Like-for-like comparison
| Property | Luna | Terra |
|---|---|---|
| API ID | gpt-5.6-luna | gpt-5.6-terra |
| Context / output ceiling | 1,050,000 / 128,000 | 1,050,000 / 128,000 |
| Default reasoning effort | Medium | Medium |
| Input / output per 1M | $0.20 / $1.20 | $2.00 / $12.00 |
| Cached input / cache write | $0.02 / $0.25 | $0.20 / $2.50 |
| Above 272K input | 2× input, 1.5× output | 2× input, 1.5× output |
The premium in dollars
| Illustrative workload | Luna | Terra |
|---|---|---|
| 12K input + 2K output | $0.0048 | $0.0480 |
| 80K input + 10K output | $0.0280 | $0.2800 |
| Five 50K-input/4K-output calls, uncached | $0.0740 | $0.7400 |
Illustrative billable quantities, standard short-context rates, no cache. These are not measured agent runs and exclude tools, runtime, retries, and people.
What public evidence can—and cannot—say
OpenAI reports Terra somewhat above Luna on multiple coding-related evaluations and a larger separation on a long-context retrieval evaluation. The direction supports testing Terra on tasks with dispersed constraints. It does not establish a repository-size threshold, statistical significance for your work, or minutes saved in review.
Do not average unlike benchmarks into a homemade “coding ability” score or divide illustrative token cost by a provider score to manufacture cost per successful task.
Run the upgrade test
- Choose a repeatable failure from a real bounded task.
- Confirm the environment, retrieval, tools, and acceptance tests work.
- Compare Luna at the current and a higher effort with Terra at a declared effort.
- Start every run from the same commit and preserve failures.
- Measure accepted output, complete billed usage, interventions, and active review/correction time.
See the proposed protocol or calculate the billing baskets in the cost analysis.
Reader questions
Frequently asked questions
Is Terra worth ten times Luna for coding?
The token-rate premium is 10× under like-for-like standard conditions, but the payoff is task-specific. It is worth retaining only where matched runs show accepted work or measured reduction in rework that matters to you.
Should I increase Luna reasoning effort before switching to Terra?
Yes, include a higher-effort Luna configuration so the experiment does not confuse model identity with inference effort. Record settings and billed output for every run.
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.