OpenAI released GPT-5.6 today — not as one model but as three: Sol, the flagship; Terra, the mid-tier; and Luna, the cheapest. OpenAI's framing: "more intelligence from every token, stronger performance per dollar, and more capability on demand for your hardest work." It's available across ChatGPT, Codex, and the API.

The launch page leads with vendor benchmarks — a couple of claimed new highs, plus an efficiency claim worth noticing: on one independent index, OpenAI says the flagship comes within a point of the leader while completing tasks in "61% less time at roughly half the estimated cost." Treat all of it the way you'd treat any vendor's launch numbers: directionally interesting, independently unverified on day one. The durable story is the pricing table.

The menu is the message

Sol runs $5 per million input tokens and $30 per million output. Terra is half that. Luna is $1 and $6 — a fifth of the flagship's price.

Three tiers of the same generation, priced five-to-one apart, is OpenAI saying something businesses should already know: most work doesn't need the biggest model. Drafting a follow-up email, extracting fields from an intake form, tagging inbound requests — that's Luna work. Analyzing a messy contract or debugging a gnarly integration — that's where Sol earns its premium. Paying flagship prices for routine work is the AI equivalent of sending every package overnight.

This maps exactly to a concept we use in every engagement: Solution Levels. Not every use case takes the same effort — or, now, the same model. The discipline of matching the tool to the task is where the savings actually live, and the vendors have just made that discipline explicit in their price lists.

What GPT-5.6 is good at, per OpenAI

The claimed strengths cluster around knowledge work at lower cost: coding, science, and — notably — cybersecurity. The efficiency claims are the ones worth watching: less time and fewer tokens for the same outcome means the cost of a unit of AI work fell again, whichever vendor's numbers you trust.

That trend line — capability per dollar rising every quarter — is the planning assumption we'd build on. Whatever was marginally too expensive to automate in your business last year probably isn't anymore.

What to do with this

  • Audit your model spend. If everything in your stack routes to a flagship model by default, a routing pass (cheap model first, escalate when needed) often cuts AI costs meaningfully with no quality loss anyone notices.
  • Re-run the math on parked ideas. Anything shelved for cost reasons deserves a second look at Luna-tier prices — through the same three-question filter as always.
  • Don't switch vendors over a launch page. Benchmarks move week to week; switching costs are real. Pick per task, measure on your own work, and stay boring about it.

If you want help deciding which tier — or which vendor — your actual workload needs, that's a 30-minute conversation.

Source: GPT-5.6: Frontier intelligence that scales with your ambition — OpenAI, July 9, 2026