GPT-5.6 Goes GA: The Locked Preview Opens Two Weeks Later

What Changed on July 9

This is the general-availability follow-up to the locked preview we covered on June 29. The short version of the unlock: a June U.S. executive order directs agencies to design a voluntary framework for government access to certain frontier models before trusted-partner releases. It also explicitly says that the framework is not a licensing or preclearance requirement. OpenAI said the government requested its initial limited preview, so GPT-5.6 went first to a small group of trusted partners whose participation was shared with the government. OpenAI announced general availability on July 9, with a global rollout expected to take roughly 24 hours.

What GA means in practice: all three models — Sol (flagship), Terra (balanced), and Luna (fast and cheap) — are self-serve in the API with official model IDs (gpt-5.6-sol, -terra, -luna), available in Codex, and rolling out across ChatGPT's paid tiers. This is the first time GPT-5.6 appears in ChatGPT at all; the preview was API and Codex only, for invited partners.

Pricing held at the numbers shared during the preview: Sol at $5 per million input tokens and $30 per million output, Terra at $2.50 / $15, and Luna at $1 / $6. Cache reads keep a 90% discount with a 30-minute minimum retention window, which matters for agent sessions that reuse the same repository context across many calls. These are list prices, not launch promotions — but as we argued in the intro-pricing trap, it's still worth writing down which price your unit economics assume before a dependency hardens.

What GA Ships That the Preview Didn't

Two additions stand out for developers. The first is programmatic tool calling in the Responses API: instead of emitting one tool call at a time and waiting for your code to respond, the model can write and run an in-memory program that coordinates multiple tools and processes intermediate results. OpenAI says this can reduce tokens and model round trips because only the results that matter need to return to the context window. If your agent burns most of its budget shuttling tool output back into the prompt, this is the feature to evaluate first.

The second is that ultra mode — the subagent orchestration feature from the preview — now has concrete availability and numbers. It runs four agents in parallel by default and lifts Sol's self-reported Terminal-Bench 2.1 score from 88.8% to 91.9%. It's available in Codex from the Plus tier up and in ChatGPT Work for Pro and Enterprise plans. Note what ultra mode actually is: a spend-more-to-get-more lever. Four parallel agents consume tokens accordingly, so treat it as a budget decision per task, not a default switch to flip. There's also a new max reasoning-effort setting above the existing levels, and the usual caveat applies to all the launch benchmarks: they're OpenAI's own numbers until independent evals reproduce them.

What This Means for How You Build

• Run the evaluation you postponed. Two weeks ago the right move was to plan an eval and wait. Now the model is callable, so the plan comes due: run Sol — and Terra, which OpenAI positions as GPT-5.5-class at half the price — against your own tasks behind your own evals before deciding anything. Vendor benchmarks decide what to test, not what to ship.
• The lock resolved in 13 days — this time. The useful datapoint from this launch cycle isn't just that access opened; it's that a voluntary government-access step became a real stage in this model's release pipeline. The executive order does not impose preclearance, and the same process may not recur unchanged, but model availability plans should still treat "announced" and "callable" as separate events with an unpredictable gap between them — the same discipline as in when your AI model gets pulled, just pointing forward instead of backward.
• Budget ultra mode like infrastructure, not like a setting. A mode that runs four agents in parallel changes your cost per request by a multiple, for a benchmark gain of about three points. If you turn it on, do it per task class with a cost ceiling — the parallel-work decomposition habits from running AI agents in parallel apply directly.
• Programmatic tool calling is the quiet headline. Orchestration moving from your application code into a sandboxed, model-written script is a bigger architectural signal than any benchmark delta. If it holds up, some of the glue code agent frameworks exist to provide gets absorbed by the API layer. Worth a prototype before you extend your own orchestration layer.

The short version: the best coding model you couldn't use is now a model you can evaluate, and the two-week lock ends as a footnote rather than a precedent — until the next launch, when the review window may be longer. Build the eval, price the modes, and keep the fallback path you already have.