Identity and current terms
| Provider / API ID | OpenAI / gpt-5.6-terra |
|---|---|
| Context / maximum output | 1,050,000 / 128,000 tokens |
| Reasoning effort | None through max; medium is documented as default |
| Standard short-context price | $2.00 uncached input, $0.20 cached input, $12.00 output per 1M tokens |
| Long-context condition | Above 272K input tokens, 2× input and 1.5× output apply to the full request |
The premium is clear; the payoff is not
Terra’s like-for-like token rates are 10× Luna’s. The illustrative 80K-input/10K-output basket costs $0.28 on Terra and $0.028 on Luna. This fixed basket is arithmetic, not an agent-workflow prediction.
OpenAI reports Terra above Luna on several coding-related evaluations. Those results can justify a matched test, but they do not disclose your codebase’s acceptance rate, retries, or review minutes. Do not convert a benchmark difference into imagined labor savings.
A defensible escalation rule
First reproduce the cheaper model’s failure. Confirm the environment and tools are working, the relevant files were available, and the acceptance test describes the intended behavior. Then rerun the identical task with the stronger configuration.
Keep Terra only for a category where it repeatedly creates accepted work Luna cannot, or where measured reduction in rework matters more than the extra spend. A single impressive patch is not a stable breakpoint.
What to record
- Starting commit, exact task, and acceptance checks.
- Requested and returned model IDs, effort, tools, and permissions.
- Every attempt, including failures and assisted retries.
- Input, cache, output/reasoning, tool, and runtime charges.
- Active review and correction effort, without inventing an hourly value.
Use the complete Luna–Terra comparison and proposed protocol to structure the decision.
Reader questions
Frequently asked questions
Is GPT-5.6 Terra ten times more expensive than Luna?
At the standard like-for-like input, cached-input, cache-write, and output token rates checked on 12 September 2026, yes. Actual workflows may consume different quantities or use other service conditions.
Does Terra’s higher coding score prove the upgrade is worth it?
No. A provider-reported benchmark difference does not establish accepted repository outcomes, total workflow cost, or human effort in your harness.
Primary reading
Sources and update conditions
Rates, availability, model identity, and provider behavior can change. Recheck the linked sources before making a consequential purchase or sending repository data.