GPT-5.6 — OpenAI's Three-Tier (Sol / Terra / Luna) Frontier Family + Token-Efficiency Claims
OpenAI launched GPT-5.6 as three tiers (Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per Mtok) on 2026-07-09; claims 54% higher token efficiency on agentic coding vs rivals plus a 90% cached-read discount — a live instance of tier-differentiated frontier pricing.
GPT-5.6 — OpenAI's Three-Tier Frontier Family + Token-Efficiency Claims
Fetch note: OpenAI's own announcement page (openai.com/index/gpt-5-6/) returned HTTP 403 to direct fetch. This entry is built from news coverage (Tech Startups, CryptoBriefing) that directly quotes OpenAI's release and CEO Sam Altman — treated as
newssourceType per KB conventions (lower authority than a first-party technical report), not a substitute for the primary page.
Core Facts
- Limited preview: 2026-06-26. Public availability: 2026-07-09, following a government review period.
- Three independent capability tiers, replacing a single flagship-model release cadence:
| Tier | Input $/Mtok | Output $/Mtok |
|---|---|---|
| Sol (flagship) | $5 | $30 |
| Terra (mid) | $2.50 | $15 |
| Luna (budget) | $1 | $6 |
- Caching: 90% discount on cached reads; 1.25x input rate on cache writes.
- Token-efficiency headline claim: Sam Altman states 54% more token-efficient on agentic coding than prior/rival models. Concretely, GPT-5.6 Sol reportedly matched Anthropic's "Mythos Preview" on ExploitBench while emitting only about one-third as many output tokens.
- Framing: OpenAI frames Terra as delivering GPT-5.5-equivalent performance for "half the cost and 16% fewer tokens" for many agent workloads — tiering is explicitly pitched as intelligent routing by task complexity, not just a cheaper/worse model ladder.
Key Quotes
- Sam Altman: "Every enterprise now is thinking about spend and the value they're getting in exchange for AI, and this is what we really want to do."
- Altman on the government-review gate: "If you want broad access...you really want to be able to be confident in your safety claims, because otherwise the world is going to get uncomfortable very fast."
Limitations
- Primary OpenAI page not directly fetchable (403) — all figures here are journalism-mediated, not independently verified against OpenAI's own model card or system card.
- Context window and full benchmark suite (SWE-bench, GAIA, etc.) not captured in the sources fetched this pass — a gap for a future ingest once the primary page or a system card is reachable.
- "54% more token-efficient" is stated as a headline claim without the underlying eval methodology being reproduced in the secondary sources used here.
Why this matters for the KB
Direct hit on the AI LENS's required numeric cell ("Inference price per million input/output tokens for top-5 providers") and the content-focus.md spine. Reading GPT-5.6's tiering alongside arXiv:2603.28576's "Tiered Super-Moore" finding (also ingested this pass) gives the KB a real-time instance of the paper's thesis: flagship (Sol) pricing resists decline while lower tiers (Terra, Luna) compress faster — exactly the differentiated tier-decay pattern the paper documents empirically across 3,000+ models.
Sources: OpenAI's GPT-5.6 is 54% more token efficient — Tech Startups, 2026-07-09; pricing table corroborated by CryptoBriefing. Primary page: openai.com/index/gpt-5-6/ (403 at fetch time).