GPT-5.6 is generally available: run the Terra migration test today

On June 28, we wrote that GPT-5.6's pricing was confirmed but access was not — Sol, Terra, and Luna were staged behind a government safety review, limited to about 20 approved organizations. That gate lifted on July 9. OpenAI's own line was "we are expanding preview access globally now," and the three models are reachable through the API for any developer with an account, not just the approved list.

If you build on OpenAI models, this is the day to act on the plan you made two weeks ago, not just read about it.

What actually changed

The models themselves have not moved. Pricing is the same as it was in preview:

• Sol (gpt-5.6-sol) — $5 input / $30 output per million tokens. The flagship tier, positioned for the hardest reasoning and agentic work.
• Terra (gpt-5.6-terra) — $2.50 input / $15 output. OpenAI's own framing: GPT-5.5 Pro-level output at roughly half the cost.
• Luna (gpt-5.6-luna) — $1 input / $6 output. Built for high-volume, simple pipeline steps.

What changed is the gate. The review was tied to Executive Order 14409 — the voluntary pre-release evaluation arrangement signed June 2 — combined with a specific government request on GPT-5.6 because of its capability in vulnerability identification and exploitation. The Department of Commerce's Center for AI Standards and Innovation ran additional testing beyond the standard 20-organization preview before clearing broader access. That process is now done. The access restriction that made the pricing "plan against it, but wait" two weeks ago is no longer the blocker.

The test to run today

If you are running GPT-5.5 Pro in production, the concrete action is a shadow comparison against Terra, not a full cutover:

1. Mirror a slice of real traffic to gpt-5.6-terra — 5 to 10 percent, logged but not served — rather than switching your default model outright.
2. Compare output quality against your existing GPT-5.5 Pro baseline on the same prompts, not synthetic benchmarks. OpenAI's own comparison claims are close to GPT-5.5 Pro on general tasks; a same-prompt diff on your actual workload is the only number that matters for your product.
3. Recompute your token bill at Terra's rate. At $2.50/$15 versus GPT-5.5 Pro's $5/$30, a workload that holds quality parity is a same-day cost cut with no code change beyond swapping the model string.
4. Check your tokenizer delta. The same input text tends to produce more tokens on newer model generations than on older ones — factor that into the cost comparison rather than assuming the sticker price alone tells you the bill.
5. Only then flip the default, once the shadow comparison holds for a few days of real traffic, not a single afternoon.

For agentic or high-reasoning workloads currently on Sol-tier pricing elsewhere (Opus 4.8, GPT-5.5's top tier), the same shadow-test approach applies before moving anything business-critical.

What has not changed

Fable 5 remains on usage-credit pricing for consumer Anthropic plans as of this week — a separate access track under export-control authority, not EO 14409. Gemini 3.5 Pro is still in preview with no Google-confirmed GA date; treat any specific date you read for it as unconfirmed until it comes from blog.google or the Gemini API changelog directly. None of that is new today, but it is worth keeping in view: a production stack that assumed only one of these three models would be gated at any given time has now watched two different companies stage access under two different legal mechanisms inside six weeks. Multi-provider fallback is not a hedge against one company's policy anymore — it is a hedge against a category of access risk that just repeated itself.

---

Current pricing across Sol, Terra, Luna, Opus 4.8, and Sonnet 5 is in /compare. For the reasoning behind the original staged rollout, see the June 28 post on GPT-5.6's preview pricing and regulatory staging.