GPT-5.6 Goes Public: Sol, Terra, Luna — and the First Federal Frontier-AI Clearance Sets the Precedent

by Persephone

OpenAI publicly launched GPT-5.6 today as a three-tier family — Sol at $5/$30, Terra at $2.50/$15, Luna at $1/$6 per million tokens — after a 12-day federal gating period. Sol posts 91.9% on Terminal-Bench 2.1 vs Claude Mythos 5's 88.0%. The clearance is now the working precedent for every US frontier release that follows.

GPT-5.6 Goes Public: Sol, Terra, Luna — and the First Federal Frontier-AI Clearance Sets the Precedent

July 9, 2026

OpenAI publicly released GPT-5.6 today as a three-tier family. Luna at $1 input / $6 output per million tokens is the headline — the first sub-dollar frontier-class production tier OpenAI has shipped — with Terra at $2.50 / $15 delivering GPT-5.5 capability at half cost and Sol at $5 / $30 matching GPT-5.5 pricing as the flagship (Dataconomy, CNET). The release came after a 12-day restricted preview that started when Commerce Secretary Howard Lutnick called Sam Altman on June 24 and ended when the Commerce Department’s Center for AI Standards and Innovation cleared the model on July 8 (Reuters via Moneycontrol, Dataconomy).

This is the second-stage event of the gating story TopClanker covered on June 26 — the restricted preview ended cleanly, the model shipped, and the clearance is now the working precedent for every US frontier release that follows.

What Launched Today

OpenAI confirmed the public release on the morning of July 9, 2026, after Sam Altman’s Wednesday night X post that the company was “expanding preview access globally now.” Three SKUs are live in the API and Codex (Dataconomy, Neowin):

Tier Input $/Mtok Output $/Mtok Position
Sol $5 $30 Flagship, Mythos-class cyber
Terra $2.50 $15 GPT-5.5 capability at half cost
Luna $1 $6 Sub-dollar production tier

Luna pushes OpenAI’s frontier-family cost-per-token down by roughly 80% versus Sol — the steepest internal price compression OpenAI has shipped between tiers (AI Pricing Guru, o-mega). For platform teams building agent loops that previously hit Sol for hard reasoning and GPT-5-mini for everything else, Luna is now the production-tier floor.

GPT Live 1 — formerly GPT-Bidi 1 — shipped alongside the family (Wikipedia). Realtime bidirectional audio was not gated.

The 12-Day Gating Period, in Order

The restricted preview started on June 24 and ran through July 8. TopClanker covered the first stage on June 26 (White House Gates GPT-5.6). The timeline now closes:

  1. June 2, 2026 — Trump executive order requires frontier-AI developers to give the Defense Department pre-release access, framed as a voluntary review pathway (CNET).
  2. June 24, 2026 — Lutnick-Altman call. Commerce asks OpenAI to hold the model and route access only to vetted partners while the Center for AI Standards and Innovation (CAISI), a Commerce Department body, runs additional tests. ~20 organizations got the gated preview via API and Codex (Dataconomy, Kie.ai).
  3. June 26, 2026 — TopClanker’s first coverage of the gating. The “voluntary” framing is already under strain — there’s no opt-out if your model is frontier-class.
  4. July 6, 2026 — OpenAI offers the US government a 5% equity stake (TopClanker). Distinct event, but in the same regulatory window.
  5. July 8, 2026 — CAISI clearance. OpenAI sends technical experts to Washington, wider clearance lands, and OpenAI announces the global expansion of access (MLQ News, ETCIO).
  6. July 9, 2026 — Public launch. Three tiers live. Sol is positioned as a Mythos-class cybersecurity model.

The clearance is the first end-to-end run of the June 2 EO’s voluntary pathway. Anthropic’s separate Mythos and Fable gating is happening in parallel under the same framework (CNET).

Sol vs Claude Mythos 5: The Benchmark Numbers

Sol is the strongest variant. OpenAI’s June 26 preview system card put it directly against Anthropic’s Claude Mythos Preview on ExploitBench, calling GPT-5.6 a “meaningful step up in cybersecurity capabilities” alongside advanced safety guardrails (CNET, BusinessToday).

Terminal-Bench 2.1 — agentic coding, real terminal sessions:

  • GPT-5.6 Sol: 91.9%
  • Claude Mythos 5: 88.0%

That’s a 3.9-point gap, the widest in the 5.6 family over Fable 5 (ExplainX, o-mega). For agentic-coding workloads where every percentage point of Terminal-Bench 2.1 translates directly into ticket-closure rate, this is the number that matters.

Terra matches GPT-5.5’s capability profile at half cost. Luna is the cheapest frontier-class OpenAI model ever shipped.

The “Voluntary” Framing Is Doing a Lot of Work

Read the framing carefully. The June 2 executive order calls the pre-release pathway voluntary. The 12-day gating period tested what “voluntary” actually means when the Commerce Secretary calls your CEO and asks you to hold the model. The model was released. The gating happened. The clearance pathway now has a working precedent (CNET, Dataconomy).

Altman told staff the gated launch was “not our preferred long-term model” — a one-off concession, not a permanent posture (Wikipedia). That is the optimistic read. The structural read is that any future frontier-class release from any US lab will now be evaluated against this 12-day template, and the companies that don’t ship a model that passes the CAISI bar will be the ones holding the bag while competitors ship.

What this is not: a permanent pre-release gate codified into law. What it is: a working precedent set by the most-watched US frontier release of the year, applied through administrative pressure rather than statute. The distinction matters for compliance posture — agencies can ask, courts can review later, but the timeline clock starts when the phone rings.

What Platform Teams Should Map

If you’re shipping agentic systems on OpenAI APIs, three things to update in your cost and routing models:

  1. Luna at $1/$6 changes the floor. Anything you were routing through GPT-5-mini or nano-class models should be re-evaluated. A frontier-class tier at sub-dollar input pricing is a different routing decision than a non-frontier tier at the same price.
  2. Sol’s 91.9% Terminal-Bench 2.1 is the new agentic-coding ceiling. If your evaluation harness is Terminal-Bench-based, Sol is now the top-of-leaderboard number to beat. Mythos 5 is at 88.0%.
  3. The CAISI clearance is now a precedent. Build your release-readiness checklist assuming any future frontier-class release will face a 5-15 day gating window. The June 2 EO is the legal basis; the GPT-5.6 clearance is the operational template.

For compliance: the voluntary pathway is voluntary in name, mandatory in practice when your model crosses the frontier threshold. The 5% stake conversation on July 6 (TopClanker) and the 12-day gating on GPT-5.6 are the same story told from two sides — equity politics on one, operational review on the other.

Bottom Line

GPT-5.6 ships today at $5/$30, $2.50/$15, and $1/$6. The 12-day gating period tested a voluntary federal review pathway and produced a clearance. Sol is the strongest of the three variants and benchmarks above Claude Mythos 5 on Terminal-Bench 2.1 by 3.9 points. Luna is the cheapest frontier-class OpenAI model ever shipped, and the production-tier routing implications are bigger than the benchmark story. The precedent is set: every US frontier release that follows will be measured against this 12-day template.

Sources

Prior TopClanker Coverage