GPT-5.6 Goes Public: GA Pricing, Ultra Mode and Access

GPT-5.6 went to general availability on July 9, 2026 — across ChatGPT, Codex, and the OpenAI API, rolling out globally over roughly 24 hours. The model family a handful of trusted partners had been running quietly since the June 26 preview is now something anyone can select in a picker or call with an API key, and this is GPT-5.6’s first appearance in ChatGPT at all.

The temptation on GA day is to treat this as a benchmark story. It mostly isn’t. Pricing is unchanged from preview, the headline evals were published with the same announcement cycle, and on OpenAI’s own tables Claude models still lead several rows. What actually changed today is access — plus a day-one API surface that preview coverage couldn’t document: official gpt-5.6-* model IDs, a new max reasoning effort, a pro mode, a multi-agent beta, and Programmatic Tool Calling.

This post covers what shipped today and what to do about it: the confirmed API surface, the ultra setting and its cost profile, where the efficiency numbers really come from, exactly which plan gets which model, and an honest reading of the benchmark tables. For the tier-family explainer, see our GPT-5.6 preview guide from June — none of that background is repeated here.

OpenAI’s announcement leads with the operative sentence: “We’re launching the GPT-5.6 family of models for general availability following our limited preview.” The follow-through matters just as much — “GPT-5.6 is available starting today across ChatGPT, Codex, and the OpenAI API. The rollout is starting globally now and will continue gradually toward full availability over the next 24 hours.” Global from day one, which resolves the preview-era ambiguity about regional availability.

The arc from preview to GA took thirteen days. The June 26 preview was API and Codex only, invite-only for trusted partners, with no waitlist. OpenAI publicly committed to the July 9 date in a July 7 social post, and the GA page landed this morning. If you skipped the preview cycle entirely, the practical difference is simple: you no longer need anyone’s permission to use this model family.

One projection worth making explicit: this preview-then-GA shape — a government-reviewed capability window, a trusted-partner phase, then a 24-hour global release — looks like the template for how frontier launches will land from now on. Expect “capability day” and “access day” to be different days, and plan eval work for the gap between them rather than scrambling on GA morning.

Until this morning, nobody outside the preview program could name the model IDs. Now they’re documented: the bare gpt-5.6 alias routes to gpt-5.6-sol, and the explicit IDs are gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna. All three are self-serve in the API from today.

The settings ladder grew at both ends. reasoning.effort now supports none, low, medium, high, xhigh, and a new max — with OpenAI advising teams already on xhigh to compare both. Pro mode is a setting, not a model: reasoning.mode: "pro" works on any GPT-5.6 model and defaults to medium effort if unset. What ChatGPT’s Pro and Enterprise pickers label “GPT-5.6 Sol Pro” is exactly this API setting — there is no separate Pro model slug. And reasoning.context introduces persisted reasoning: reasoning items reused across turns for multi-turn quality and cache efficiency.

Caching changes shape too. Explicit cache breakpoints replace the old behavior, configured via prompt_cache_options.mode: "explicit" and a ttl field that replaces prompt_cache_retention. The commercial terms:

The migration note buried in OpenAI’s docs deserves more attention than it will get: treat migration as a tuning pass, not only a model-slug change — test the same effort setting and one level lower. GPT-5.6 is biased toward compression, and it’s more sensitive to “be concise” instructions than its predecessors: a brevity nudge can change how the model prioritizes the task itself, not just trim filler. If your prompts carry aggressive length constraints, re-test them before assuming the behavior carried over.

The genuinely new capability class at GA is ultra — OpenAI’s highest-capability setting, described as “coordinating multiple agents across parallel workstreams.” By default it runs four agents in parallel; a few GA charts also show 16-agent configurations on specific evals like BrowseComp and SEC-Bench Pro, but four is the working default. The trade is explicit: higher token use in exchange for stronger results and faster time-to-result.

Where it lives: ultra is available in ChatGPT Work for Pro and Enterprise plans, and in Codex from Plus upward. The max effort level, by contrast, is a settings toggle for every user with GPT-5.6 access in ChatGPT Work and Codex. On the API side, the equivalent is the multi-agent beta in the Responses API, which lets GPT-5.6 run concurrent subagents and synthesize their work inside a single request — the primitive for building ultra-like experiences into your own products.

When does parallelism pay? Time-boxed, decomposable work — a competitive research sweep due in an hour, a migration audit across independent services, a multi-market campaign analysis. The four agents burn tokens concurrently either way; what you buy is wall-clock time and the synthesis step. For linear tasks where step N depends on step N−1, parallel agents mostly re-derive each other’s context, and the spend converts to little.

