GPT-5.6 is OpenAI's next-generation model family, previewed on June 26, 2026 in three variants: Sol (flagship), Terra (mainstream), and Luna (cost-optimized). It is the successor to the GPT-5.5 family released in April 2026 and is the first OpenAI flagship released under a White House access-control agreement that gives the US government a window to evaluate frontier models before public release. Sol sets a new benchmark record on Terminal-Bench 2.1 (91.9% for Sol Ultra, 88.8% for base Sol) and delivers roughly 3x output-token efficiency on ExploitBench (vulnerability research) versus the prior Mythos Preview frontier. Pricing is $5 input / $30 output per 1M tokens for Sol (same as GPT-5.5), $2.50 / $15 for Terra, and $1 / $6 for Luna. As of June 27, 2026 access is limited to vetted partners; broader rollout is expected within weeks.

What is the difference between Sol, Terra, and Luna?

Sol is the flagship — highest capability, designed for long-horizon coding, agentic security work, and frontier reasoning. Priced at $5 input / $30 output per 1M tokens. Terra is the mainstream tier — performance similar to GPT-5.5 but at half the cost, intended for general ChatGPT and API use. Priced at $2.50 input / $15 output per 1M. Luna is the cost-optimized tier — OpenAI's cheapest GPT-5.6 model, designed for high-volume background tasks, batch processing, and embedded applications. Priced at $1 input / $6 output per 1M. The naming pattern (Sol/Terra/Luna = sun/earth/moon) replaces the prior numeric-suffix convention (GPT-5.5, GPT-5.5-Cyber). Expect this naming scheme to extend forward: future siblings could include compute-tier or specialization variants.

Why is GPT-5.6 only in limited preview?

The US government requested a limited initial rollout under a frontier-AI evaluation agreement. OpenAI gave the White House a preview of GPT-5.6's capabilities before launch, and the public release is being delayed while the government conducts national-security and safety evaluations. OpenAI has publicly described the constraint as 'unsustainable' but is complying. Initial access is restricted to vetted enterprise and research partners — likely overlapping with the same critical-infrastructure organizations that the US government cleared for Anthropic Claude Mythos 5 access on the same day (June 27, 2026). OpenAI expects broader rollout in the coming weeks, meaning roughly July 2026 for ChatGPT Plus/Pro tier subscribers and the API.

What new features does GPT-5.6 introduce?

Four notable additions beyond raw capability gains. (1) More predictable prompt caching — explicit cache breakpoints and a 30-minute minimum cache life, which makes caching costs and behavior much easier to plan for in production. (2) Major output-token efficiency improvements — Sol matches Mythos Preview-class security performance using roughly 1/3 the output tokens, materially reducing agentic-workflow costs. (3) Strengthened layered safeguards for high-risk activities — the Sol variant in particular has additional protections tuned for cybersecurity, bioweapons, and other high-risk surfaces that triggered the government access controls. (4) Tighter integration with Codex (OpenAI's coding agent) and ChatGPT agent mode — Sol is the default model for new Codex sessions in preview accounts. The pricing structure also shifts: Terra at half GPT-5.5's cost is a deliberate move to capture mid-market workloads.

Quick Answer

What is GPT-5.6? Sol, Terra, Luna Explained (June 2026)

Published: June 27, 2026

What is GPT-5.6? Sol, Terra, Luna Explained (June 2026)

On June 26, 2026, OpenAI previewed GPT-5.6 — a new flagship family with three variants: Sol, Terra, and Luna. The release came with three unusual features: a White House access-control agreement limiting initial availability, a naming convention break (sun/earth/moon instead of numeric suffixes), and a benchmark gap larger than typical point releases. This page covers what GPT-5.6 is, how the three variants differ, pricing, access timeline, and how to evaluate whether to wait or use alternatives today.

Last verified: June 27, 2026.

TL;DR

GPT-5.6: OpenAI’s next-generation model family, previewed June 26, 2026
Three variants: Sol (flagship, $5/$30 per 1M), Terra (mainstream, $2.50/$15), Luna (cheap, $1/$6)
Benchmark leader: Sol Ultra 91.9% Terminal-Bench 2.1 (vs Mythos 5 88.0%, GPT-5.5 83.4%)
Efficiency leap: ~3x fewer output tokens than Mythos Preview on ExploitBench at competitive scores
Access: limited preview only as of June 27; White House cleared trusted partners only
Public rollout: “coming weeks” per OpenAI — likely July 2026
Available alternative today: Claude Fable 5 (GA), Gemini 3.5 Pro (GA), GPT-5.5 (GA)

The three variants

Sol — the flagship

Sol is OpenAI’s strongest model yet. It’s the variant the US government’s access controls are really about. Key specs:

Pricing: $5.00 input / $30.00 output per 1M tokens (same as GPT-5.5)
Benchmarks: Sol Ultra 91.9% Terminal-Bench 2.1; base Sol 88.8%
Specialization: long-horizon coding, agentic security workflows (vulnerability research, exploitation reasoning), frontier reasoning
Safeguards: strengthened layered protections for cybersecurity, bioweapons, and other high-risk activity surfaces
Default in: Codex preview, ChatGPT agent mode (for cleared accounts)

The headline efficiency claim: on ExploitBench, Sol is competitive with the prior Mythos Preview frontier while generating only ~1/3 the output tokens. For agentic workflows priced per token, that’s a ~3x cost compression at frontier capability.

