GPT-5.6 (Sol, Terra, Luna)
OpenAI's next-generation family, generally available since 9 July 2026 — three models, a 1.05M-token context window, and a split benchmark record against Claude Fable 5.
GPT-5.6 specs at a glance
| Maker | OpenAI |
| General availability | 9 July 2026 (global rollout over ~24 hours) |
| Preview began | 26 June 2026 — ~20 partner organizations, government safety review |
| Variants | Sol (flagship) · Terra (balanced) · Luna (fast/affordable) |
| Context window | 1,050,000 tokens — all three variants (preview-era "1.5M" figure was incorrect) |
| Max output | 128,000 tokens |
| Knowledge cutoff | 16 February 2026 |
| API pricing | Sol $5/$30 · Terra $2.50/$15 · Luna $1/$6 per 1M tokens (input/output) |
| Reasoning modes | max (extended chain-of-thought budget) · ultra (spawns internal subagents to decompose and parallelize work) |
| Availability | ChatGPT, ChatGPT Work, Codex, OpenAI API |
Calculate API costs for this model → Compare its monthly bill against 18 other rate cards.
From preview to GA in 12 days
GPT-5.6 had an unusually public gating episode. OpenAI began a limited preview on 26 June 2026, shipping Sol, Terra and Luna to roughly 20 trusted partner organizations at the request of the US government, behind a formal safety review. Those restrictions were lifted on 8 July 2026, and the family went generally available the next day, 9 July — rolling out across ChatGPT, the new ChatGPT Work agent tool, Codex and the API over about 24 hours.
One correction from the preview period worth flagging: an oft-cited 1.5-million-token context window circulated during preview coverage. The confirmed GA figure, per OpenAI's own API documentation, is 1,050,000 tokens across all three variants — independent trackers have flagged the 1.5M number as incorrect.
The three models: Sol, Terra, Luna
Sol is the flagship — OpenAI's new frontier model, launching with what the company calls its most robust safety stack to date, and positioned as the strongest model yet for long-horizon cybersecurity work (evaluated on ExploitGym, a benchmark built with UC Berkeley researchers and other frontier labs).
Terra is the balanced everyday-work model, priced at $2.50/$15 per 1M tokens — roughly half of GPT-5.5's $5/$30 rate card for reportedly comparable quality on everyday tasks. That price also lands within a few cents of Claude Sonnet 5's standard $3/$15 rate card.
Luna is the fast tier: OpenAI's lowest-cost model at $1/$6 per 1M tokens, aimed at the same high-volume segment as Claude Haiku 4.5 and Gemini Flash models — and its output rate exactly matches Grok 4.5's $6 per-1M output price.
GPT-5.6 Sol benchmarks
The figures below are Sol's published scores, cross-checked against independent coverage of OpenAI's GA benchmark charts. As with any vendor-published suite, treat these as directional rather than definitive — and note that benchmark versions matter (Terminal-Bench 2.0 vs 2.1 are not directly comparable).
| Benchmark | Sol | Comparison (same or adjacent chart) |
|---|---|---|
| Terminal-Bench 2.1 | 91.9% | Claude Fable 5 (adaptive reasoning): 88.0% · Gemini 3.1 Pro: 70.7% |
| SWE-bench Pro | 64.6% | Claude Fable 5: 80% · Grok 4.5: 64.7% · Claude Opus 4.8: 69.2% |
| Agents' Last Exam | 53.6 | Claude Fable 5 (adaptive reasoning): 40.5 — Sol leads by 13.1 points |
| Artificial Analysis Coding Agent Index (max reasoning) | 80 | 54% fewer output tokens and 57% less time than the next-highest-scoring model |
| Cerebras inference speed (Sol) | up to ~750 tokens/sec | Rolling out on Cerebras wafer-scale compute through July 2026 |
Key takeaway
Sol and Claude Fable 5 split the benchmark record roughly evenly. OpenAI's own tally has Sol leading on Terminal-Bench, BrowseComp, OSWorld, Agents' Last Exam and cybersecurity (ExploitGym); Fable 5 leads on SWE-bench Pro, GDPval-AA v2, FrontierMath, HealthBench Professional and the Artificial Analysis Intelligence Index v4.1. Sol's headline advantage is efficiency and price: Sam Altman told CNBC the model is 54% more token-efficient on agentic coding, at half of Fable 5's per-token input cost.
