GPT-5.6 Sol is out. OpenAI’s newest flagship model went to general availability on July 9, 2026, after a bumpy two-week limited preview that had the U.S. government deciding who got early access. That alone made this a strange launch. Here is what actually matters once you get past the announcement noise.
Direct answer: GPT-5.6 Sol is OpenAI’s top-tier reasoning and coding model, released generally on July 9, 2026, after a restricted preview that started June 26. It costs $5 per million input tokens and $30 per million output tokens, matches GPT-5.5’s price, and edges ahead of Claude Mythos 5 and GPT-5.5 on the Terminal-Bench 2.1 coding benchmark. It ships alongside two cheaper tiers, Terra and Luna, so “GPT-5.6” is really a family of three models, not one.
What Is GPT-5.6 Sol?
GPT-5.6 Sol is the flagship tier in OpenAI’s GPT-5.6 family. Instead of shipping one model with a dial you turn up or down, OpenAI split this release into three separate tiers: Sol, Terra, and Luna. Each one targets a different job.
Sol is built for the hard stuff — long agentic coding runs, security research, science workflows, and tasks where getting the answer wrong actually costs you something. Terra sits in the middle, aimed at everyday production traffic. Luna is the cheap, fast option for high-volume work like classification and routing.
This is a real shift in how OpenAI names and prices its models. The old mini/nano suffix system is gone for this generation. Now you pick a tier by name, and the tier decides your rate card.
Why the Release Was Delayed by the U.S. Government
Direct answer: The Trump administration asked OpenAI to restrict early access to GPT-5.6 to a small, government-vetted group of partners while federal agencies reviewed the model’s cybersecurity capabilities.
OpenAI previewed Sol, Terra, and Luna on June 26, 2026. But before opening things up, the company said it had “previewed our plans and the models’ capabilities” with the U.S. government first. The Commerce Department’s Center for AI Standards and Innovation ran the evaluation, and OpenAI reportedly sent technical staff to Washington to answer questions directly.
The concern centered on Sol’s coding and cybersecurity ability. In testing against Chromium and Firefox, the model found real bugs and exploitation primitives — the building blocks an attacker would need for an exploit — though it did not put together a full working exploit chain on its own. OpenAI says that stopped short of its own “Cyber Critical” threshold, but the company still paired the release with what it calls its most advanced safety stack yet, and layered that with the phased rollout.
For about two weeks, access was limited to roughly 20 trusted organizations. OpenAI was open about not loving this arrangement. “We don’t believe this kind of government access process should become the long-term default,” the company said in its announcement. Sam Altman reportedly told employees internally that OpenAI didn’t see the government-curated list as sustainable long term either.
The restriction lifted on July 9, and Sol, Terra, and Luna are now rolling out globally across ChatGPT, the API, and Codex. Worth noting: Anthropic went through something similar with Claude Fable 5 and Claude Mythos 5 just weeks earlier, so this kind of pre-release government check may be becoming a pattern for frontier model launches, not a one-off.
GPT-5.6 Sol, Terra, and Luna: How the Tiers Differ
Here’s the breakdown, and the pricing is where this gets interesting for anyone budgeting an API bill.
| Tier | Best for | Input price (per 1M tokens) | Output price (per 1M tokens) |
|---|---|---|---|
| Sol | Hard reasoning, agentic coding, security, science | $5.00 | $30.00 |
| Terra | Everyday production traffic, RAG, chatbots | $2.50 | $15.00 |
| Luna | Classification, routing, high-volume drafting | $1.00 | $6.00 |
All three share a roughly 1 million token context window with a 128,000 token max output — so you’re not losing capacity by dropping down a tier. You’re trading reasoning depth for cost.
There’s a trap worth flagging here. The plain gpt-5.6 alias in the API routes to Sol by default. If you’re just swapping model names in existing code without checking, you could end up paying flagship rates for work Luna would have handled fine. Pin the exact tier — gpt-5.6-sol, gpt-5.6-terra, or gpt-5.6-luna — if cost matters to you.
Sol also introduces six reasoning effort levels: none, low, medium, high, xhigh, and max. This is really a cost dial disguised as a quality setting. Higher effort means more output tokens, and output costs five times more than input on Sol. Two identical prompts run at different effort levels can cost several times apart. OpenAI’s own guidance suggests benchmarking your workload one effort level below what you’d normally pick — a lot of tasks hold up fine, and the savings add up fast at volume.
There’s also a new “ultra” mode. Instead of one agent working through a task, ultra runs multiple subagents in parallel on different parts of the same problem. OpenAI positions it for the most demanding agentic work, but it multiplies token spend on purpose, so it’s not something you’d want running by default on routine tasks.
