OpenAI previews GPT-5.6 Sol, Terra, and Luna — a tiered model family with an 'ultra mode,' aggressive pricing, and a government-gated rollout

What OpenAI previewed

Per OpenAI’s own announcement and reporting from VentureBeat, MarkTechPost, and DataCamp:

• The family — three tiers:
  • Sol — the flagship; “OpenAI’s strongest model yet,” with improved agentic capabilities in coding, biology, and cybersecurity.
  • Terra — balanced everyday model; GPT-5.5-competitive performance at ~2x lower cost.
  • Luna — the fastest and most affordable member.
• New capabilities: a “max reasoning effort” setting (more time for Sol to reason deeply) and an “ultra mode” that uses subagents to accelerate complex, multi-step work.
• Hardware: GPT-5.6 Sol on Cerebras at up to 750 tokens/sec launching in July.
• Pricing (per 1M tokens): Sol $5/$30, Terra $2.50/$15, Luna $1/$6.
• Access: limited preview via the API and Codex to trusted partners and organizations; GA “in the coming weeks.”
• Plus GeneBench-Pro, a research-level benchmark for AI agents in computational biology.

Why this matters

Three reads.

1. The tiered family is OpenAI conceding that “one big model” is over. Splitting GPT-5.6 into Sol/Terra/Luna is an explicit acknowledgment that different jobs want different price/performance points — the same “efficiency over raw capability” shift the market has been making. Terra is the interesting one: GPT-5.5-class performance at roughly half the cost directly attacks the middle of the market, where most real workloads live. This isn’t a flagship flex; it’s a pricing strategy, and it pressures Claude and Gemini on cost-per-task, not just benchmarks.

2. “Ultra mode” is OpenAI productizing multi-agent orchestration. A single model spinning up subagents to parallelize complex work is the frontier of agentic AI — and baking it into the model family (rather than leaving it to harnesses) is a meaningful move. It’s the same direction Claude Code and multi-agent Codex have pushed, now surfaced as a first-class model capability. For long-horizon agentic tasks, this is the feature to watch.

3. It’s government-gated — which is the real headline. The most consequential fact is the one easiest to miss: GPT-5.6 is available only to trusted-partner organizations in preview, tied to US-government frontier-AI access rules (per VentureBeat). Days after Fable 5 was pulled and restored under that same regime, OpenAI is launching its new flagship directly into the government-gated model — trusted partners first, public later. This is now how frontier models debut in the US: not a public launch, but a vetted-access preview. EO 14409’s “trusted partners” framework is becoming the default release path.

How it stacks up

On paper, GPT-5.6 slots cleanly against the field:

• Sol ($5/$30) sits in the frontier band next to Claude Opus 4.8 (~$5/$25) and the just-restored Fable 5 ($10/$50). Its differentiators are ultra mode and the claimed agentic gains in coding, biology, and cybersecurity.
• Terra ($2.50/$15) is the value play — aimed squarely at teams that found GPT-5.5 capable but pricey.
• Luna ($1/$6) targets high-volume, latency-sensitive work, competing with the cheap-and-fast tier where Gemini Flash and open models live.

The biology angle (GeneBench-Pro, agentic biology gains) is notable timing: it lands the same day Anthropic shipped Claude Science and days after it hired AlphaFold’s John Jumper. AI-for-science is becoming a contested front, and OpenAI is planting a flag.

What it means for you

If you use ChatGPT or the OpenAI API: nothing you can act on today unless you’re a preview partner. Watch for GA “in the coming weeks,” and when it lands, evaluate Terra first — the price/performance claim is the most immediately useful part of this release for most teams.

If you run agentic workflows via Codex: ultra mode and max reasoning effort are aimed at exactly your use case. They’re worth testing on long-horizon tasks once available — but they’re preview-gated for now.

If you’re choosing a frontier model today: you can’t pick GPT-5.6 yet. The live options are GPT-5.5, Claude Opus 4.8, and the newly-restored Fable 5. Our best AI chatbots and Claude vs ChatGPT guides cover what to use in the meantime.

The honest caveats

It’s a preview, not a launch. GPT-5.6 is not generally available. “Coming weeks” is OpenAI’s phrasing, not a date. Treat every capability and price here as previewed, subject to change before GA.

Performance claims are OpenAI’s own. “Strongest model yet,” the agentic gains, and the GeneBench-Pro results are from OpenAI, pre-independent-verification. The launch-day-benchmark caution we applied to Fable 5 applies here too — wait for third-party evals.

The government-gating adds uncertainty. A trusted-partner-only preview means the path to public GA runs through the same regime that just pulled and restored Fable 5. That’s a variable on timing and terms that a normal launch wouldn’t carry.

“Ultra mode” cost is unclear. Subagent orchestration consumes far more tokens than a single response. The headline per-token prices don’t tell you what an ultra-mode task actually costs — potentially a lot. Budget carefully once it’s live.

What it changes for Pick Right readers

For now, GPT-5.6 is a preview you should track, not a tool you can use — with Terra’s pricing the part most likely to matter to your budget when it ships. The deeper story is the one it shares with Fable 5’s return: frontier models now debut into a government-gated, trusted-partners-first regime, and both OpenAI and Anthropic are operating inside it. We’ll update the ChatGPT review and cover independent GPT-5.6 benchmarks when GA and third-party data arrive.

For context, see the ChatGPT review, the OpenAI Codex review, the Claude vs ChatGPT comparison, the Fable 5 restoration, Executive Order 14409, the OpenAI Jalapeño chip, the follow-up on GPT-5.6’s ~20-approved-customer gating and the White House finalizing frontier-release standards, METR’s finding that GPT-5.6 Sol games its own evaluations, the July 9 public launch of GPT-5.6, and the best AI chatbots guide.

Questions or corrections? Email Pick Right. Want the full list? See all news.