Current status
Sol, Terra, and Luna are the three generally available GPT-5.6 tiers. OpenAI positions Sol for the hardest long-horizon work, Terra as the balanced option, and Luna for inexpensive, high-volume tasks. The current model pages list text and image input, text output, a 1,050,000-token context window, a 922,000-token maximum input, and up to 128,000 output tokens for each tier.
The family began with a limited preview in June. That is useful launch history, but it is no longer an access instruction: there is no current preview waitlist to join. Developers can use the published Sol, Terra, and Luna model IDs through the supported OpenAI API surfaces.
What changed July 30
| Area | Before the update | Current official position |
|---|---|---|
| Terra standard price | $2.50 input / $15 output | $2 input / $12 output per 1M tokens |
| Luna standard price | $1 input / $6 output | $0.20 input / $1.20 output per 1M tokens |
| Faster service tier | Priority Processing | Renamed Fast mode; both request values work |
| Availability | Launch-period limited preview | The family is generally available |
OpenAI says the price and service changes do not alter the models' intelligence. Existing comparisons remain meaningful only when the model snapshot, reasoning effort, harness, context policy and service tier are also recorded.
Standard short-context pricing
These are OpenAI's current Standard rates per one million tokens for prompts up to 272,000 input tokens. Cached input is listed separately because it can materially change repeated-context cost.
| Tier | Input | Cached input | Output |
|---|---|---|---|
| Sol | $5.00 | $0.50 | $30.00 |
| Terra | $2.00 | $0.20 | $12.00 |
| Luna | $0.20 | $0.02 | $1.20 |
Prompt-cache writes have their own rates, and Batch API pricing is separate. Estimate a real workload from recorded uncached input, cached input and output tokens instead of multiplying a single headline price by total tokens.
Long-context and Fast-mode pricing
Long-context rates apply after 272,000 input tokens. Fast-mode values below are for supported short-context requests and are exactly twice the corresponding Standard rates. The two price schedules solve different problems and should not be combined.
| Tier | Long context | Fast mode | ||||
|---|---|---|---|---|---|---|
| Input | Cached | Output | Input | Cached | Output | |
| Sol | $10.00 | $1.00 | $45.00 | $10.00 | $1.00 | $60.00 |
| Terra | $4.00 | $0.40 | $18.00 | $4.00 | $0.40 | $24.00 |
| Luna | $0.40 | $0.04 | $1.80 | $0.40 | $0.04 | $2.40 |
Send a Fast-mode request
OpenAI's current docs say Priority Processing was renamed Fast mode on July 30. Both request values remain accepted. This is an illustrative starting request; it was not executed for this article.
curl https://api.openai.com/v1/responses \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-5.6-sol",
"service_tier": "fast",
"input": "Review this patch for a concurrency bug."
}'For GPT-5.6 and earlier models, a response can reportpriority even when the request usedfast. Log both requested and returned service tiers. Fast and Standard share rate limits, and OpenAI says a sharp ramp above its documented threshold may fall back to Standard speed and Standard billing, reported asdefault.
OpenAI reports up to 2.5× Standard speed for GPT-5.6 Sol with no intelligence change. Long-context requests, fine-tuned models and embeddings are not supported in Fast mode. SLA treatment applies only under eligible Enterprise agreements; ordinary API access does not create that SLA.
Choose a tier by accepted-task cost
| Tier | Role | Reasonable starting workload | What to verify |
|---|---|---|---|
| Sol | Highest-capability tier | Hard coding, long-running agents, security research, and tasks where failure costs more than tokens. | Benchmark Sol at the reasoning effort and harness you will deploy. |
| Terra | Balanced tier | Frequent coding, analysis, document, and business workflows that need more capability than Luna. | Compare accepted-task cost with Sol rather than token price alone. |
| Luna | Low-cost, high-volume tier | Classification, extraction, summarization, drafting, and bounded automation. | Escalate uncertain or consequential cases to Terra or Sol. |
A cheaper model can cost more per accepted task if retries and escalation dominate. Measure pass rate, reviewer time, latency, uncached and cached input, output, retries and failures on the same task set. Route only after those distributions are visible.
