In this article
Quick summary
OpenAI has now published new standard API GPT-5.6 pricing for GPT-5.6 Terra and GPT-5.6 Luna. Compared with the preview-era rate card, Terra is down 20% and Luna is down 80% on both input and output tokens.
| Model | Input / 1M tokens | Cached input / 1M | Output / 1M tokens |
|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $0.50 | $30.00 |
| GPT-5.6 Terra | $2.00 | $0.20 | $12.00 |
| GPT-5.6 Luna | $0.20 | $0.02 | $1.20 |
| GPT-5.5 | $5.00 | $0.50 | $30.00 |
The table above shows standard, short context pricing. For long context, Terra is $4 input / $18 output and Luna is $0.40 input / $1.80 output per 1M tokens.
This is a meaningful change for teams putting AI agents into production: Terra is now roughly 60% cheaper than GPT-5.5, and Luna roughly 96% cheaper on a like-for-like input or output rate.
More than "a model price cut"
At preview, OpenAI listed Terra at $2.50 input / $15 output and Luna at $1 input / $6 output per 1M tokens. The current API pricing page shows:
| Model | Preview price | Current standard price | Reduction |
|---|---|---|---|
| Terra – input | $2.50 | $2.00 | 20% |
| Terra – output | $15.00 | $12.00 | 20% |
| Luna – input | $1.00 | $0.20 | 80% |
| Luna – output | $6.00 | $1.20 | 80% |
The strategic signal is cost-performance: OpenAI is positioning Luna as the ultra-cheap option for high-volume workloads, Terra as the balanced tier for daily agent work, and Sol for hard reasoning and escalation.
If every task goes through the flagship model, your product will soon ship a new feature: real-time invoice streaming.
Pick the model by workflow, not by feeling
Luna: high-volume, deterministic work
Use Luna for well-structured tasks that need speed and volume:
- Classification, intent routing, tagging.
- Data extraction against a JSON schema.
- Content normalization, short summaries, first drafts.
- FAQ first-pass and pre-processing before escalation.
At $0.20 input and $1.20 output per 1M tokens, Luna is the right tool for cutting cost at the "bottom layer" — where call volume tends to grow fastest as a system scales. Mapping those layers explicitly helps; see our five-layer map for AI workflow automation.
Terra: the worker agent for everyday work
Terra fits workflows that need better reasoning or coding quality but do not yet justify the flagship:
- Document analysis, technical writing, code review.
- Ticket-handling agents, multi-step task planning.
- Drafting proposals, comparing options, supporting internal operations.
At $2 input / $12 output per 1M tokens, Terra is 60% cheaper than GPT-5.5 or Sol — but it still needs to be evaluated against your own real-world dataset.
Sol: only when escalation is worth it
Sol suits hard reasoning, long-horizon agentic work, and controlled final decision support. Being more expensive is not the problem; using Sol for everything is.
Cost model: the gap is bigger than you think
Assume a workflow handling 10,000 requests per month, each averaging 2,000 input tokens and 500 output tokens.
| Model | Estimated monthly cost |
|---|---|
| Luna | $10 |
| Terra | $100 |
| Sol | $250 |
The math:
- Total input: 20M tokens/month.
- Total output: 5M tokens/month.
- Luna, for example: 20 × $0.20 + 5 × $1.20 = $10.
These figures exclude long context, tool calls, retries, and reasoning tokens. But they are enough to show that model routing can create a 25× cost spread between Luna and Sol on the same hypothetical workload. Deciding which workflows deserve which tier is the same exercise as scoring processes to prioritize for AI automation.
Prompt caching: a FinOps lever you have to measure
The current rate card also breaks out cached input and cache writes separately:
- Terra: cached input $0.20, cache write $2.50 / 1M tokens.
- Luna: cached input $0.02, cache write $0.25 / 1M tokens.
- Sol: cached input $0.50, cache write $6.25 / 1M tokens.
With long system prompts, repeated tool schemas, or context shared across requests, cache-hit rate belongs in your observability dashboard — alongside token usage, latency, retry rate, and outcome.
Implementation recommendations
Luna → classify / extract / route / draft
Terra → analyse / code / daily multi-step work
Sol → complex reasoning / high-risk escalation / final decision support
Do not route on intuition. Design a policy you can measure:
- Define task classes and quality SLOs.
- Set a token budget, retry cap, and escalation threshold for each class.
- Track cost by workflow and business outcome, not just by model.
- Re-evaluate against an internal test set before changing the default model. The discipline here is the same one covered in measuring, piloting, and governing AI projects.
Conclusion
The Terra repricing — and especially Luna's — makes multi-model architectures far more practical for teams building AI agents. The advantage is not in using the cheapest model; it is in using the right model for the right unit of work, then controlling cost through routing, caching, and observability.
