The $1 model beat the $2.50 model on OpenAI's own launch chart. Read that number first.

On Thursday, thirteen days after it began letting a government-vetted list of partners try its new models, OpenAI opened GPT-5.6 to everyone — Sol, Terra and Luna, live across ChatGPT, the API and Codex. Within hours the same chart was everywhere: Terminal-Bench 2.1, with Sol's "Ultra" configuration at 91.9 percent, a new high-water mark, up and to the right. It is a genuinely striking number, and I want to be clear before I take it apart that some of what launched on Thursday is real and checkable. But the chart everyone reposted also contains a smaller, quieter number that almost nobody led with: Luna, the $1 budget model, scored 84.3. Terra, the $2.50 mid-tier model — the one OpenAI positions above it — scored 82.5. The cheap model beat the expensive one on the launch benchmark. That footnote is the most honest thing in the whole announcement, and it is worth reading before the big number, because it tells you what these scores are actually measuring.

First, the claims, stated precisely, because auditing a number starts with pinning it down. GPT-5.6 ships as three models. Sol is the flagship, priced at $5 per million input tokens and $30 per million output — the same list price as GPT-5.5 before it. Terra is the mid-tier at $2.50 and $15, which OpenAI describes as competitive with GPT-5.5 at half the cost. Luna is the new floor at $1 and $6, a budget tier below anything in OpenAI's current production lineup. On Terminal-Bench 2.1, OpenAI reports Sol Ultra at 91.9 percent, standard Sol at 88.8, Luna at 84.3 and Terra at 82.5, against 88.0 for GPT-5.5. Around the edges sit the softer claims: a reported 54 percent improvement in token efficiency on agentic coding, an unconfirmed context window in the neighbourhood of 1.5 million tokens, serving speeds of 750 tokens per second on Cerebras hardware, and estimated scores on SWE-bench Verified and FrontierMath that OpenAI has not formally published. Keep those categories separate — priced, reported, estimated — because they are about to matter.

Who was allowed to check

A launch benchmark is a claim about a comparison, and the first question to ask of any comparison is who ran it. Here the answer is unusually clean: OpenAI did, and for the entire preview period, almost nobody else could. GPT-5.6 spent its thirteen days before Thursday inside a government-coordinated preview — the arrangement, reported by The Information and confirmed in outline by CNN and Axios, in which federal offices approved access customer by customer after all three variants crossed into OpenAI's "High" capability band for both biological and cyber risk. The initial cohort has been described in reporting as roughly twenty government-vetted organisations. Whatever you think of that gate as policy — my colleague Sam Brenner has written about what it does to the market — consider what it does to measurement. The organisations chosen to see this model early were selected for security clearance, not statistical hygiene. They were enterprises and testing bodies, not the leaderboard maintainers, academic eval groups and independent measurement shops that normally spend a launch week re-running a vendor's claims.

So on the day the public finally got access, every number in circulation traced back to one source: the party selling the model. That is not an accusation of dishonesty. It is a statement about error bars. A vendor-run benchmark is a single lab's result, on a test the lab chose, in a configuration the lab tuned, published without variance. The ordinary corrective — a dozen independent groups re-running the suite within days and quietly flagging what doesn't replicate — was structurally delayed here, because the preview design kept precisely those people out. The re-runs will now happen; the first independent numbers should land within days of public API access. Until they do, the correct status of every performance claim in Thursday's launch is not "false" but "unreplicated," and the history of unreplicated launch numbers is not flattering.

The footnote that scrambles the ladder

Now the number I promised. Luna at 84.3, Terra at 82.5. OpenAI's product ladder says Terra sits above Luna: it costs two and a half times as much per input token, and the marketing positions it as the balanced model for everyday work, with Luna as the fast, cheap one. On the launch benchmark, the ladder inverts. There are two readings of this, and both are more informative than the headline score.

The first reading is statistical, and it is the one I would put money on: a 1.8-point gap on a single benchmark, reported with no variance, no confidence interval and no stated number of runs, may not be a gap at all. Terminal-Bench is a finite suite of tasks; scores on it move by a point or two between runs of the same model depending on temperature, harness and luck. If OpenAI ran each model once and printed the result, Luna-beats-Terra could be noise wearing a ranking. The problem is that the same is then true of every other 1.8-point gap on the chart — including some of the ones doing promotional work. You do not get to keep the noise when it flatters the flagship and discard it when it embarrasses the mid-tier.

The second reading is worse for the marketing and better for your understanding: the gap is real, and it reveals what the tiers actually are. Model tiers are priced by their cost to serve and positioned by their average performance across a broad mix of tasks — OpenAI's own commentary around the launch says as much, noting that the tiers prioritise balanced performance rather than dominance on any specific benchmark. A smaller model distilled and tuned hard on agentic terminal tasks can absolutely beat a larger sibling on exactly those tasks while losing to it broadly. Which means the tier names are a pricing structure, not a capability ranking — and a benchmark chart organised by tier is a price list wearing a lab coat. Either reading should change how you buy: test your own workload on all three, because the ladder printed on the pricing page demonstrably does not predict the ordering on at least one axis OpenAI itself chose to publish.

