GPT-5.6 Is Now Live for Everyone: Plan-by-Plan Access, ultra Mode and the Day-One Routing Math

July 10, 2026. GPT-5.6 is no longer a permission slip. OpenAI moved Sol, Terra and Luna to general availability on July 9 across ChatGPT, Codex and the OpenAI API, with the rollout completing globally over roughly 24 hours. It is the family's first appearance in ChatGPT, exactly 13 days after the June 26 preview that launched behind a government gate. We covered the launch-day facts as they broke in yesterday's brief. Today, with the full GA announcement and API documentation public, the useful story for operators is what actually changed: who gets which model on which plan, a day-one API surface nobody outside the preview could document, and pricing mechanics that reward teams who re-tune rather than re-label.

What shipped at GA

1. General availability everywhere. OpenAI's announcement is unambiguous: GPT-5.6 is available starting July 9 across ChatGPT, Codex and the API, rolling out globally toward full availability over the next 24 hours. That resolves the preview era's regional ambiguity, and it puts the family in ChatGPT for the first time.
2. Pricing held flat from preview. Sol is $5 input and $30 output per 1M tokens, Terra $2.50 and $15, Luna $1 and $6. No GA premium, no repricing.
3. Official model IDs. The API slugs are gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna, and the bare gpt-5.6 alias routes to Sol.
4. A taller settings ladder. Reasoning effort now tops out at a new max level above xhigh, and a pro reasoning mode works on any GPT-5.6 model. What ChatGPT's Pro and Enterprise pickers label Sol Pro is that API setting, not a separate model.
5. ultra, a multi-agent setting. OpenAI's highest-capability configuration coordinates four agents in parallel by default, trading higher token use for stronger results and faster time-to-result. Developers get the same primitive through a multi-agent beta in the Responses API.
6. Programmatic Tool Calling. GPT-5.6 can write and run lightweight programs that coordinate tools, filter large intermediate outputs and return a compact result to the model, cutting round trips. It is compatible with Zero Data Retention accounts.
7. Caching changed shape. Explicit cache breakpoints arrive with a 30-minute minimum cache life. Cache writes are billed at 1.25x the uncached input rate, while reads keep the 90 percent discount.
8. Two sibling launches the same morning. OpenAI shipped ChatGPT Work, a new agentic workspace we cover in a separate brief, and Microsoft named GPT-5.6 the preferred model in Microsoft 365 Copilot.

Who gets what, plan by plan

In the standard Chat surface, Plus, Pro, Business and Enterprise users get Sol at medium and higher effort settings, and Pro and Enterprise add the Sol Pro option for the hardest tasks. OpenAI's access notes list no free-tier option in plain Chat. In ChatGPT Work and Codex the map is different: Free and Go users get Terra, every paid plan from Plus up can choose among Sol, Terra and Luna with a per-model effort setting, and the max toggle is available to everyone with GPT-5.6 access there. The ultra setting is the real plan differentiator: ChatGPT Work reserves it for Pro and Enterprise, while Codex offers it from Plus up. On the API there is no plan gating at all, all three models are self-serve for any account.

Read that map twice and a quiet signal appears: the cheapest guaranteed path to GPT-5.6 is not the chat window, it is OpenAI's new agentic surfaces. That is where OpenAI wants the next wave of usage to land.

The benchmark picture, read honestly

Every headline number so far is OpenAI-reported, with no independent audit yet. On OpenAI's own tables, Sol leads long-horizon agentic work with 52.7 percent on Agents' Last Exam (the announcement prose cites 53.6 from a maxed-out configuration, the table is the auditable figure), sets a new state of the art of 80 on the Artificial Analysis Coding Agent Index, 2.8 points above Claude Fable 5 while using less than half the output tokens and costing about one third less, and posts 62.6 percent on OSWorld 2.0 computer use with 85 percent fewer output tokens than Claude Opus 4.8. Sam Altman told CNBC that Sol is 54 percent more token efficient on coding tasks, and TechCrunch notes the marketing aims squarely at Anthropic.

The same tables refuse a victory lap. Claude Mythos 5 leads SWE-Bench Pro at 80.3 percent versus Sol's 64.6, Fable 5 keeps the GDPval Elo lead at 1,759.6 versus 1,747.8 and edges the Artificial Analysis Intelligence Index at 59.9 versus 58.9, and HealthBench Professional stays with Fable 5. OpenAI's real pitch is performance per dollar, not a sweep: within one point of Fable 5 on the Intelligence Index at roughly half the estimated cost and 61 percent less time. That is the same cost-curve war Sonnet 5 reopened last week, now fought from the other side.

What it means for operators

GA day is an access story, and the bill is decided by how you build. Six moves, in order:

1. Run migration as a tuning pass, not a slug swap. OpenAI's own guidance says to test your current reasoning effort and one level lower. GPT-5.6 is biased toward compression and reacts more strongly to brevity instructions, so re-test any prompt that carries hard length constraints before trusting the old behavior.
2. Price per finished task, then route. Point high-volume classification, tagging and extraction at Luna, make Terra the production default, and reserve Sol and Sol Pro for work that measurably needs the depth. Sticker price per token is the wrong unit of account.
3. Treat ultra as a budget decision. Four parallel agents buy wall-clock time and a synthesis step on decomposable, deadline work. On linear tasks where each step depends on the last, they mostly re-derive each other's context. Keep it off for routine throughput.
4. Redo the caching math. Writes now cost 1.25x uncached input, reads keep the 90 percent discount, and segments live at least 30 minutes. Long, stable system prompts win. Prompts that churn every call now pay a premium on every write.
5. Adopt Programmatic Tool Calling where control flow is predictable. The efficiency numbers OpenAI published are customer results from adopting the pattern, not from renaming a model: Clio cut prompt tokens 38 percent, PlayCo cut total tokens 63.5 percent, Rogo finished 28 percent faster. Exploratory agents that decide step by step should stay on direct tool calls.
6. Budget for the safety layer. Generation can pause mid-stream for several seconds while classifiers review output, and ChatGPT and Codex offer a retry on a lower-capability model when a safeguard adds friction. Benchmark those pauses before putting Sol behind latency-sensitive product UX.

The meta-lesson has not moved since the stack went usage-metered: stay model-agnostic, keep an abstraction layer, and re-run the routing math every time a release like this lands. If you want that math run on your actual workloads, with routing, caching and tool-calling architecture handled, that is exactly what our AI automation team and dedicated AI engineers do for clients every week.

The gate that set the template

The 13-day arc from gated preview to global GA is the first complete run of the voluntary pre-release review created by June's executive order, and it landed on its predicted date. The White House's formal framework has still not been announced, with reporting pointing to a window that closes around July 11, so the de facto rule for operators is already visible: capability day and access day are now different days, and eval work belongs in the gap between them. Our June 29 deep covers how the gate worked. One more calendar item hides in the safety notes: individuals who want the most cyber-capable access must enable hardware-backed passkeys by September 1 under the Trusted Access for Cyber program. If your agency runs security-adjacent automations, put that date in the ops calendar now.