GPT-5.6 Is Here — Same Price as GPT-5.5, Twice the Brain

OpenAI quietly dropped something interesting this week. Three new models under the GPT-5.6 umbrella: Sol (flagship), Terra (cheap sidekick), and Luna (fast and light). Sam Altman called Sol “a significant step forward” and was quick to point out it’s the same price as GPT-5.5. Terra, meanwhile, delivers 5.5-level performance at half the cost.

So what’s actually new?

The Headlines

Sol is built for agentic coding — the kind of task where you’re not just asking one question and getting an answer, but where the model needs to plan, switch contexts, and chain multiple steps together without hand-holding. OpenAI also pushed hard on biology workflows and cybersecurity, which is why the US government asked for a controlled rollout. Sol isn’t available worldwide on day one.

The interesting part: Sol can now use subagents to tackle problems too big for a single-agent setup. You can also dial up a “max reasoning effort” knob for deeper problem-solving on hard tasks.

Benchmarks tell part of the story. Sol set a new state of the art on Terminal-Bench 2.1 (coding tasks) and posted stronger GeneBench v1 scores than GPT-5.5 while using fewer tokens. That’s the efficiency angle Altman was pointing at.

The Three Models

The lineup isn’t just Sol and a couple of siblings — it’s three tiers designed for different budgets:

• GPT-5.6 Sol — $5M input / $30M output. The heavy hitter.
• GPT-5.6 Terra — $2.50M input / $15M output. Half the price of Sol, same performance tier as GPT-5.5. That’s a real downgrade in cost for teams that don’t need the absolute top.
• GPT-5.6 Luna — $1M input / $6M output. The cheap-and-fast option.

There’s also a caching play here: GPT-5.6 adds explicit cache breakpoints with a 30-minute minimum cache life, and cache reads come at a 90% discount on the cached-input rate. If you’re running repeated queries on the same context, that’s meaningful.

The Cybersecurity Angle

This is where things get complicated. OpenAI classified Sol, Terra, and Luna as High capability in both Cybersecurity and Biological/Chemical risk under its Preparedness Framework. Not High in AI self-improvement or Cyber Critical — just High in the dual-use categories that make governments twitchy.

In browser exploit testing (Chromium and Firefox), Sol identified bugs and exploitation primitives. It didn’t autonomously produce a full-chain exploit — but the fact that it got close is why the rollout is staged rather than a big bang.

OpenAI ran over 700,000 A100-equivalent GPU hours on automated red teaming before launch. That’s not casual.

What This Means for You

If you’re paying for GPT-5.5 API access right now, Terra is the thing to watch. Half the price, same performance tier — that’s a real operational cost reduction for high-volume applications. If you’re on ChatGPT Plus, Sol access should be coming to your account in the next few weeks as the rollout widens.

If you’re building agentic workflows — multi-step coding tasks, research pipelines, automated analysis — Sol’s subagent architecture and deeper reasoning mode are worth testing against your current setup.

And if you’re in biology or security research, the capability ceiling here is real. The safeguards are real, too, but they’re not a wall — they’re a gate that sometimes asks you to wait.

The model drops today in the US and UK. Broader access follows in weeks.