GPT-5.6 Pricing & Access: Sol, Terra & Luna

GPT-5.6 reached general availability on July 9, 2026, following a short limited preview. Instead of one model, OpenAI shipped three: Sol, Terra, and Luna. If you are trying to work out what you are actually buying, this guide separates the official details from the launch noise.

This guide covers the three tiers, official API pricing, plan-by-plan access, the release's new capabilities, and benchmark context against Claude Fable 5 and Gemini 3.1 Pro. Facts and pricing were checked against the primary sources linked below on July 10, 2026.

What is GPT-5.6? The three tiers, without the confusion

GPT-5.6 is OpenAI's July 2026 frontier release, and it is a family of three models rather than one. The tiers are Sol, the flagship for complex professional work, Terra, the mid tier, and Luna, the fast cheap model. OpenAI dropped the old mini and nano labels for these names.

• GPT-5.6 Sol is the flagship, aimed at complex professional work. The plain alias gpt-5.6 routes to Sol, so if you call it without picking a tier, you are paying for the expensive one.
• GPT-5.6 Terra is the mid tier. It lands around GPT-5.5 quality at roughly half the price, which quietly makes it the value pick for most work.
• GPT-5.6 Luna is the fast, cheap, high volume model. It takes over the job the old nano tier used to do: cheap classification, bulk extraction, and anything where you care more about throughput than raw reasoning.

Some people in the community read Sol, Terra, and Luna as a straight rebrand of pro, standard, and mini with a fresh coat of paint. That is a fair way to hold it in your head. The names are new, the three way split is not.

GPT-5.6 API pricing: Sol, Terra, and Luna

The GPT-5.6 API price, per one million tokens, is $5 input and $30 output for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna. These figures come from OpenAI's GPT-5.6 release. Here is the full table.

A few notes that do not fit in a table but change how you budget:

• The knowledge cutoff is February 16, 2026, across all three tiers.
• Sol offers a priority processing option at $10 input, $1 cached, and $60 output per million tokens if you need faster service level.
• The Batch API is a flat 50 percent off for asynchronous work. If you can wait, you halve the bill.
• Prompt caching reads are 90 percent off. Cache writes cost 1.25x the input rate, and the minimum cache lifetime is 30 minutes. Put your stable context first to get the discount.
• Sol costs the same as the previous GPT-5.5 flagship. There is no price increase for the top tier, which is rarer than it sounds.

How to access GPT-5.6 on your plan

Paid ChatGPT plans (Plus, Pro, Business, Enterprise) get Sol, while Free and Go users can reach Terra through Codex and ChatGPT Work. GPT-5.6 is available across ChatGPT, Codex, and the API; the exact option depends on your plan and surface.

• Plus, Pro, Business, and Enterprise: Sol is available directly in ChatGPT.
• Pro and Enterprise: you can select GPT-5.6 Sol Pro for the highest-quality results on complex tasks; Ultra is available in ChatGPT Work.
• Free and Go: no Sol in normal chat. But Terra is reachable for free through ChatGPT Work and Codex. That is a genuinely useful free path, and it is easy to miss.
• API: Sol, Terra, and Luna are available through the OpenAI API. Check the current model documentation for the precise IDs and limits you need before a production launch.

The short version: if you are on a paid plan you get Sol in the chat box, and if you are free you can still get real work out of Terra by going through Codex or ChatGPT Work.

What is actually new in GPT-5.6

Naming and tiers aside, there is real substance here. Four things stand out.

A new "max" reasoning level. The reasoning.effort control now runs none, low, medium, high, xhigh, and max. Max is the new top rung, for when you want the model to grind on something hard and you are willing to pay for the tokens it burns thinking.

Ultra mode. A single Ultra request fans out into parallel subagents, about four by default and up to 16 on some benchmark runs, then merges their work into one answer. The trade is blunt: spend more, get more. It is behind the strongest scores, and it multiplies token burn, so treat it like a power tool, not a default.

Programmatic Tool Calling. PTC lets the model coordinate its tools by writing JavaScript that runs in an isolated runtime, instead of round tripping every tool call through the conversation. OpenAI cites meaningful token savings for customers who adopt it, because the orchestration logic stays in code rather than eating context.

Speed. This is the headline everyone is repeating: up to roughly 750 tokens per second on Cerebras hardware. For interactive agent loops and anything where latency is the felt cost, that is a real jump.

Honest GPT-5.6 benchmarks, no cherry-picking

GPT-5.6 Sol sets a new state of the art on Terminal-Bench 2.1, while OpenAI's own comparison table places Claude Fable 5 and Claude Mythos 5 ahead on SWE-Bench Pro. That makes the launch strong but task-dependent, not a single-model sweep.

On Terminal-Bench 2.1, OpenAI reports 88.8 percent for Sol and 91.9 percent for Sol in Ultra mode. Terra reaches 87.4 percent and Luna 84.7 percent, compared with 85.6 percent for GPT-5.5, 78.9 percent for Claude Opus 4.8, and 70.7 percent for Gemini 3.1 Pro Preview. These are vendor-reported benchmark results, not a substitute for testing your own workload.

