Co-Founder & Lead Programmer of AcceleratedLogic AI
OpenAI announced GPT-5.6 on July 9, 2026 as a family with three API models: Sol, Terra, and Luna. This update replaces the original release-day summary with a source-based guide to the current model IDs, limits, prices, and orchestration features. It does not present unverified benchmark scores as AcceleratedLogic test results.
Choose among Sol, Terra, and Luna
OpenAI describes Sol as the flagship for complex professional work, Terra as a balance of capability and cost, and Luna as the efficient option for cost-sensitive or high-volume workloads. The OpenAI model table lists the API IDs, context limits, and current per-token prices:
- GPT-5.6 Sol (gpt-5.6-sol) — 1,050,000-token context; $4 per million input tokens and $20 per million output tokens in the current model table.
- GPT-5.6 Terra (gpt-5.6-terra) — 1,050,000-token context; $2 per million input and $12 per million output tokens.
- GPT-5.6 Luna (gpt-5.6-luna) — 1,050,000-token context; $0.20 per million input and $1.20 per million output tokens.
OpenAI has announced price updates since launch, so these are a dated snapshot rather than a permanent quote. Check the current model pricing documentation before budgeting. Long prompts, cached tokens, and tool use can change total cost.
Reasoning and API features
The model documentation lists reasoning effort settings from none through max, with settings available depending on product and endpoint. OpenAI's GPT-5.6 model guidance recommends testing the same workload at the current setting and one lower setting before switching a production route. The API changelog describes Programmatic Tool Calling and beta multi-agent orchestration in the Responses API; these are features to configure and test, not automatic behavior of every model call.
For long or tool-heavy requests, estimate the total work across all calls and agents. Lower per-token pricing on Luna does not guarantee lower cost if the task needs many more tokens, retries, or tools; a higher-tier model is not automatically more economical if a smaller one passes the same acceptance tests.
A practical model-selection test
1. Build a small held-out set from real user requests and define success checks before running it.
2. Run Sol, Terra, and Luna with the same input, tools, and acceptance tests. Record the model ID, reasoning effort, latency, token usage, retries, and human corrections.
3. Compare cost per accepted result, not just price per million tokens or a single aggregate leaderboard.
4. Repeat tasks where output variation could change the decision, then retest after model or prompt updates.
OpenAI's launch post includes vendor-run evaluations and their methodology notes. Those results describe OpenAI's published setup. They are a useful shortlist signal, but they do not replace a comparison on your own task distribution.
Takeaway
The three-tier family makes it possible to tune capability and cost per workload. Start with the least expensive tier that meets your task's quality and latency requirements, then escalate only where measured failures justify the change. Recheck official IDs, context limits, and prices before shipping because availability and pricing can change after a model launch.