GPT-5.6 Sol, Terra & Luna: What Developers Need to Know
July 6, 2026
OpenAI launched its biggest model family of 2026 on June 26, and most developers can't touch it yet. GPT-5.6 ships as three models — Sol, Terra, and Luna — sitting behind a US government access gate while OpenAI negotiates a broader rollout.
Here's what's in each tier, what it costs, and what to do when access opens.
Three Models, Three Jobs
Sol is the flagship — built for hard coding, long-horizon agentic work, and security research. Priced at $5 input / $30 output per million tokens, same sticker as GPT-5.5 but OpenAI claims substantially better performance. On Terminal-Bench 2.1, Sol Ultra scores 91.9% — above Claude Mythos 5's 88% and Claude Opus 4.8's 74.6%.
Terra is the everyday workhorse at $2.50/$15 per million tokens — roughly half the cost of GPT-5.5 for similar general-purpose performance. The tier to watch for replacing GPT-5.5 in production pipelines.
Luna is the cheap speed tier at $1/$6 per million tokens. Strangely, it outscores Terra on Terminal-Bench 2.1 (84.3% vs 82.5%) — likely because it was optimized for throughput rather than balanced general tasks. For cost-sensitive agent fleets, benchmark Luna against Gemini 3.5 Flash and Claude Haiku before committing.
New Reasoning Controls and Cache Pricing
All three tiers support a reasoning effort parameter — dial it from low to max to trade latency and cost for depth. Sol gets an extra mode called ultra, which spawns subagents to run complex tasks in parallel — OpenAI's answer to multi-agent orchestration, baked into the model runtime.
Cache pricing changed. Writes are now billed at 1.25x the uncached input rate; reads still get the 90% discount. There are also explicit cache breakpoints and a 30-minute minimum cache lifetime. Run the math on your workload before assuming caching saves the same amount as before.
Why It's Gated
All three models score "High" under OpenAI's Preparedness Framework for cybersecurity and biological/chemical capabilities — enough for the US government to require a staged rollout. Roughly 20 government-approved organizations have access now, via API and Codex only. ChatGPT doesn't have any of them.
OpenAI was direct about its view in the release docs: "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." General availability is expected mid-to-late July 2026.
It's the same pattern we saw with Claude Fable 5's government-ordered suspension — frontier models are now regulated assets, and access can shift based on policy with zero warning.
What to Do Now
If you don't have preview access: keep building on GPT-5.5, Claude Sonnet 5, or Fable 5 — but wire in a model router so you can swap Sol in the day the API opens. For high-volume workloads, plan a Terra migration path; the 50% cost cut at comparable performance is worth the integration work. For cheap agent fleets, evaluate Luna against the existing cheap-tier options once GA lands.
Sources: OpenAI, VentureBeat
Frequently Asked Questions
What are GPT-5.6 Sol, Terra, and Luna?
GPT-5.6 is OpenAI's 2026 model generation in three tiers: Sol for maximum intelligence and hard agentic work, Terra for balanced everyday tasks at half the cost of GPT-5.5, and Luna for fast high-volume work at the lowest price. All three are currently restricted to about 20 government-approved partners.
How much does GPT-5.6 cost?
Sol is $5 input / $30 output per million tokens. Terra is $2.50 input / $15 output. Luna is $1 input / $6 output. Cache writes are now charged at 1.25x the uncached input rate; cache reads still get a 90% discount.
When will GPT-5.6 be publicly available?
OpenAI says broader access is coming in 'the coming weeks' from the June 26, 2026 preview. Mid-to-late July 2026 is the most widely expected window for API general availability, though no firm date has been announced.
What is GPT-5.6 Sol's ultra mode?
Ultra mode is exclusive to Sol. It spawns subagent processes that work in parallel on different aspects of a complex problem, similar to a coordinated multi-agent system — but baked into the model runtime rather than requiring external orchestration.