GPT-5.6 is generally available: Sol, Terra and Luna arrive in ChatGPT, Codex and API
OpenAI is broadly rolling out the GPT-5.6 family. Sol, Terra and Luna are coming to ChatGPT, Codex and API, with the biggest changes around agentic work, cost control and reasoning modes.
OpenAI has announced general availability for the GPT-5.6 family. This is the broader rollout that follows the limited GPT-5.6 Sol preview in June, when the company first showed the strongest variant of its new generation to selected partners. Now GPT-5.6 is moving into the products where people will actually use it: ChatGPT, Codex and the OpenAI API.
The launch was confirmed on July 9, 2026 in OpenAI's official announcement and the OpenAI API changelog. OpenAI is not shipping one model, but three: GPT-5.6 Sol, GPT-5.6 Terra and GPT-5.6 Luna. Sol is the flagship model, Terra is the balanced quality-and-cost option, and Luna is the fastest and most cost-efficient tier for higher-volume use.
The short version: GPT-5.6 is no longer only a capability preview. It is becoming a production model family that users, developers and teams can start working into their daily workflows.
GPT-5.6 at a glance
| Area | Detail |
|---|---|
| Company | OpenAI |
| GA announcement date | July 9, 2026 |
| Models | GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna |
| Availability | ChatGPT, ChatGPT Work, Codex, OpenAI API |
| API alias | gpt-5.6 points to gpt-5.6-sol |
| API surfaces | v1/responses, v1/chat/completions, v1/batch |
| Technical additions | Programmatic Tool Calling, explicit prompt caching, persisted reasoning, max, Pro mode, multi-agent beta |
| API pricing | Sol: $5 input / $30 output; Terra: $2.50 / $15; Luna: $1 / $6 per 1M tokens |
What changed after the preview
The June preview was a signal that OpenAI had a successor to GPT-5.5, but access was cautious and controlled. GPT-5.6 Sol was positioned mainly for selected partners, API usage and Codex, with OpenAI emphasizing additional safety testing before a wider rollout.
The new announcement changes the center of gravity. OpenAI now describes GPT-5.6 as a generally available family across its main products. The rollout is global and gradual, with full availability expected across eligible users over roughly 24 hours. Some accounts may see the new options earlier than others, but this is no longer a narrow preview for a small group.
The more important product decision is the three-tier structure:
- Sol is for the hardest tasks: coding, research, analysis and agentic workflows.
- Terra is for everyday work where teams want strong quality without paying for the maximum tier every time.
- Luna is for fast, high-volume workloads where the cost of each answer matters.
That is increasingly how the model market works. Teams do not want one model for everything. They want routing: cheaper models handling simple tasks, balanced models covering most work, and the flagship model reserved for problems where the extra capability justifies the cost.
Availability in ChatGPT and Codex
In ChatGPT, access depends on plan and reasoning mode. According to OpenAI, Plus, Pro, Business and Enterprise users get access to GPT-5.6 Sol through medium and higher reasoning settings. Pro and Enterprise users can also select GPT-5.6 Sol Pro for the most demanding work.
In ChatGPT Work and Codex, the structure is more flexible. Free and Go users get GPT-5.6 Terra, while Plus, Pro, Business and Enterprise users can choose between Sol, Terra and Luna and set the reasoning effort for each. The max setting is available to users with GPT-5.6 access in ChatGPT Work and Codex. ultra, OpenAI's mode based on parallel work by multiple agents, is available in ChatGPT Work for Pro and Enterprise and in Codex from Plus upward.
That matters for Codex in particular. GPT-5.6 is not just a model for more polished chat answers. OpenAI is presenting it as a model for longer tasks that involve tools, checks, iteration and work over real files. Codex is one of the places where those differences should show up faster than they do in a simple text prompt.
What developers get in the API
The API now exposes three concrete model IDs: gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna. The gpt-5.6 alias points to Sol, which is useful for testing the flagship model. In production, though, teams will probably want to choose the exact model ID so they can control cost and behavior.
