GPT-5.6 Sol, Terra, and Luna: Full Pricing, Benchmark Breakdown, and What the Government-Controlled Preview Means for Developers

OpenAI previewed GPT-5.6 Sol (flagship), Terra (balanced), and Luna (affordable). Here are the specs, pricing, benchmark results, and what the US government's limited-release request means for developers and enterprises waiting for access.

OpenAI dropped a big announcement on June 26, 2026: the GPT-5.6 family — three models codenamed Sol, Terra, and Luna — with tiered pricing and a phased rollout that the US government asked them to keep narrow for now. This isn't just another model update. The release dynamics themselves are changing, and if you build on top of these models, you need to know what's actually different.

The Three Models

OpenAI is splitting the GPT-5.6 lineup into three tiers — a smart move that mirrors how Anthropic has been positioning Claude Mythos and Fable as distinct products for different use cases.

Sol is the flagship. It's OpenAI's most capable model to date, designed for deep reasoning, complex coding workflows, biology research, and cybersecurity analysis. It introduces two new modes:

• Max reasoning effort — gives Sol more time to think deeply on hard problems

• Ultra mode — goes beyond single-agent limits by coordinating sub-agents for complex multi-step tasks

Terra is the balanced model. OpenAI says it's competitive with GPT-5.5 but costs half as much. For teams that found GPT-5.5 too expensive for daily use but still want strong performance, Terra sits in the middle.

Luna is the affordable option. OpenAI describes it as offering "strong capability at our lowest cost." No specific pricing yet, but expect it to compete with fast/cheap models like Claude Haiku or Gemini Flash.

Pricing (What We Know)

The TechCrunch coverage confirmed the starting price: Sol costs $5 per million input tokens, positioning it firmly as a premium reasoning model. As a baseline:

| Model | Tier | Price per M input tokens | Comparable to | |-------|------|------------------------|---------------| | Sol | Premium reasoning | ~$5 | GPT-5.5 reasoning, Claude Mythos | | Terra | Balanced | ~$7.50 (half of GPT-5.5 per OpenAI) | GPT-5.5, Claude Fable | | Luna | Fast/cheap | TBD, likely under $1 | Claude Haiku, GPT-4o mini |

Full API pricing hasn't been published on OpenAI's platform pricing page as of this writing, but these are directionally accurate based on the company's own statements.

Benchmark Performance

OpenAI shared a targeted set of benchmarks rather than a full suite — they promise expanded results when the models go broadly available:

Terminal-Bench 2.1 — Sol sets a new state of the art on this test of command-line agentic workflows. This is the benchmark that measures a model's ability to plan, iterate, and coordinate tools in real terminal environments. If you use AI coding agents like Codex or Cursor's agent mode, this directly impacts how well Sol will handle multi-step dev tasks.

GeneBench v1 — Sol shows strong improvements in long-horizon genomics and quantitative-biology analyses, achieving better results than GPT-5.5 while using fewer tokens. This is narrower than most developers need, but it hints at broader efficiency gains in the architecture.

ExploitBench — Sol is competitive with Anthropic's Mythos Preview (the model that was restricted by the US government earlier this month) while using roughly one-third of the output tokens. That's a significant efficiency gain for cybersecurity-related work.

ExploitGym — Sol, Terra, and Luna all show "strong improvements in cyber capabilities as reasoning increases."

Critically, OpenAI says Sol does not cross the Cyber Critical threshold under its Preparedness Framework. In tests against Chromium and Firefox, it identified bugs and exploitation primitives but did not autonomously produce a functional full-chain exploit. That said, the company acknowledges benchmark thresholds can't capture every misuse scenario — hence the phased release.

The Government Angle

This is the part that matters if you're a developer or enterprise waiting for access. The Trump administration requested that OpenAI limit the initial rollout to a "small group of trusted partners." OpenAI complied but made its frustration clear in the announcement:

> "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."

This follows the June 12 incident where Anthropic was ordered to pull Mythos 5 and Fable 5 from the market entirely — and the subsequent ban on foreign nationals accessing certain models. The government's concern centers on cybersecurity capabilities, where frontier models are increasingly capable of finding and chaining vulnerabilities.

OpenAI is positioning the restricted preview as a short-term compromise while it works with the administration on a "cyber Executive Order framework and a repeatable process for future model releases."

What This Means for Developers and Enterprises

1. If you're not a "trusted partner" on the government-approved list, you'll have to wait weeks, not days, for API access. OpenAI said the preview is for a small group "whose participation has been shared with the government."

2. Terra might be your sweet spot — if Terra truly matches GPT-5.5 at half the cost, that's a significant operational savings for teams that upgraded to GPT-5.5 recently.

3. Safety guardrails are built into the model behavior, not bolted on top as a separate filter. OpenAI is consciously avoiding the classifier-based approach that caused issues for Fable 5, where non-security prompts were invisibly routed to older models.

4. Expect more government involvement, not less. Both Anthropic and OpenAI now have their most capable models under government-negotiated access agreements. If you're building products on frontier models, build with API key rotation and multi-provider fallbacks as a hedge against availability disruptions.

Practical Takeaways

• For everyday coding tasks: Terra is likely the best value unless you need maximum reasoning depth

• For security research, complex debugging, or agent workflows: Sol's Ultra mode could be a significant step up from GPT-5.5

• For high-volume, low-cost use: Luna should compete well with existing cheap models

• Don't assume any frontier model stays available — multi-provider routing is becoming table stakes

OpenAI said GPT-5.6 will be generally available "in the coming weeks" across ChatGPT, Codex, and the API. Watch the platform pricing page for exact per-model pricing when the broader release happens.