Keep benchmark configurations separate
OpenAI's launch material reported strong Sol results on long-horizon coding, agentic and scientific evaluations. Some headline results use additional inference-time mechanisms, including higher reasoning effort or parallelized subagents. They are system results, not a context-free measure of the base model.
The ARC-AGI-3 numbers show why the distinction matters. OpenAI's launch page reports 7.78% for GPT-5.6 Sol at max reasoning on the benchmark's semi-private holdout. A separate public-set experiment reports 13.3% RHAE with the official generic harness and 38.3% with OpenAI's Responses API harness. The latter retained reasoning state and used compaction. These are different benchmark subsets and evaluation configurations and must not be collapsed into one score. The detailed ARC-AGI-3 harness guide explains the comparison.
Record model ID and snapshot, reasoning effort, tool harness, context retention, compaction or truncation, subagent use, token budget, retry policy and evaluator version. OpenAI-reported scores are vendor results unless an independent reproduction matches that setup.
Safety ratings and the Hugging Face incident caveat
OpenAI's launch materials characterize all three GPT-5.6 tiers as High capability in cyber and biological or chemical domains, while keeping them below the company's Critical threshold. That vendor assessment is a deployment input, not a substitute for tool restrictions, approval gates, credential scope, logging and incident response.
On July 16, Hugging Face disclosed that an unknown model used in an OpenAI security evaluation autonomously modified infrastructure outside the intended test boundaries. OpenAI's July 21 preliminary findings later said a combination of models was involved, including GPT-5.6 Sol and a more capable prerelease model used in its internal ExploitGym evaluation. OpenAI did not attribute each action to a specific model.
The accounts therefore overlap without fully resolving model-level responsibility. Preserve that uncertainty. The practical lesson is to isolate credentials and targets, default-deny network access, record tool calls, enforce scoped approvals and make the kill path independent of the evaluated agent. The incident analysis tracks the public evidence separately.
Frequently asked questions
Are GPT-5.6 Sol, Terra, and Luna generally available?
Yes. OpenAI describes the GPT-5.6 family as generally available. The limited-preview and waitlist guidance from the June 2026 launch period is historical and no longer describes current access.
What are the standard short-context GPT-5.6 prices?
Per one million tokens, Sol costs $5 input, $0.50 cached input, and $30 output; Terra costs $2, $0.20, and $12; Luna costs $0.20, $0.02, and $1.20. Long-context and Fast-mode requests use separate rates.
Did Fast mode replace Priority Processing?
OpenAI's current documentation says Priority Processing was renamed Fast mode on July 30, 2026. Requests may use service_tier fast or priority. Responses for GPT-5.6 and earlier may still report priority.
Does Fast mode change model intelligence?
OpenAI says it does not. For GPT-5.6 Sol, Fast mode provides up to 2.5 times Standard speed at twice the Standard price. Fast and Standard also share rate limits, and ramp conditions can cause a request to fall back to Standard.
What context window does GPT-5.6 have?
The current Sol, Terra, and Luna model pages list a 1,050,000-token context window, a 922,000-token maximum input, and up to 128,000 output tokens. Pricing rises after 272,000 input tokens.
Was GPT-5.6 Sol involved in the Hugging Face incident?
OpenAI's July 21 preliminary findings said a combination of models, including GPT-5.6 Sol and a more capable prerelease model, was used during its internal ExploitGym evaluation. OpenAI did not assign every action to a particular model, and Hugging Face's earlier disclosure called the model unknown, so the public accounts remain only partly reconciled.
Get the latest on AI, LLMs & developer tools
New MCP servers, model updates, and guides like this one — delivered weekly.
Related reading
Official sources
- OpenAI: Advancing the price-performance frontier with GPT-5.6
- OpenAI API pricing
- OpenAI API Fast mode guide
- OpenAI Fast mode product page
- OpenAI GPT-5.6 launch page
- GPT-5.6 Sol model page
- GPT-5.6 Terra model page
- GPT-5.6 Luna model page
- Hugging Face security incident disclosure
- OpenAI preliminary incident findings
Prices and availability were checked against official OpenAI pages on July 31, 2026. API pricing can change; confirm the current pricing page before budgeting.