Which of the six models fits your task?
| Model | Official role | Good starting point | Check before using it broadly |
|---|---|---|---|
| GPT-6 Astra | OpenAI's highest-capability tier | Cross-repository changes with unclear requirements or high repair cost | Whether fewer failed attempts justify the higher rate |
| GPT-6 Sol | Complex coding and agent work | Implementing a feature and checking tests across several files | Whether Luna passes the same checks with less usage |
| GPT-6 Luna | Focused, high-volume work | A specified edit or extraction job with an automated validator | Failures and human repair time on edge cases |
| GPT-5.6 Sol | Flagship GPT-5.6 tier | A working prompt-and-review pipeline already calibrated to Sol | Whether GPT-6 Sol reaches the same acceptance rate for less |
| GPT-5.6 Terra | Middle GPT-5.6 tier | An evaluated mid-tier route you cannot yet replace safely | GPT-6 Sol's lower published output rate |
| GPT-5.6 Luna | High-volume GPT-5.6 tier | An existing batch job with measured pass and retry rates | GPT-6 Luna's lower published rates |
The task suggestions follow OpenAI's workload guidance. For example, a one-file change with a failing-test check fits Luna better than an open-ended migration across services. If the Luna attempt misses a constraint and requires a Sol retry, count both attempts. The unsuffixed API alias gpt-5.6 points to GPT-5.6 Sol; use an explicit model name if you mean Terra or Luna.
What Codex usage costs in credits
Codex subscriptions use a shared allowance, and eligible usage beyond the included limit can use ChatGPT credits. OpenAI publishes these Standard-speed credit rates per 1 million tokens. Cached input is a separate rate for input that qualifies for a cache read.
| Model | Input credits | Cached input credits | Output credits | Plus local messages per 5 hours, estimate |
|---|---|---|---|---|
| GPT-6 Astra | 250 | 25 | 1,250 | 5-45 |
| GPT-6 Sol | 50 | 5 | 250 | 15-150 |
| GPT-6 Luna | 2.5 | 0.25 | 12.5 | 350-3,000 |
| GPT-5.6 Sol | 100 | 10 | 500 | 10-100 |
| GPT-5.6 Terra | 50 | 5 | 300 | 25-200 |
| GPT-5.6 Luna | 5 | 0.5 | 30 | 250-2,000 |
The message ranges are OpenAI's estimates for local work, not a fixed number of turns. A long repository session can use more of your allowance than a short request. Your current limit and reset time are shown in the usage dashboard, and Codex CLI users can check /status. Model choice, context length, reasoning, tool results, and caching all affect consumption. GPT-6 Fast mode, where available, uses 2.5 times the Standard Codex credit rate.
In the pictured example, every model receives 100,000 uncached input tokens and produces 10,000 output tokens. At Standard speed, the published rate formula is (input tokens / 1,000,000 × input rate) + (output tokens / 1,000,000 × output rate). It yields 37.5 credits for GPT-6 Astra, 7.5 for GPT-6 Sol, 0.375 for GPT-6 Luna, 15 for GPT-5.6 Sol, 8 for GPT-5.6 Terra, and 0.8 for GPT-5.6 Luna. Real runs can use different token amounts, even on the same task.
API token prices are a separate bill
If you use an API key in Codex CLI, SDK, or an application, API token pricing applies. The table shows Standard prices in US dollars per 1 million text tokens for requests with at most 272,000 input tokens. OpenAI lists the GPT-6 rates and separate pages for GPT-5.6 Sol, Terra, and Luna.
| Model | Input | Cached input | Cache write | Output |
|---|---|---|---|---|
| GPT-6 Astra | $10.00 | $1.00 | $12.50 | $50.00 |
| GPT-6 Sol | $2.00 | $0.20 | $2.50 | $10.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.125 | $0.50 |
| GPT-5.6 Sol | $4.00 | $0.40 | $5.00 | $20.00 |
| GPT-5.6 Terra | $2.00 | $0.20 | $2.50 | $12.00 |
| GPT-5.6 Luna | $0.20 | $0.02 | $0.25 | $1.20 |
A cache write is an input-token rate, not an additional fee added on top of the input rate. A later matching cache read uses the discounted cached-input rate. For all six models, a request with more than 272,000 input tokens moves to long-context pricing for the full request: input and cache rates double, while the output rate rises by 50%. Tool calls and other processing modes may add costs. GPT-5.6 Sol's published API price is promotional through at least November 21, 2026. Check the current rate card before committing a budget.
Why a large context window does not predict usage
Each of the six model pages lists a 1,050,000-token context window and a 128,000-token maximum output. Those are ceilings, not the tokens charged for every turn. Files, conversation history, tool definitions, and tool results can raise input usage. The visible answer and internal reasoning count toward output; OpenAI's reasoning guide says reasoning tokens are billed as output and appear in the API usage object's output_tokens_details. A long debugging session can therefore cost more than its final answer suggests. On an API route, inspect the usage object; in Codex, check the usage dashboard or /status rather than estimating from the reply length.
When is GPT-5.6 still worth keeping?
Keep a GPT-5.6 route when its prompts, validators, and review process already produce reliable accepted work. To test a replacement, give both models the same representative inputs and the same pass/fail checks. Record attempts, accepted outputs, input and output tokens, tool charges, elapsed time, and minutes of human repair. A lower list price helps only if it survives those checks; for a batch extraction pipeline, compare cost per accepted record rather than cost per raw response.
The closest API-price comparison is GPT-5.6 Terra against GPT-6 Sol: both cost $2 per million uncached input tokens, but Sol costs $10 rather than $12 per million output tokens. GPT-6 Sol is half the input and output price of GPT-5.6 Sol; GPT-6 Luna is half the input price and less than half the output price of GPT-5.6 Luna. These are equal-token comparisons, not proof of equal quality or equal total usage.
For a retained hands-on comparison within the newer family, see our Astra, Sol, and Luna Codex test. Each model completed four small tasks in one run. That test did not include the three GPT-5.6 models, long-context work, or a per-run credit debit, so it cannot rank all six models. For ways to reduce unnecessary context before selecting a more expensive tier, read the Codex usage planning guide.
