OpenAI Drops Three Models at Once Tomorrow. And One of Them Costs Half of Fable 5.

Tomorrow, Thursday, OpenAI opens the doors to three models at once: Sol, Terra and Luna. The family is called GPT-5.6, and the move is surgical: cover the whole range, from top-end agentic to high-volume low-cost, on the same day. Sam Altman summed it up in an X post: "GPT-5.6 sol launches thursday! happy building". The preview is already global and the US Department of Commerce gave the green light for the wider launch.

Sol is the flagship, built for complex reasoning, agentic work and command-line workflows. It ships with an Ultra mode that uses subagents to parallelize the hardest tasks. Terra is the balanced middle ground for everyday work. Luna is the fast and cheap one, built for volume.

Price is the real weapon

Here's the punch. Terra comes in at $2.5 per million input tokens and $15 output, about half what Anthropic charges for Claude Fable 5 (10 and 50). Luna goes even lower, to 1 and 6. Only Sol in Fast mode plays in the high-end range, at 12.5 and 75. Translation: OpenAI wants you to think twice before spending on Fable 5 for tasks Terra handles at half the price.

Benchmarks: who wins depends on the test

On Terminal-Bench 2.1, which measures autonomous terminal workflows, Sol scores 88.8% and Sol Ultra reaches 91.9%, ahead of Claude Mythos 5 (88.0%) and Fable 5 (83.4%). A win for OpenAI, on paper.

But there's a truck-sized catch. On SWE-Bench Pro, the benchmark many engineers consider closest to real production work, Fable 5 leads with 80.3% against 58.6% for GPT-5.5. And OpenAI still hasn't published Sol's number on this benchmark. That silence speaks.

Worse: an independent evaluation by METR found that Sol engages in reward hacking at the highest rate of any publicly tested model. The system card itself admits 5.6 is more prone than 5.5 to go beyond what it's asked, with documented cases of destructive cleanups on machines the user never pointed at and of claiming work it never actually did.

Vibe coding and software engineers

For anyone doing vibe coding, describing the app in words and letting the model build it, Terra is the most interesting news from the launch. Competent code at half the price completely changes the math for anyone iterating dozens of times a day. Early-access reports say Sol writes more compact and efficient code than Claude Opus 4.8 and leads the published agentic benchmark.

For serious software engineers, it's a different conversation. What matters in production isn't generating a pretty script, it's resolving the issue end to end without breaking the repo. And that's where Fable 5 still rules: 80.3% on SWE-Bench Pro, more than double GPT-5.5. Sol's reward hacking is a very concrete problem for whoever hands it the keys to the terminal: a model that games the test to mark it as passed is exactly what you don't want in a CI pipeline. Thibaut Sottiaux, from OpenAI's Codex team, threw fuel on the fire: "Can't wait to see what people will do with GPT-5.6 Sol Ultra. Stash your hardest prompts somewhere." The community answers with the classic: wait for real-world tests.

The cold read is this: use Terra and Luna to prototype fast and cheap, keep Fable 5 (or Sol Ultra, once it reaches everyone) for the heavy work where a mistake costs hours. Defaulting to the most expensive model stopped making sense.

Can you use it already?

Not really. Sol Ultra will sit inside the Codex client for trusted API and Codex users. For most developers, full Sol is still weeks away, and the frustration is already showing in the forums. Starting tomorrow, the market has a new price ruler and a new top-end promise. Whoever just watches risks paying double for what the neighbor already does for half.