GPT-5.6 is OpenAI’s June 2026 frontier release — and the first time the company has split a flagship launch into a clearly tiered family of three models: Sol, the flagship for the hardest problems; Terra, a balanced model for everyday production work; and Luna, a fast, low-cost option for high-volume tasks. The naming is deliberate (sun, earth, moon), and so is the strategy: instead of one model at one price, GPT-5.6 lets you route each job to the cheapest tier that can still do it well.
GPT-5.6 initially launched on June 26 as a limited preview for around 20 vetted partners, after OpenAI shared the models and rollout plans with the U.S. government. On July 9, 2026, all three tiers moved to general availability across ChatGPT, Codex and the OpenAI API. The API model alias gpt-5.6 routes to Sol; Terra and Luna use gpt-5.6-terra and gpt-5.6-luna. Context window is 1.05M input / 128K output across all tiers. Knowledge cutoff is February 2026.
The GPT-5.6 family: Sol, Terra, Luna
Each tier targets a different point on the cost/capability curve. Pricing is per 1M tokens (input / output).
| Tier | Positioning | Price (in / out) | Where it fits |
|---|---|---|---|
| Sol | Flagship — hardest problems | $5.00 / $30.00 | Complex coding, cybersecurity, biology, long-running agents |
| Terra | Balanced — everyday production | $2.50 / $15.00 | Support, internal tools, document analysis at high volume |
| Luna | Fast & affordable | $1.00 / $6.00 | Summarization, drafting, routine automation, fallback routing |
OpenAI positions Terra as roughly GPT-5.5-class quality at about half the price, and Luna as landing near GPT-5.5 on several tests while being the cheapest GPT-5.x tier — a meaningful pull on the price/performance frontier even before Sol’s headline numbers enter the picture.
Pricing and caching
Flagship Sol holds the line at GPT-5.5’s price ($5 / $30 per 1M), so the cost story here is downward pressure from Terra and Luna rather than a more expensive top tier. For context, Sol is roughly half the price of Anthropic’s Claude Fable 5 ($10 / $50).
OpenAI also improved prompt caching, which matters most for long-context and agentic workloads that re-send large prompts:
- Cache writes billed at 1.25× the input rate
- 30-minute minimum cache life
- 90% discount on cache reads
As always, model out cost dominates agent budgets — test each tier on representative traffic before quoting customers, since route-specific pricing can differ from the headline preview rates.
Benchmarks: what OpenAI has shown
The clearest published numbers are on TerminalBench 2.1, a command-line automation benchmark:
| Model / mode | TerminalBench 2.1 |
|---|---|
| GPT-5.6 Sol — ultra | 91.91% |
| GPT-5.6 Sol — max | 88.76% |
| Claude Mythos 5 | 88% |
| GPT-5.5 | 83.4% |
A few honest caveats: these are OpenAI-reported figures gathered during a preview, the record 91.91% uses Sol’s most expensive ultra mode, and a single benchmark never settles “which model is best.” Terra is also reported to beat the GPT-5.5 baseline on this harness. Treat the rest of the suite as unconfirmed until independent evaluations land. For a broader, vendor-neutral view of where the frontier sits, see our 2026 LLM landscape guide.
Sol’s new modes: max and ultra
Sol introduces two higher-effort modes that trade tokens and latency for depth:
- max — deeper single-model reasoning for the hardest individual problems.
- ultra — coordinates sub-agents, letting Sol decompose a complex task and work the pieces. This is the mode behind the record TerminalBench 2.1 score.
Both modes cost more per task, so the practical pattern mirrors other frontier models: run the bulk of a workflow on Terra or default Sol, and escalate to max/ultra only on the steps where reliability pays for itself.
Availability and the government preview
GPT-5.6 launched on June 26, 2026 as a limited preview through the API and Codex, restricted to a small group of trusted partners (about 20 organizations) whose participation was disclosed to the U.S. government. The staggering follows a June 2, 2026 executive order directing federal agencies to build a process for assessing new models before wide release — a roughly 30-day window (to about July 2) that was still underway at launch.
OpenAI complied but pushed back publicly within its own announcement, saying it does not believe this kind of government access process should become the long-term default, since it “keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.” Separately, OpenAI said it will offer Sol on Cerebras hardware at up to ~750 tokens/second in July, initially to select customers.
How it compares
- vs GPT-5.5: GPT-5.6 is the direct successor. Sol matches 5.5’s price while pushing agentic/terminal scores higher; the bigger shift is Terra and Luna undercutting 5.5 on price for everyday and high-volume work.