Terra — the mainstream tier

Terra is the variant most users will actually use once GPT-5.6 hits ChatGPT Plus and Pro. Key specs:

Pricing: $2.50 input / $15 output per 1M tokens (half of Sol and GPT-5.5)
Performance target: similar to GPT-5.5 at half the cost
Use case: general ChatGPT use, API for everyday tasks, content generation, summarization, conversational agents

Terra is the deliberate mid-market move. By pricing it at half of GPT-5.5 with comparable performance, OpenAI is forcing a market-wide repricing of mainstream LLM API access. Anthropic Claude Sonnet at ~$3/$15 and Gemini 3.5 Pro at $1.25-$2.50/$10-$15 are the closest competitors.

Luna — the cost-optimized tier

Luna is OpenAI’s cheapest GPT-5.6 model. Key specs:

Pricing: $1 input / $6 output per 1M tokens
Use case: high-volume background tasks, batch processing, embedded applications, classification, simple agent steps
Competitive position: competes with Gemini 3.5 Flash, Claude Haiku, and DeepSeek V3.5 for the cost-sensitive segment

At $1/$6, Luna is in the same price band as the established cheap-tier models but with GPT-5.6 family lineage. Expect Luna to be the workhorse for production teams running large agent fleets where per-task cost dominates.

Pricing comparison vs other June 2026 flagships

Model	Input ($/1M)	Output ($/1M)
GPT-5.6 Sol	$5.00	$30.00
GPT-5.6 Terra	$2.50	$15.00
GPT-5.6 Luna	$1.00	$6.00
GPT-5.5 (reference)	$5.00	$30.00
Claude Fable 5	$10.00	$50.00
Claude Mythos 5	~$15	~$75 (estimated, cleared access only)
Gemini 3.5 Pro (≤200K)	$1.25	$10.00
Gemini 3.5 Pro (>200K)	$2.50	$15.00
Gemini 3.5 Flash	$0.35	$2.80
DeepSeek V3.5	$0.27	$1.10

Two things jump out:

Sol at the same price as GPT-5.5 with better benchmarks is OpenAI applying pricing pressure on Anthropic.
Terra at half GPT-5.5 price is OpenAI applying pricing pressure on Google’s Gemini Pro pricing.

Why the limited rollout matters

The White House access-control agreement is the most consequential part of this release for the industry, not just for OpenAI. It establishes a precedent:

The US government previewed GPT-5.6 capabilities before launch
It requested a limited initial rollout for evaluation
OpenAI complied (and called it “unsustainable” publicly)
Anthropic’s Mythos 5 went through a similar process — suspended June 12, 2026; partially restored to vetted US organizations on June 27, 2026 (the same day as the Sol preview)

This is the new pattern for frontier-AI releases in the US: government preview, government access list, staged public release. Expect Gemini 4 (whenever it launches) to follow the same path.

Prompt caching improvements

A quietly important change: GPT-5.6 introduces more predictable prompt caching with:

Explicit cache breakpoints — you control where the cache boundary sits in your prompt
30-minute minimum cache life — caches persist long enough to plan for in production
Predictable cost behavior — easier to estimate per-request cost

For agent workflows that reuse system prompts and codebase context across many turns, this can cut input-token costs by 90% or more. Combined with Sol’s output-token efficiency, the total per-task cost in production can drop substantially.

What to do today

If you have preview access: start with Sol for agentic coding and security workflows. Compare against your current Mythos 5 or Fable 5 deployment on real workloads, not just benchmarks. Measure output-token consumption — the efficiency claim is the real headline.
If you don’t have preview access (most teams): keep using Claude Fable 5, Gemini 3.5 Pro, or GPT-5.5 for now. Build behind a model router so you can swap Sol in the day API access opens.
For mainstream API use cases: plan a migration path from GPT-5.5 to Terra once Terra is generally available. The 50% cost reduction at similar performance is worth the integration work.
For high-volume cheap workloads: evaluate Luna against Gemini 3.5 Flash, Claude Haiku, and DeepSeek V3.5 once Luna is GA. Pricing puts it in the same band; quality differences will decide the choice.

Timeline expectations

Late June 2026: Sol/Terra/Luna available only to White-House-cleared trusted partners
Early-to-mid July 2026: likely expansion to ChatGPT Plus and Pro subscribers
Mid-to-late July 2026: likely API general availability for paying customers
August 2026 or later: Sol Ultra (the higher-capability Sol variant referenced in benchmarks) to broad availability

This timeline is OpenAI’s “coming weeks” language interpreted against the Mythos 5 precedent (15-day suspension before partial restoration).

What is GPT-5.6? Sol, Terra, Luna Explained (June 2026)

TL;DR

The three variants

Sol — the flagship

Terra — the mainstream tier

Luna — the cost-optimized tier

Pricing comparison vs other June 2026 flagships

Why the limited rollout matters

Prompt caching improvements

What to do today

Timeline expectations

Related