OpenAI’s current API ladder is led by GPT-5.6 Luna, Terra, and Sol. The models share the same large context window and tool-use stack but target different economics: Luna is the volume route, Terra balances capability and cost, and Sol is the maximum-intelligence tier.
The live chart, calculator, and pricing table above are generated from pricing.json at build time. For the complete provider table, use the OpenAI pricing page; for cross-provider shopping, open the AI API pricing comparison.
How the GPT-5.6 tiers differ
| Tier | Start here when | Escalate when |
|---|---|---|
| GPT-5.6 Luna | The task is defined, repeatable, high-volume, or easy to verify | Quality misses create retries or human correction |
| GPT-5.6 Terra | The workflow needs more judgment, analysis, or reliable tool use | Errors are costly or the task remains ambiguous |
| GPT-5.6 Sol | Maximum reasoning, coding, cybersecurity, or agent capability matters | Use Fast mode only if latency changes the outcome |
OpenAI cut Luna’s standard rates by 80% and Terra’s by 20% on July 30, 2026; Sol’s standard rate did not change. Read the event-level analysis in OpenAI’s GPT-5.6 price cut: impact and what it means.
All three tiers support a 1.05-million-token context window and the GPT-5.6 tool-use stack. Do not treat equal context limits as equal model value: a cheaper tier can be the better route when the prompt is clear and the result is easy to check.
Hidden OpenAI API costs
Cache writes and reads are billed differently. GPT-5.6 cache reads receive a large discount, while writes cost more than normal input. Stable system instructions, tool schemas, and reference material should sit in a deterministic prefix so later calls reuse the cached region.
Long context changes the rate. When GPT-5.6 input crosses OpenAI’s long-context threshold, the full request receives higher input and output rates. Retrieval, summarization, and compaction can be cheaper than repeatedly sending an entire repository or document set.
Fast mode is a latency premium. It replaces Priority Processing and charges more than Standard without changing intelligence. Existing priority requests remain compatible, but teams should monitor whether an automatic migration changes spend.
Reasoning and tool loops affect completed-task cost. A low per-token model can still be expensive if it retries, emits long reasoning traces, or makes unnecessary tool calls. Track total tokens, tool calls, retries, latency, and accepted results together.
Regional processing can add an uplift. Eligible data-residency endpoints may cost more than the global route. AWS-hosted OpenAI models can also differ from direct API pricing, so verify the billable platform rather than assuming one universal rate.
Best OpenAI model by use case
| Workload | First model to test | Why |
|---|---|---|
| Classification, extraction, routing | GPT-4.1 nano or GPT-5.6 Luna | Cheap, easy-to-evaluate utility work |
| Customer support and RAG | GPT-5.6 Luna | Strong volume economics with escalation available |
| Data analysis and multi-step automation | GPT-5.6 Terra | Balanced capability for production workflows |
| Complex coding and agent planning | GPT-5.6 Sol | Maximum capability can reduce failed iterations |
| Offline bulk processing | Luna or Terra with Batch | Latency is less valuable than throughput savings |
| Latency-critical premium calls | Sol with Fast mode | Use only where response time has measurable value |
Do not pick a default from a benchmark alone. Build a small evaluation set from real production tasks, define an accepted-result rubric, and compare cost per accepted task. A router often wins: start cheap, detect uncertainty or failure, then escalate.
How to reduce OpenAI API spend
- Route by difficulty. Send routine calls to Luna or a nano model and reserve Sol for the hardest fraction of traffic.
- Use Batch or Flex for delay-tolerant work. Both can materially lower token rates when immediate responses are unnecessary.
- Keep prefixes cacheable. Avoid timestamps, random IDs, or reordered tool definitions inside otherwise stable context.
- Cap outputs. Output is usually the more expensive side of the request, and verbose agent loops compound quickly.
- Trim tool results. Return only the log lines, records, or document passages the next step needs.
- Measure accepted-task cost. Include retries, human correction, latency, and downstream failures—not only token volume.
For workload math, enter your actual prompt, cache, and response volumes in the AI token calculator.
Open-model quote to compare: Before standardizing on one provider, price the same evaluation set through Novita, an OpenAI-compatible managed route for DeepSeek, Llama, Qwen, image, and video models.
Free tier, subscriptions, and API billing
OpenAI does not advertise an ongoing free production API tier. Promotional credits may appear for new or eligible accounts, but applications should be budgeted as paid usage.
ChatGPT and Codex subscriptions do not replace API billing. Subscription prices and quota budgets are separate from the token-metered developer API. After the July price cut, Terra and Luna consume fewer paid subscription credits, but the subscription sticker prices did not change.
Bottom line
GPT-5.6 Luna is the first model to test for defined, high-volume work. Terra is the middle route for production tasks that need more judgment. Sol belongs on the hardest calls, and Fast mode belongs only where latency creates measurable value.
The best OpenAI architecture is usually a routing stack, not one model everywhere. Keep prompts cacheable, move delay-tolerant jobs to discounted processing, cap outputs, and judge every tier on cost per accepted result.
Sources: OpenAI’s official API pricing documentation, GPT-5.6 price-performance announcement, and AI Pricing Guru’s live pricing dataset.