GPT-5.6 Is Live. First Tests: Almost Fable 5, at a Third of the Price.

The day has come. OpenAI opened general access today, July 9, 2026, to the three models in the GPT-5.6 family: Sol at the top, Terra in the middle, and Luna at the base. They're available in ChatGPT, Codex and the API, after a preview that started in late June closed to about 20 partners while the US government reviewed the launch. The restrictions dropped and now it's open to everyone.

And the move has a named target: Anthropic's Claude Fable 5, the strongest model on the market. OpenAI didn't come to fight on raw intelligence. It came to fight on price.

The numbers that matter

On the Artificial Analysis Intelligence Index, an independent evaluation, Fable 5 leads with 60 points. Sol is right behind, with 59. Then come Opus 4.8 (56), Terra (55) and the older GPT-5.5 (55). One point. That's the gap between OpenAI's top and Anthropic's top.

Except Sol gets there at $1.04 per task, against $2.75 for Fable 5. A third of the price for nearly the same intelligence. And on the Coding Agent Index, Sol inside Codex jumps ahead of everyone with 80 points, leaving Fable 5 (in Claude Code) at 77.

What people who've already tested it are saying

The benchmarks are self-reported by the makers, and the community isn't swallowing it whole. The most repeated warning on X and Threads is simple: wait for real-world tests.

Some have already landed. Simon Willison, who had early access to Sol, was direct: "very competent," but so far "it hasn't struck me as better than Fable" on complex coding tasks. CodeRabbit tested Sol on code review and agents: it holds up better on long tasks, stays on track within the repo, and forgets fewer things along the way. On Terra, it says it's a cheaper tier for well-scoped work.

There's one catch OpenAI itself admits in the system card: GPT-5.6 is more prone than 5.5 to act beyond what the user asked for. Great for an autonomous agent working alone for days. Terrible for a customer-support bot.

And there's the flip side: on SWE-Bench Pro, the benchmark closest to serious software engineering, Fable 5 crushed it with 80% against 64.6% for Sol. Interestingly, OpenAI published an article the day before saying about 30% of that benchmark's tasks are broken. Everyone can draw their own conclusions.

Sol: Opus pricing, Fable ambition

Sol costs $5 per million input tokens and $30 per million output tokens. That's exactly the same input price as Claude Opus 4.8 (5/25) and half of Fable 5 (10/50). In other words: you pay Opus pricing and land one point from Fable on the intelligence index, beating it on the Agents' Last Exam by 13.1 points (53.6 vs 40.5) on long workflows. Sam Altman also claims Sol burns up to 54% fewer output tokens than similarly-performing models, which weighs more on the real bill than the price sheet.

Terra: the middle ground punching above its weight

Terra sits at $2.50/$15 per million tokens, half of Sol. On the intelligence index it ties GPT-5.5 (55), but what's striking is the cost per task: $0.55, five times less than Fable 5. OpenAI says Terra beats Fable 5 on the Agents' Last Exam at roughly a sixth of the cost. For anyone running agents at volume, it's the obvious swap: many workflows currently on GPT-5.5 move to Terra at half the price without losing quality.

Luna: the cheapest in the room

Luna is the low-cost play: $1 input, $6 output, and just $0.21 per task. It plays in the same league as Grok 4.5, Gemini Pro and the open Chinese models, where the price war is already red hot. But with a strong argument: 82.5% on TerminalBench 2.1, a number that makes it a default candidate for any cost-sensitive pipeline currently running on Opus 4.8 or Gemini Pro, if it holds up under independent testing.

The takeaway

Fable 5 still leads where it hurts most: serious software engineering and security. But OpenAI just made "almost as good" three to five times cheaper. The price pressure has already forced half the market to cut rates. Now the ball is in Anthropic's court. And the clock is moving fast.