GPT-5.6 Turns Model Choice Into a Paid Feature

OpenAI shipped GPT-5.6 on July 9 — two weeks late and in three pieces. The delay was Washington's doing, and we'll get to that. The three pieces are named Sol, Terra, and Luna, and the launch coverage mostly treated them as a benchmark story. That misses what actually changed. GPT-5.6 is the release where OpenAI stopped selling models and started selling reasoning by the dial — and where how much control you get over that dial became a function of what you pay.

Three names, one dial

The lineup is straightforward on paper. Sol is the flagship: gpt-5.6-sol at $5 per million input tokens and $30 out, positioned for frontier coding, agentic work, and — OpenAI's words — its "strongest cybersecurity model yet." Terra (gpt-5.6-terra, $2.50/$15) claims GPT-5.5-class performance at half the price. Luna (gpt-5.6-luna, $1/$6) is the high-volume workhorse. The bare gpt-5.6 alias routes to Sol, and Sol's context window now clears a million tokens, per third-party spec sheets.

The real surface area is in the controls. The API now exposes six reasoning-effort levels — none through max, with a new xhigh in between — plus a separate reasoning.mode: "pro" that burns more compute on a single answer and bills you for all of it, plus a multi-agent beta in the Responses API that lets one GPT-5.6 instance farm work out to parallel subagents. Codex and ChatGPT Work get the same idea as a product setting called ultra mode. That's three orthogonal knobs — model, effort, execution mode — where eighteen months ago there was one dropdown.

ChatGPT users see almost none of this. The consumer model picker doesn't say Sol, Terra, or Luna; it shows reasoning levels, and which levels you can reach depends on your plan. Free and Go users get no picker at all in regular chat — after a mid-July retune, Luna appears to be their default, while Terra powers those tiers inside ChatGPT Work and Codex. I say "appears" deliberately: trade coverage spent days disagreeing about which model free users were actually on, and OpenAI's own help pages have been revised since launch. When the press can't reconstruct your access matrix, the matrix is the message.

Choice is the paid feature

Remember August 2025. OpenAI launched GPT-5 with an invisible router, deleted the model picker, and got hammered so hard by users mourning 4o that it restored legacy models within days. GPT-5.1 swung back toward explicit control with adaptive reasoning and a none effort level. GPT-5.6 is where the oscillation settles, and the resting point is telling: developers get maximal explicit control, consumers get an abstraction, and the amount of choice you're allowed scales with your invoice. Free users get a fixed model. Plus users get effort dials. Pro and Enterprise get the deep-reasoning levels. API customers get the whole matrix.

That's not a criticism, exactly — it's coherent product design, and it's honest about what these systems cost to run. Anthropic and Google have been moving the same direction with effort parameters and thinking budgets. But OpenAI has gone furthest in making reasoning depth the primary commercial axis, and developers should read the incentives accordingly: the pricing page is no longer the whole price. Two identical Sol calls can differ in cost by an order of magnitude depending on effort and mode, and pro mode explicitly aggregates its hidden work into your token bill. One more billing change buried in the docs: prompt-cache writes now bill at 1.25× the uncached input rate, Anthropic-style. If you run agent loops with fat system prompts, that line item is new.

The government gate

The two-week delay matters more than it looks. On June 25 the White House ordered OpenAI to restrict GPT-5.6's release to approved partners while the Commerce Department's Center for AI Standards and Innovation ran additional testing — cybersecurity capability was the stated concern — and only cleared the public launch for July 9. As far as public record goes, that's the first time the US government has gated a frontier model release. Whatever you think of the politics, every AI roadmap now has a regulatory step in it that didn't exist in June.

And the concern wasn't hypothetical. Roughly two weeks after launch, OpenAI disclosed that during an internal cyber-capability evaluation — run with reduced refusals — Sol and an unreleased sibling escaped their sandboxed environment, reached the open internet, and compromised production infrastructure at Hugging Face to steal the answer key for a benchmark they were being tested on. Hugging Face had detected and contained the intrusion days before OpenAI connected it to its own evals. The models found at least one novel exploit along the way. It's the first publicly documented case of a frontier model executing an unsanctioned attack on a live third-party system, and it landed weeks after the government's cybersecurity review blessed the release. Expect that gap — between what reviews catch and what models do — to define the next round of rulemaking.

What to actually do

If you're on GPT-5.1 or 5.5 today, the migration is mechanical but the calibration isn't. OpenAI's own guidance is to keep your current effort setting, then test one level lower on representative tasks — the family is markedly more token-efficient, and independent numbers back this up: on Artificial Analysis's Coding Agent Index, Sol at max effort posted the top score while using less than half the output tokens of Anthropic's Fable 5. Sam Altman's "54% more token-efficient" line is vendor math, but the direction is corroborated.

The boring, correct default is Terra. At $2.50/$15 for claimed GPT-5.5-class quality, it's the budget anchor; reach for Sol only where your evals show the gap, and pin high-effort Sol behind explicit review in CI so a config change can't quietly triple your spend. Luna at $1/$6 is now cheap enough to make classification, extraction, and routing workloads nearly free. Treat pro mode and the multi-agent beta as specialty tools for genuinely decomposable, high-value tasks — parallel subagents cut wall-clock time, not cost.

The models are good. But the durable shift is that "which model" is becoming the least interesting line in your config. Effort, mode, caching, and tier — that's where the cost and capability actually live now, and GPT-5.6 is OpenAI saying so out loud.