the same level and one level lower. Lower effort can reduce cost and latency if quality holds.
6. Roll out with routing and monitoring. Send simple accepted cases to Luna, ambiguous work to Terra, and the hardest or highest-risk cases to Sol. Keep a fallback path, but cap retries so one request cannot trigger an uncontrolled cascade.

Re-run the evaluation when prompts, tools, retrieval data, model behavior, or pricing changes. A routing policy is an operating system, not a one-time purchasing decision.

Limitations to plan around

GPT-5.6 Sol, Terra, and Luna do not support fine-tuning in their current model documentation. They also output text rather than audio, image, or video, although they can accept image input and invoke supported tools through the Responses API.

The large context window is a capacity limit, not evidence that every task benefits from maximum context. Very long prompts cost more, may increase latency, and can bury relevant evidence. The February 16, 2026 knowledge cutoff also means current applications still need retrieval or web search for newer facts.

Finally, vendor positioning is not an independent guarantee of performance. Production decisions should rest on your own test set, observed error costs, and total workflow spend.

Conclusion

For most teams, the best GPT-5.6 API strategy is not one model. Use Sol to set the quality target, Terra as the main production candidate, and Luna for validated high-volume tasks. Then apply caching to stable repeated prefixes and Batch to work that can wait.

The largest saving comes from routing each task to the smallest tier that passes a measurable acceptance test. Price per token matters, but cost per successful outcome is the metric that should decide the deployment.