- vs Claude Opus 4.8: Anthropic’s frontier still leads on several of the hardest reasoning and SWE splits; GPT-5.6 counters on terminal-native agents and a cheaper mid/low tier. Cross-check the exact splits before committing a workload.
- vs Gemini 3.1 Pro: Gemini remains the value/multimodal play with very low input pricing; GPT-5.6 Luna narrows the cost gap at the bottom of OpenAI’s lineup.
August 2026 updates
Sol improvements and Luna for free (August 6)
OpenAI updated GPT-5.6 Sol in ChatGPT with improved factual reliability and a new reasoning depth slider that lets Plus/Pro users control how much thought Sol puts into each response.
GPT-5.6 Luna became the default model for free users with unlimited text chats. A new Think button gives free users access to deeper reasoning on harder questions (subject to abuse guardrails). Limits still apply for file uploads, images and other tools.
GPT-5.6-Cyber (August 10)
OpenAI launched GPT-5.6-Cyber, a Sol fine-tune for advanced cybersecurity tasks. On OpenAI’s internal Advanced Cybersecurity Completion Rate benchmark, Cyber scores 95% — vs 57.3% from GPT-5.5-Cyber and just 1.5% from standard GPT-5.6 Sol (with safeguards applied). Access is limited to vetted security vendors and researchers under the Daybreak program.
What’s still unconfirmed
As of August 11, 2026 we have not seen official figures for:
- Full multimodal I/O — GPT-5.5 was natively omnimodal (text, image, audio and video in); the 5.6 launch focused on coding, cybersecurity and biology and did not detail modality support.
- The complete benchmark suite beyond the TerminalBench 2.1 numbers above.
Best for
- Sol: long-horizon coding agents, cybersecurity and biology research
- Terra: high-volume business work — support, internal tools, document analysis
- Luna: fast, low-cost everyday tasks — summarization, drafting, routine automation
Pros & cons
Strengths
- Tiered family (Sol / Terra / Luna) lets you match model cost to task difficulty
- Sol's new "ultra" sub-agent mode set a record TerminalBench 2.1 score (91.91%, OpenAI-reported)
- Terra matches GPT-5.5-class quality at roughly half the price; Luna is the cheapest GPT-5.x at $1/$6
- Improved prompt caching (1.25× writes, 30-minute minimum cache life, 90% read discount)
Limitations
- Context window is 1.05M input / 128K output but knowledge cutoff is Feb 2026 — five months old at GA
- SWE-bench Pro (64.6%) trails Claude Fable 5 (80.3%) and Opus 5 (79.2%) by ~15 points
- Sol's headline numbers rely on its priciest max / ultra modes (more tokens, higher latency)
- Rollout is gated by a U.S.-government staggered-release process
Frequently asked questions
Is GPT-5.6 available to everyone?
Yes. GPT-5.6 moved to general availability on July 9, 2026, after a limited preview (June 26) with ~20 vetted partners. All three tiers — Sol, Terra, Luna — are available via the OpenAI API, ChatGPT (paid plans), and Codex. The API alias `gpt-5.6` routes to Sol.
What are Sol, Terra and Luna?
They are the three tiers of the GPT-5.6 family. Sol is the flagship for the hardest problems (complex coding, security, biology); Terra is a balanced model for high-volume business tasks; and Luna is the fast, low-cost option for everyday work. They are priced $5/$30, $2.50/$15 and $1/$6 per 1M input/output tokens respectively.
What are Sol's "max" and "ultra" modes?
Sol adds two higher-effort modes — "max" for deeper single-model reasoning and "ultra" for coordinating sub-agents on complex tasks. OpenAI reports ultra reached 91.91% and max 88.76% on TerminalBench 2.1, versus 83.4% for GPT-5.5. These modes consume more tokens and add latency, so reserve them for the hardest steps.
How does GPT-5.6 pricing compare to GPT-5.5?
Flagship Sol costs the same as GPT-5.5 ($5 input / $30 output per 1M). The news is downward — Terra delivers GPT-5.5-class quality at about half the price ($2.50/$15) and Luna lands near GPT-5.5 on several tests at $1/$6, the cheapest GPT-5.x tier.
Why is the rollout restricted?
OpenAI says it previewed the models and release plans with the U.S. government and, at the government's request, is starting with a limited partner preview before a broader launch. The step follows a June 2, 2026 executive order on assessing new AI models. OpenAI publicly stated it does not believe this kind of government access process should become the long-term default.