How to read an AI model announcement: benchmarks, pricing and availability

A model announcement is a press release wearing a lab coat. It is written by the vendor, about the vendor, on the vendor’s happiest day, and it follows a template so stable you can learn to read it once and reuse the skill on every launch: a superlative headline, a benchmark table, a pricing line, an availability paragraph. This guide takes the template apart using a current example, OpenAI’s GPT-6 Astra announcement of 3 September 2026, and shows which parts you can verify the same day.

One term up front: a benchmark is a fixed exam for models, a published set of tasks that different models can be scored against so the results are comparable.

The benchmark table is the seller marking its own exam

The table is real data presented by an interested party. The vendor chose which exams to sit, which scores to print, and which rivals to sit beside, so read it for its choices as much as its values.

Look first at what is being compared. A table that measures the new model against the vendor’s own previous release will show progress almost by definition. A table that includes rival models is bolder, but check the fine print for which version of each rival was used and when.

Look next at the footnotes. Wording such as “with tools”, “best of several attempts” or “with extended thinking” means the score was achieved under a setup you may not get by default, and the same footnote may not apply to the rival’s column. And notice the missing rows: a benchmark the vendor quoted proudly at the last launch and omits at this one is usually a score that did not move. None of this makes the table false. It makes it an opening claim, priced accordingly, until independent results arrive in the weeks after launch.

The pricing line is the only sentence with a unit

Somewhere below the charts sits the sentence that can be checked the same afternoon, because it has a unit: dollars per million tokens. Our guide to how model pricing works explains the meter itself; on launch day, what matters is placing the announced figure against the vendor’s existing list, which the announcement will not do for you.

The worked example: OpenAI’s GPT-6 Astra announcement leads on capability, and the rate card it points to prices the model at $10 per million input tokens and $50 per million output tokens, with cached input at $1. Read the row to the end, though, because the same page carries a second pair of columns for long context, where the model’s input line becomes $20 per million and its output line $75. A headline rate is the cheapest way the vendor can describe itself truthfully.

Set against the OpenAI models our pricing ledger tracks, read the same day, that $10 sits at the top: the cheapest OpenAI models in our ledger are $0.20 per million input tokens. The vendor’s full page is wider in both directions than any subset of it, running from $0.05 per million on its smallest current model to $150 on the dearest reasoning model it still sells, so the honest comparison is the one the chart above draws, between the flagship and the models a reader would otherwise pick. Neither figure is a secret; the announcement simply has no reason to draw the comparison, and the ledger exists because somebody should.

The pricing line also tells you who the launch is for. A flagship price signals a model aimed at businesses and heavy tools, not at the casual subscriber, whatever the headline promises everyone.

The availability line tells you who actually gets it

The last paragraph decides whether the announcement matters to you this month. It answers, usually in this order: is the model in the API, the by the token service developers buy, from today; when does it reach the consumer app, and which paid tiers see it first; which countries wait; and is it labelled preview, meaning the vendor reserves the right to change or withdraw it without the usual notice.

The rollout order is information, and reading it as strategy is our inference rather than anything the vendor states. API first means developers are the priority audience. Dearest tier first tells you the vendor is using scarcity to sell upgrades, a pattern our plans comparison records across vendors.

Where that scarcity becomes a number, read the number carefully. OpenAI’s Codex pricing page prints a table of local messages per five hour period, and the GPT-6 Astra row on the Plus plan reads 5 to 45. Those are Codex messages drawn from the ChatGPT plan allowance rather than a general cap on chat, which is the sort of distinction a launch post never makes and our plans ledger keeps in the note beside the row.

Then read the quiet companion lines. A launch is also when older models get their retirement dates, but the notice itself lives in the vendor’s deprecation policy rather than the announcement: OpenAI publishes at least six months before retiring a generally available model, at least three months for specialised variants such as its Codex and deep research models, and as little as two weeks for anything labelled preview, with affected customers told by email and in the documentation.

What it means for you

A routine for the next launch day, in the order the page should be read: check the availability paragraph first to learn whether this reaches anything you use, and when. Check the pricing line against the vendor’s own list, or against our dated ledger, to see where the model really sits in the range. Take the benchmark table as the seller’s opening claim and wait for independent scores before repeating it. And if you subscribe to the vendor’s consumer product, watch what its published allowance table says in the following days, because that page, not the announcement, is the commitment your money actually buys.