Benchmarks: Does Sol Actually Beat Claude and Gemini?
Direct answer: On Terminal-Bench 2.1, a real-world command-line coding benchmark, Sol Ultra scored 91.9% and base Sol scored 88.8%, both ahead of Claude Mythos 5 and GPT-5.5, which tied at 88.0%.
That 3.9-point gap for Ultra mode is a genuine jump in a field where half-point improvements usually make headlines. Base Sol’s edge is smaller — 0.8 points over Claude Mythos 5 — real, but not something you’d notice in daily use without controlled testing.
On the developer side, gpt-5.6-sol is also reported scoring around 94% on GPQA Diamond (graduate-level science questions built to resist simple lookup) and in the mid-80s on tool-use benchmarks like τ²-Bench. OpenAI also ran the model against SecureBio biology benchmarks and reported meaningful gains there, though it hasn’t published exact numbers yet, which makes that claim harder to verify independently.
Here’s the honest part competitors won’t say: these are OpenAI’s own benchmark results, run under OpenAI’s own conditions. Independent, long-term production testing is still thin this early. Investor Matt Shumer, who tested Sol directly, said it was excellent but that Anthropic’s Fable 5 came out ahead on “almost every task” he tried. Other early testers, including some who work at OpenAI, have been far more enthusiastic. Take the split reviews as a sign to test Sol on your own workload before betting a whole pipeline on the benchmark charts.
Should You Actually Switch to GPT-5.6 Sol?
If you’re already running GPT-5.5 in production, the honest recommendation is to look at Terra first, not Sol. Terra is priced at half of GPT-5.5’s rate and is positioned by OpenAI as matching its quality — that’s a real cost cut if it holds up in your own tests. Sol only earns its premium on the tasks where GPT-5.5 was actually failing you: complex agentic coding, long multi-step reasoning, or security-adjacent work.
For high-volume, low-stakes jobs — tagging, extraction, first-pass drafts — Luna at $1/$6 per million tokens is the one worth testing. It undercuts most comparable fast models on the market right now.
One more practical point: caching. Cache reads get a 90% discount on all three tiers, while cache writes cost 1.25 times the standard input rate, with a 30-minute minimum cache life. If your workflow reuses the same system prompt across many calls, caching alone can shrink your bill more than picking the cheapest tier does.
The Bottom Line
GPT-5.6 Sol is a solid step up for hard agentic and coding work, and the government-restricted preview turned what should have been a routine model launch into a genuinely unusual news story. But Sol isn’t automatically the model to route everything through. Test Terra against your current GPT-5.5 spend first. Save Sol for the jobs that actually need it. And if you’re weighing it against Claude or Gemini, run your own workload — the early reviews are split, and the benchmark charts only tell part of the story.
FAQ SECTION:
Q1: When was GPT-5.6 Sol released? A1: GPT-5.6 Sol was previewed on June 26, 2026, in a restricted rollout to about 20 government-vetted partners. It reached general availability on July 9, 2026, across ChatGPT, Codex, and the OpenAI API.
Q2: Why was GPT-5.6 Sol’s release restricted at first? A2: The U.S. government, through the Commerce Department’s Center for AI Standards and Innovation, asked OpenAI to limit early access while it reviewed Sol’s coding and cybersecurity capabilities. OpenAI complied, then lifted the restriction after two weeks.
Q3: How much does GPT-5.6 Sol cost? A3: Sol costs $5 per million input tokens and $30 per million output tokens, the same rate as GPT-5.5. Terra costs $2.50/$15, and Luna costs $1/$6 per million tokens.
Q4: Is GPT-5.6 Sol better than Claude Mythos 5? A4: On Terminal-Bench 2.1, Sol Ultra scored 91.9% and base Sol scored 88.8%, both ahead of Claude Mythos 5’s 88.0%. But early hands-on reviews are mixed, with at least one tester rating Anthropic’s Fable 5 higher on most real tasks. Test both on your own workload before deciding.
Q5: What is the difference between GPT-5.6 Sol, Terra, and Luna? A5: Sol is the flagship for hard reasoning, coding, and security tasks. Terra is a balanced mid-tier priced at half of GPT-5.5’s rate for everyday production work. Luna is the cheapest and fastest tier, built for high-volume jobs like classification and routing.
Q6: What is the context window for GPT-5.6 Sol? A6: GPT-5.6 Sol has roughly a 1 million token context window with a maximum output of 128,000 tokens. Terra and Luna share the same context window size.