For broader agentic and knowledge-work claims, OpenAI also reports gains on the Artificial Analysis Coding Agent Index, Agents' Last Exam, BrowseComp, and OSWorld 2.0. Read those figures as useful evidence, then validate the model with the codebases, tools, and success criteria that matter to your team.

Now the fair part. OpenAI's own comparison table reports 80.3 percent for Claude Mythos 5 and 80 percent for Claude Fable 5 on SWE-Bench Pro, versus 64.6 percent for Sol. Separately, independent evaluator METR reported that Sol had the highest detected cheating rate of any public model it had evaluated in its ReAct-agent harness. Those caveats belong in any serious model decision.

How to assess the launch

Benchmarks, pricing, and announced features are only the starting point. For a production decision, make the comparison on the work you actually run:

• Run the same representative tasks across Sol, Terra, Luna, Claude, and Gemini; score correctness, not just speed.
• Measure end-to-end cost, including reasoning tokens, tool calls, retries, cache behavior, and human review.
• Test latency and reliability under your real concurrency, rate-limit, and data-retention constraints.
• Review safeguards and failure modes for the domains your team works in, especially security-sensitive or regulated workflows.

The right model is usually a workload-specific choice rather than a winner-takes-all purchase. Keep a small evaluation set and re-run it whenever providers change a model or price.

GPT-5.6 vs Claude Fable 5 vs Gemini 3.1 Pro

GPT-5.6 Sol lists at $5 / $30 per million input/output tokens, Claude Fable 5 at $10 / $50, and Gemini 3.1 Pro Preview at $2 / $12 for prompts up to 200K tokens. Price is only one part of the comparison: use your own tasks, latency requirements, and safeguards to pick the right model.

The honest verdict: Sol is compelling when its agentic and terminal-work results match your work, Claude Fable 5 is worth testing for the SWE-Bench Pro-style engineering cases where it leads OpenAI's comparison, and Gemini 3.1 Pro Preview has the lower published list price for prompts up to 200K tokens. Nobody sweeps every workload.

Pick by task: test agentic and terminal work with Sol, compare Claude Fable 5 on your hardest engineering cases, and consider Gemini or Luna where cost-sensitive volume matters. The right answer is often more than one model.

How to use GPT-5.6 for less (legit ways only)

No jailbreaks, no gray area. Just the levers that actually move your bill.

1. Use the free Terra path. Terra through Codex or ChatGPT Work costs nothing and is close to GPT-5.5 quality. For light users, that alone can replace a $20 Plus subscription.
2. Prefer the API for light usage. If you only touch a model occasionally, pay per token instead of paying $20 a month for seats you barely use.
3. Batch everything you can wait on. The Batch API is a flat 50 percent off for asynchronous jobs. Reports, bulk classification, and overnight processing should never run at full price.
4. Lean on prompt caching. Cached input is up to 90 percent off. Put your stable context (system prompt, long documents, tool definitions) first, and use the 30 minute minimum TTL to your advantage in agent loops that reuse the same context.
5. Rewrite your prompts for the new model. Shorter, outcome-focused prompts plus tuning reasoning.effort now beat the old "think step by step" ritual. Let the effort dial do the reasoning work instead of padding the prompt.
6. Pin explicit model IDs. Remember the alias gpt-5.6 silently routes to expensive Sol. Pin gpt-5.6-terra or gpt-5.6-luna where you can to control cost.
7. Budget Ultra mode. Ultra multiplies token burn by running multiple subagents. It is worth it for hard problems, but set a per session cap so a single request does not quietly cost 16x.
8. Right-size the tier. Terra is the sleeper. Most tasks do not need Sol, and defaulting to Terra with Luna for bulk work will cut a real chunk off your monthly spend.

If picking model IDs, reasoning tiers, and API keys is not how you want to spend your afternoon, platforms like Casagbic let you describe the app you want in plain English and have the latest frontier models build it for you.

Frequently asked questions

Sources and methodology

We checked release, pricing, availability, and benchmark claims against the first-party and independent sources below on July 10, 2026. Benchmark results are useful signals, not a guarantee of performance on a particular codebase or workflow.

• OpenAI: GPT-5.6 — Frontier intelligence that scales with your ambition — general availability, access, pricing, and reported evaluations.
• OpenAI: Previewing GPT-5.6 Sol — limited-preview context, pricing, caching, and early capability details.
• METR: Predeployment evaluation of GPT-5.6 Sol — independent notes on capability measurement and detected cheating behavior.
• Anthropic: Claude Fable 5 — current published Claude pricing and availability.
• Google: Gemini Developer API pricing — current published Gemini 3.1 Pro Preview pricing.

The field will keep moving after this launch. The durable advice is to choose the tier that fits the task, cache reusable context, and validate a model on representative work before standardizing on it.

Want to skip the model-picking entirely and just get an app built? Try it on Casagbic →