Three weeks after GPT-5.6 Sol surfaced on prediction markets and in pre-launch leaks, OpenAI made it official on June 26: a three-tier model family, a 1.5-million-token context window on the flagship, and a pricing ladder designed to cover everything from bulk API calls to frontier reasoning. By the end of the day, almost none of it was accessible to the public.
The Trump administration invoked a June 2 executive order requiring federal evaluation before wide release of powerful new AI models. GPT-5.6 is currently restricted to approximately 20 pre-approved organizations via Codex and the API. OpenAI complied but pushed back immediately: “We don’t believe this kind of government access process should become the long-term default.”
General availability is expected in the coming weeks — analysts estimate July 10–17 — at which point standard API customers will gain access. Here’s what you’re getting when it arrives.
Three Models, Three Philosophies#
Sol is the frontier offering. It targets coding, biological research, and cybersecurity — domains where extended reasoning over vast context produces outsized returns. The 1.5M-token context window makes it the largest production context in OpenAI’s lineup, 43% larger than GPT-5.5 Pro’s 1.05M tokens. Sol includes a “max” reasoning effort mode and an “ultra” sub-agent mode for multi-step autonomous workflows. Pricing: $5 input / $30 output per million tokens.
Terra is the workhorse. OpenAI positions it as “2x cheaper than GPT-5.5 with comparable performance” for high-volume business workloads — a direct challenge to Anthropic’s Sonnet tier and Google’s Gemini 3.5 Flash. Pricing: $2.50 input / $15 output per million tokens.
Luna is the speed and cost tier: $1 input / $6 output, built for latency-sensitive and high-frequency deployments. Autocomplete, classification, lightweight chat.
The naming breaks from OpenAI’s prior convention (version numbers attached to the base model) toward something more evocative. Astronomical bodies. Whether the performance lives up to the branding is something we’ll know once benchmarks drop.
The Government’s Pattern Is Now Clear#
June 26, 2026 was a remarkable day in AI policy. In the space of a few hours, the US government simultaneously:
- Restricted GPT-5.6 to ~20 pre-approved organizations under the June 2 executive order
- Authorized Anthropic to restore Claude Mythos 5 access for 100+ US critical infrastructure institutions
Both actions landed on the same day. Neither was announced in advance. Both affected the most capable models from the two leading AI companies.
The legal instruments are different. Fable 5 and Mythos 5 were suspended June 12 via export control law, citing a specific “fix this code” jailbreak demonstrated to the government. GPT-5.6’s restriction comes from a broader June 2 executive order that simply requires federal evaluation of powerful new models before wide release — no specific jailbreak required.
But the pattern is the same: the US government is now actively intervening in frontier model launches for both Anthropic and OpenAI, operating under different statutes but converging on the same result. Frontier AI is being treated as a regulated capability, not a software product.
Anthropic’s experience is instructive for what OpenAI should expect. The Mythos 5 partial reinstatement (100+ institutions in critical infrastructure) came 15 days after the ban. Fable 5 remains restricted for general users. The government’s resolution path: fix the specific concern, negotiate the scope of reinstated access, and wait. OpenAI’s timeline will likely follow a similar shape — weeks to general availability, not days.
What This Means for Developers#
Sol’s 1.5M-token context window is the number most worth watching, because it’s large enough to change what’s architecturally possible in agentic coding.
A 1.5M context can hold:
- An entire mid-size codebase (150K–300K lines of code with tests and documentation)
- End-to-end security audit of a monorepo without chunking
- Multi-file refactors with full dependency graphs in context simultaneously
The practical question is whether that context holds coherently under agentic load. Every large-context model released in the last two years has degraded at the edges — retrieving information from the beginning of context less reliably than from the middle or end. If Sol maintains coherent reasoning across 1.5M tokens under multi-step agentic tasks, it represents a real capability jump. If it doesn’t, the spec is impressive and the performance is ordinary.
Sol’s “ultra” sub-agent mode is the other benchmark for whether OpenAI has meaningfully closed the agentic gap with Claude Code. No independent data is available at the limited-preview stage. OpenAI’s own numbers haven’t been published.
Terra Is the Competitive Threat to Watch#
Sol gets the headlines, but Terra is probably the model that matters most for the developer ecosystem.
At $2.50/$15 per million tokens, Terra lands between Claude Haiku 4.5 ($0.80/$4.00) and Claude Sonnet 4.6 ($3/$15). OpenAI’s claim of “comparable to GPT-5.5 performance at 2x lower cost” positions Terra as a direct substitute for Sonnet 4.6 on price-sensitive agentic workloads.
If that benchmark claim holds under independent evaluation, Terra becomes a serious competitor to Anthropic’s mid-tier on the economics. GPT-5.5 was already competitive with Sonnet 4.6 on many coding tasks at comparable pricing. A model that matches GPT-5.5 at half the price would shift the cost calculus for teams running high-volume agents.
What Got Retired#
GPT-4.5 was simultaneously retired from ChatGPT on June 26. Existing conversations migrate automatically to GPT-5.5. OpenAI’s model portfolio is now a clean three-generation stack: 5.3-Codex, 5.5, and 5.6 (Sol/Terra/Luna). No model is actively maintained below GPT-5.3-Codex.
The retirement also marks the end of the 4.x era — models that were still positioned as the “reliable” choice for teams that didn’t want to trust the newest generation. That comfort zone is gone. Teams still on GPT-4.5 are now on GPT-5.5 whether they chose it or not.
The Practical Stack Right Now#
For teams evaluating AI coding tools this week: the frontier has moved, but it’s not fully accessible yet.
The practical stack is Claude Opus 4.8 / Fable 5 (where accessible, pending the ongoing restrictions) for the highest-autonomy tasks, GPT-5.5 for teams in the OpenAI ecosystem, and Gemini 3.5 Flash for Google-native workflows. Gemini 3.5 Pro remains delayed — expected July 2026, per unofficial reports.
GPT-5.6 Terra, once generally available, will likely become the default for mid-tier API workloads if the performance claims survive benchmarking. GPT-5.6 Sol will be the benchmark for context-heavy agentic coding — but only once someone can actually run it.
Watch for independent evaluations in the first week of July.
Sources:
- Axios — OpenAI releases powerful new GPT-5.6 model under restrictions
- TechCrunch — OpenAI limits GPT-5.6 rollout after government request
- MacRumors — OpenAI Launches GPT-5.6 Sol, Terra, and Luna in Limited Preview
- VentureBeat — OpenAI unveils GPT-5.6 Sol, Terra and Luna models
- Decrypt — OpenAI Rolls Out GPT-5.6—But Only for Some Users Due to Trump Admin
- 9to5Mac — Anthropic cleared to release Claude Mythos 5 to US institutions
- TechTimes — GPT-5.6 Launch Window: 1.5M Token Context