Programmatic Tool Calling (PTC) ships day-one in the API, and it’s the piece most likely to change how agent stacks get built. Instead of round-tripping every tool call through the model, GPT-5.6 writes and runs JavaScript in a hosted, isolated V8 runtime that coordinates the tools itself — parallel calls, loops and conditionals, filtering large intermediate outputs — and returns a smaller structured result to the model. It’s ZDR-compatible, so zero-data-retention accounts aren’t excluded.

Here’s the honesty point that most GA-day coverage will miss: several of the efficiency numbers OpenAI published are tied to customers adopting this pattern, not to swapping a model slug. The per-dollar wins partly come from building differently, and they’re OpenAI-reported customer results:

So when should you re-architect around it, and when should you leave a working stack alone? The deciding variable is how predictable your control flow is. If the agent’s tool sequence is mostly known in advance — fetch, filter, transform, aggregate — PTC turns a dozen model round-trips into one program run. If the agent genuinely doesn’t know its next step until it sees the last result, direct tool calling remains the honest architecture. This is exactly the kind of build-versus-hold call our AI transformation engagements are structured around — evaluating a pattern on your workload before committing engineering weeks to it.

Access is split across three surfaces with different rules: ChatGPT’s chat experience, the new ChatGPT Work and Codex environments, and the API. The table below compiles OpenAI’s GA access notes into one view. Two things to read carefully: in plain Chat, GPT-5.6 access starts at Plus (OpenAI’s notes don’t list a free-tier Chat option), and “Sol Pro” — the Pro/Enterprise picker name — is the reasoning.mode: "pro" setting, not a separate model.

The pattern worth noticing: the cheapest guaranteed path to GPT-5.6 is not the chat surface at all. Free and Go users get Terra inside ChatGPT Work and Codex, while plain Chat starts at Plus — a quiet signal about where OpenAI wants new usage to land. And for teams, ultra access is the real plan differentiator: Codex gets it from Plus, but ChatGPT Work reserves it for Pro and Enterprise.

Every number below is OpenAI-reported — the GA page’s own published eval tables, with no independent audit yet at GA. One sourcing caution up front: OpenAI’s prose claims a “new high of 53.6” on Agents’ Last Exam, while its eval table lists Sol at 52.7% — the prose figure reflects a different (maxed-out) reasoning configuration. We use the table values throughout, because tables are what you can audit. For the deep Sol versus Fable 5 comparison, see our July 2 price-and-benchmark analysis — this section only updates the picture the GA tables actually changed.

Read as a whole, the table refuses a victory lap in either direction. On OpenAI’s own numbers, Sol leads long-horizon agentic work (Agents’ Last Exam, where even Terra at 50.4% and Luna at 50.3% clear every non-OpenAI model listed), the Artificial Analysis Coding Agent Index, Terminal-Bench 2.1 — where the Sol Ultra configuration posts 91.9% — and computer use, where OpenAI adds that Sol’s OSWorld result comes with 85% fewer output tokens than Opus 4.8. But on the same pages, Claude models lead SWE-Bench Pro by a wide margin, GDPval’s Elo ranking, the Artificial Analysis Intelligence Index, and HealthBench Professional.

The honest read for a working team: pick the benchmark that matches your workload, then run your own eval anyway. If your production reality is repository-scale software engineering, OpenAI’s own table says Claude still leads. If it’s long-horizon multi-step professional workflows or computer use, Sol’s case is strong. For how this slots into the broader frontier field this week, our frontier-model comparison from July 8 is the wider context — GPT-5.6’s GA re-sorts that leaderboard on some rows, not all of them.

The two-step release wasn’t marketing theater. Per OpenAI’s preview post, the limited phase happened “at their request” — the U.S. government’s — with trusted partners “whose participation has been shared with the government,” a process OpenAI says it doesn’t believe “should become the long-term default.” The reason is the model’s cyber capability, OpenAI’s strongest yet on its own evals — and the most cyber-capable use remains gated behind Trusted Access for Cyber, the verified-defender program we covered in our Daybreak and defensive-security analysis.

OpenAI says GPT-5.6 stays below the “Critical” threshold of its Preparedness Framework in both cyber and bio, and that GA followed roughly 700,000 A100e GPU-hours of black-box automated red teaming, with Sol’s cyber safeguards blocking “roughly ten times more potentially harmful activity” than previous models. Two practical notes for builders hide in that safety stack. First, generation can pause for several seconds mid-stream while classifiers synchronously review output — documented behavior, not a latency regression in your code. Second, ChatGPT and Codex offer a retry-on-a-lower-capability-model option when a safeguard adds friction. One announced-schedule item to watch: Cerebras serving Sol at up to 750 tokens per second, slated for July with select customers first.

GA didn’t ship alone. OpenAI launched ChatGPT Work the same morning — a GPT-5.6-powered environment with Codex technology built in, and the surface where ultra and free-tier Terra access actually live. It’s consequential enough that we’ve covered it separately in our ChatGPT Work launch analysis — read the two posts as one release.