The most interesting technical addition is Programmatic Tool Calling in the Responses API. In plain terms, the model can write and run lightweight in-memory programs that coordinate tools, filter intermediate results and decide what should remain in context. That can reduce the number of model-to-app round trips, especially in workflows with many tools and large raw outputs.
The second major change is better cache control. GPT-5.6 adds explicit cache breakpoints and a minimum 30-minute cache lifetime. OpenAI is also changing cache accounting: cache writes are billed at 1.25x the uncached input rate, while cache reads still receive a 90% discount. For applications with long system prompts, repeated workflows or large context blocks, this may matter more than the headline token price.
Persisted reasoning, the max level, Pro mode and multi-agent orchestration in the Responses API all point in the same direction. OpenAI wants GPT-5.6 to do more than answer. It wants the model to run longer work with better control over cost, state and parallel subtasks.
GPT-5.6 pricing
The API pricing makes the segmentation clear:
| Model | Input per 1M tokens | Output per 1M tokens | Natural role |
|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $30.00 | hardest tasks and highest quality |
| GPT-5.6 Terra | $2.50 | $15.00 | balance of quality and cost |
| GPT-5.6 Luna | $1.00 | $6.00 | fast, cheaper and high-volume workflows |
Sol is the premium option, but it should not automatically become the default for every workload. Terra may become the practical choice for many teams that want near-frontier quality without paying flagship rates for every response. Luna looks better suited for high-volume automation: classification, extraction, routing, simpler agents and embedded app assistance.
The smart strategy will be mixed. Move only the tasks that truly benefit from Sol onto the flagship model. Use Terra where it preserves enough quality at lower cost. Use Luna when scale and response economics matter more than maximum reasoning depth.
Benchmarks, safety and the real signal
OpenAI says GPT-5.6 improves coding, agentic work, cybersecurity, science and professional knowledge work. The company is also emphasizing performance per dollar rather than raw benchmark wins alone. That is an important shift: the question is not only whether a model is "best", but how much useful work it delivers for each token and each dollar.
In coding, OpenAI points to strong Sol results on coding-agent evaluations, Terminal-Bench and DeepSWE. In knowledge work, the company highlights better documents, spreadsheets, presentations and user interfaces. In cybersecurity, OpenAI says the model is more capable while also putting some of the most sensitive defensive capabilities behind Trusted Access.
That last part is worth watching. GPT-5.6 is more capable in dual-use areas, but OpenAI is describing a layered safety approach: real-time checks, monitoring, a reasoning monitor and more restricted access for the highest-risk capabilities. The company says it conducted a broad red-team program before GA, including large-scale automated testing.
What it means for readers
For everyday ChatGPT users, the main point is simple: GPT-5.6 is entering the product, but access and visible model options depend on plan, region and rollout timing. Not everyone will see the same set of choices at the same time.
For Codex users, this may be one of the more important OpenAI launches of the year. If Sol, max and ultra work as described, the difference should be clearest in long-running tasks: fewer stalls, better tool use, more effective self-correction and a higher chance of moving from a request to a finished change.
For companies and developers, three practical questions matter most:
- Which tasks actually need Sol?
- Can Terra replace GPT-5.5 in everyday workflows?
- Can Luna lower the cost of high-volume automation without a visible drop in quality?
That is where the real value of GPT-5.6 will be tested. Not in one viral demo, but in the economics of daily work: token use, response time, failed attempts and whether the model can carry a longer process with less supervision.
Our take
General availability for the GPT-5.6 family is a bigger product moment than the earlier Sol preview. The preview showed the direction. GA turns it into something teams can actually plan around.
The important change is not simply that OpenAI has another strong model on a leaderboard. It is the full package: three cost tiers, new reasoning modes, better tool use, multi-agent work in the API and real availability across ChatGPT and Codex. This looks less like a model built for one impressive answer and more like a model family built for longer work that needs to end in a usable result.
If earlier model launches asked, "Does AI answer better?", GPT-5.6 pushes the question toward: "Can AI ship more useful work, faster and with less supervision?" That is why Sol matters - and why Terra and Luna may end up being just as important.
