GPT-5.6 Terra & Luna Hit GA — Verdict Moves to Conditional

GPT-5.6 Terra and Luna reached general availability on July 9, 2026, ending the limited-preview government gate that restricted them to roughly 20 trusted partners since June 26. As of July 20, both models are confirmed available through the OpenAI API and Codex platforms, confirmed by multiple independent sources against OpenAI's live pricing page (OpenAI(opens in new tab), 2026; TechJacks Solutions(opens in new tab), 2026). Pricing held unchanged from the preview: Terra at $2.50/$15 per 1M tokens and Luna at $1/$6. Both exit pending and enter conditional — the value proposition is clear, but the absence of independent benchmark reproduction and the API/Codex-only access pattern are meaningful caveats.

What happened

The GPT-5.6 family launched June 26, 2026 as a limited preview restricted to ~20 US-government-vetted organizations, accessible only through the API and Codex platforms. On July 9, the US Commerce Department's Center for AI Standards and Innovation completed its review and the gate lifted — all three models (Sol, Terra, and Luna) became generally available across ChatGPT, Codex, and the API (OpenAI(opens in new tab), 2026; FelloAI(opens in new tab), 2026; Simon Willison(opens in new tab), 2026).

Pricing carried through from preview unchanged:

Terra is positioned as the GPT-5.5 replacement — comparable capability at half the cost. OpenAI reports TerminalBench 2.1 at 87.4%, just above Claude Fable 5, with a CTF cybersecurity score of 91.8% (OpenAI(opens in new tab), 2026). It carries activation classifiers and the same safety stack as Sol for regulated workflows. Available via gpt-5-6-terra model ID.

Luna is the budget tier that redefines the capability-per-dollar floor. OpenAI reports TerminalBench 2.1 at 84.7%, approaching GPT-5.5's 85.6% at one-fifth Sol's cost. No activation classifiers. Available via gpt-5-6-luna model ID.

Neither model is selectable in the standard ChatGPT model picker. Terra and Luna are available in Work and Codex(opens in new tab) depending on plan tier, but the classic ChatGPT interface still serves GPT-5.5 (OpenAI Help Center(opens in new tab), 2026).

Why it matters

The GA transition removes the single biggest barrier to adoption. Before July 9, both models were effectively off-limits to anyone outside the trusted-partner program — a policy hold that made the GPT-5.6 family more of a press release than a product for most teams. With the gate lifted, the economics are now actually usable.

Terra is the most commercially important model in the family. At $2.50/$15, it undercuts Claude Sonnet 5's standard rate of $3/$15 on input while claiming GPT-5.5-class capability. For teams that built their production stack on GPT-5.5, Terra is the direct cost-down path — half the per-token spend for comparable output.

Luna at $1/$6 is the strongest budget contender on the market for capability-per-dollar, competing directly with Claude Haiku 4.5 ($1/$5) and undercutting Gemini 2.5 Flash ($0.30/$2.50 on output). For model-routing architectures — Luna by default, escalate to Terra or Sol — the economics pencil out.

The caveat that prevents a stronger verdict: every benchmark figure on this page is vendor-reported. No independent third-party has reproduced the TerminalBench or CTF scores, and the preview-to-GA transition was only two weeks — not enough time for the evaluation ecosystem to catch up. Terra and Luna remain in the "promising on paper, verify with your own evals" zone.

What changes for you

• Replace GPT-5.5 with Terra for most production workloads. With pricing stable through GA, Terra is the cost-down path for teams that ran GPT-5.5 in production. Run a side-by-side eval on your own task distribution before cutting over.
• Build model-routing architectures with Luna as the cheap default tier. Route routine classification, extraction, and summarization to Luna; escalate to Terra when eval scores dip below threshold; reserve Sol for the hardest agentic work.
• Hold on ChatGPT integration. Terra and Luna remain API/Codex-only in standard ChatGPT. The model picker still serves GPT-5.5. If your team relies on the ChatGPT interface for everyday work, you're still on the previous generation.
• Run your own evals before committing spend. All benchmark claims — TerminalBench, CTF, HealthBench — are vendor-reported. The GA window is only two weeks old. Treat OpenAI's numbers as directional until your own eval pipeline confirms.

FAQ

Are Terra and Luna available in ChatGPT? Not in the standard ChatGPT model picker. Terra and Luna are available in Codex and Work depending on your plan tier, but for standard ChatGPT conversations, the model picker still serves GPT-5.5. API access via gpt-5-6-terra and gpt-5-6-luna is unrestricted for any developer with an OpenAI API key.

How do Terra and Luna compare to Claude models? Terra ($2.50/$15) undercuts Claude Sonnet 5 ($3/$15) on input pricing at comparable reported capability. Luna ($1/$6) competes with Claude Haiku 4.5 ($1/$5) — slightly pricier on output but with higher reported benchmark scores. The key difference: Claude's benchmarks have been independently validated across multiple organizations; GPT-5.6's are OpenAI-reported only.

Should I switch from GPT-5.5 to Terra today? For API workloads — yes, with the caveat that you should run your own evals first. The cost savings are real ($2.50 vs $5.00 input), but the capability claim rests entirely on vendor benchmarks. For ChatGPT-dependent workflows, you can't switch yet — Terra isn't in the standard picker.