Pricing and API features
All three variants launched with published, confirmed rate cards — a contrast with the preview period, when pricing was TBD.
| Variant | Input (per 1M) | Cached input | Output (per 1M) | Batch (in/out) |
|---|---|---|---|---|
| Sol | $5.00 | $0.50 | $30.00 | $2.50 / $15.00 |
| Terra | $2.50 | $0.25 | $15.00 | $1.25 / $7.50 |
| Luna | $1.00 | $0.10 | $6.00 | $0.50 / $3.00 |
Cached input carries the usual 90% discount off the uncached rate, with a new wrinkle: cache writes cost 1.25x the uncached input rate, and cached content now has a 30-minute minimum life with explicit cache breakpoints available in the API. Other new API surface: programmatic tool calling, native multi-agent support, and a detail: original option for image inputs.
Notably, Sol's $5/$30 rate card is identical to GPT-5.5 (Thinking)'s pricing — same price, higher published benchmarks on several suites.
Who should use GPT-5.6
- Heavy agentic-coding workloads — Sol's token-efficiency claim (54% fewer output tokens per the Artificial Analysis Coding Agent Index) compounds savings beyond the rate card itself.
- Teams choosing between OpenAI tiers — Terra at $2.50/$15 is the new default recommendation if you don't need Sol's frontier ceiling; it undercuts GPT-5.5 by half at reportedly comparable quality.
- Cost-sensitive, high-volume jobs — Luna at $1/$6 competes directly with Claude Haiku 4.5 and Grok 4.5's output rate.
- Cybersecurity and long-horizon agent research — Sol's strongest published margin over Claude Fable 5 is in this category.
If your workload leans on SWE-bench-style multi-file engineering, deep knowledge work or scientific research, Claude Fable 5 still leads those specific benchmarks, at double Sol's input price.
How GPT-5.6 compares
- GPT-5.6 vs GPT-5.5 — the direct upgrade, at the same price →
- GPT-5.6 vs Claude Fable 5 — the split-decision frontier matchup →
- GPT-5.6 vs Grok 4.5 — near-identical SWE-bench Pro scores, very different prices →
- GPT-5.5 ("Spud") — the model GPT-5.6 replaces as OpenAI's flagship →
- GPT-Live — OpenAI's voice models, announced the day before this GA launch →
Frequently asked questions
Is GPT-5.6 released?
Yes. It became generally available on 9 July 2026 across ChatGPT, ChatGPT Work, Codex and the OpenAI API, after a ~20-partner limited preview that began 26 June 2026 under a US government safety review.
What models are in the family?
Sol (flagship), Terra (balanced — roughly GPT-5.5-class performance at about half the price) and Luna (fast, OpenAI's lowest cost). All three share a 1.05M-token context window and 128K max output.
How much does GPT-5.6 cost?
Per 1M tokens: Sol $5/$30, Terra $2.50/$15, Luna $1/$6. Cached input is 90% off; cache writes cost 1.25x the uncached rate. Batch pricing is half the standard rate.
Why was it restricted before launch?
At the request of the US government, OpenAI gated the preview behind a safety review lasting about 12 days. Restrictions were lifted 8 July 2026, a day before GA.
Is GPT-5.6 better than Claude Fable 5?
It's split. Sol leads Terminal-Bench 2.1 (91.9% vs 88.0%) and Agents' Last Exam (53.6 vs 40.5); Fable 5 leads SWE-bench Pro (80% vs 64.6%), GDPval-AA v2 and FrontierMath. Sol costs half of Fable 5's input rate.
Is GPT-5.6 the same as GPT-6?
No. GPT-6 has not been released; GPT-5.6 is an intermediate family between GPT-5.5 and GPT-6.
- OpenAI — GPT-5.6: Frontier intelligence that scales with your ambition (GA announcement)
- OpenAI — Previewing GPT-5.6 Sol
- OpenAI — API pricing (Sol/Terra/Luna rate cards)
- OpenAI — Model docs (context window, max output, knowledge cutoff)
- Simon Willison — The new GPT-5.6 family: Luna, Terra, Sol
- CNBC — OpenAI's newest model is 54% more token efficient, Altman says
- Axios — Trump administration lifts restrictions on OpenAI's GPT-5.6