A 1.8-point gap on a single vendor-run benchmark, with no variance reported, is not a ranking. It is a coin that has been flipped once.

Fifty-four percent more efficient, measured how?

The claim doing the most quiet work in this launch is not a benchmark score at all. OpenAI's headline efficiency claim — reported as a 54 percent improvement in token efficiency on agentic coding — is, if true, the most economically significant number on the slide, because token efficiency compounds with price. A model that completes the same task in roughly half the tokens at the same list price is a price cut that never appears on the pricing page. Stack it: Terra at half of GPT-5.5's list price, times a meaningful efficiency gain per task, and the effective cost of a unit of agentic work has fallen a lot further than the tier prices suggest. That is presumably the story OpenAI wants told.

So ask the auditing questions. Fifty-four percent fewer tokens to complete what, exactly? Efficiency on agentic tasks is a ratio — tokens spent per task completed — and both the numerator and the denominator are choices. Which task suite? Completed to what standard, judged by whom? Averaged how — per task, per token, weighted by task length? A model can look dramatically more efficient by abandoning hard tasks early, by emitting terser reasoning that fails more often (which then costs you retries the metric never sees), or by being measured on a task mix that plays to its training. None of the reporting around the launch answers these questions, and OpenAI has not published the methodology. The claim is plausible — efficiency genuinely is where frontier labs are competing this year, and the 90-percent prompt-caching discount and the reported 750-token-per-second Cerebras serving path point the same direction. Plausible and verified are different filing cabinets. This one stays in the first.

The claim you can actually check today

Here is the part of Thursday's launch that needs no replication study, because it is falsifiable by your next invoice: the prices. A benchmark slide is marketing; a pricing page is a commitment. And read as a set of prices rather than a set of scores, GPT-5.6 is genuinely significant news.

• Terra offers what OpenAI itself frames as GPT-5.5-competitive performance at $2.50/$15 — half the incumbent flagship's list price, a claim your own workload can confirm or refute this week.
• Luna opens a production tier at $1/$6 that undercuts OpenAI's previous floor and lands directly on the territory of every budget-model competitor.
• The competitive spread is now stark: Anthropic's Claude Sonnet 5 sits at an introductory $2/$10, Claude Fable 5 at $10/$50, and the market's full range of per-million output pricing runs from under a dollar to fifty.
• Sol holds GPT-5.5's exact list price at $5/$30 — meaning OpenAI's claimed generation-over-generation gain arrives at zero nominal price increase, before any token-efficiency effect.

That is deflation at every tier, and it is the kind of claim I like, because it does not depend on trusting anyone. If Terra fails to do your GPT-5.5-grade work, you will know by Friday and so will every developer forum on the internet. If Luna is a toy, the $1 price signs its own confession. Prices are the rare launch-day numbers that come with built-in, distributed, adversarial verification — every customer is the error bar. It says something about the state of AI communication that the pricing table, the least glamorous artifact in the announcement, is also the only one published to a standard where being wrong has consequences.

Compared to what?

Every audit needs a baseline, so put Thursday's launch next to its competition. The same day GPT-5.6 went public, SpaceXAI pushed Grok 4.5 to general availability carrying an "Opus-class" performance claim supported, as of launch day, by no published benchmarks, no system card, no pricing detail and no context-window specification at all. My colleague Amy Mercer wrote last week about what it means that Grok's claims rest on internal evals no outsider can run; I will just note the comparison it sets up. Google's Gemini 3.5 Pro, meanwhile, remains in preview, past its reported June target. Against that field, OpenAI's launch is — and I want to say this plainly, because credit is part of the method — the most documented frontier release of the summer. Published tier pricing. A named public benchmark with per-model scores, including the score that embarrasses its own product ladder. A capability classification, disclosed. The bar this clears is on the floor, but it clears it, and the difference between "vendor-run numbers you can soon check" and "vibes" is a real difference.

The pattern worth watching is the one both launches share: the gap between announcement day and audit day is widening. GPT-5.6's preview structure delayed independent measurement by design; Grok 4.5 delays it by omission; and each week that a frontier model's only scorecard is its maker's, the industry gets a little more comfortable treating vendor claims as results. The fix has not changed since I started writing this column. Held-out test sets. Published methodology. Variance on every bar. Independent re-runs before the victory lap. It is boring, it does not screenshot well, and it is the entire difference between measurement and advertising.

What Thursday's numbers actually support

So, the ledger. Supported today: GPT-5.6 exists, is publicly available at every tier, and is priced aggressively enough to move the whole market's cost curve — that much is contractual. Supported with an asterisk: Sol's Terminal-Bench family of scores, vendor-run, single-suite, no variance, awaiting the independent re-runs that public access finally makes possible. Not yet supported: the 54 percent efficiency claim (no methodology), the 1.5-million-token context window (unconfirmed), the estimated SWE-bench and FrontierMath figures (unpublished), and any sentence containing the phrase "best model in the world." And quietly instructive: Luna over Terra, 84.3 to 82.5 — the footnote that tells you the tiers are prices, not rankings, and that the only benchmark that will ever rank these models for your purposes is the one you run on your own work. The chart goes up and to the right. It usually does. Look harder anyway — and this week, for once, everyone finally has the access to do exactly that.