· AI

GPT-6.2 Astra's List Price Fell 20 Percent. Its Real Cost Fell Closer to Sixfold.

The number on OpenAI's pricing page is not the number that will show up on your invoice.

OpenAI shipped two new models on November 17, GPT-6.2 Astra and GPT-6.2 Sol, and the headline most outlets ran with was the price cut: Astra's API rate dropped 20 percent, to $8 per million input tokens and $40 per million output tokens, while Sol held its predecessor's $2 and $10 rates, per OpenAI's own figures as reported by llm-stats.com. That is a real discount. It is also the smaller story. The number that actually determines what a business pays is cost per finished task, and on that measure the cut is running several times deeper than the price list suggests, at least on the benchmarks both OpenAI and independent evaluators have published so far.

My position: cost per task, not price per token, is the metric worth watching in this market, and this release is a clean example of why. A vendor can cut its per-token rate by 20 percent while cutting the cost of getting a job done by far more, because a better model often finishes the same task using fewer tokens. That is the real driver behind the price war between OpenAI, Anthropic, Google, and a wave of Chinese open-weight labs that has pushed average reported inference prices from above $2 to roughly $1.20 per million tokens in a matter of months, according to one market analysis. None of that came from a mandate. It came from four or five labs bidding for the same agentic workloads, which is competition doing exactly what competition is supposed to do.

The skeptic's case, stated fairly. A lot of the cost-per-task numbers in this launch are self-reported, drawn from a single test run on a single benchmark, and should be read that way. Artificial Analysis's own Intelligence Index, which is independently run and aggregates nine separate tests, scores Astra at 62 against Claude Fable 5.2's 65, meaning OpenAI's new model is still behind on raw capability by that measure, per Decrypt's review of the launch. Astra's score on Artificial Analysis's Coding Agent Index did not move at all, landing at 53, identical to the model it replaced, according to Wccftech's coverage. Even the press got the cost math wrong in the retelling: that same Wccftech piece calls Astra "about 8x cheaper" than Fable 5.2 on the Intelligence Index, citing $4 per task against $26.50. Divide those two numbers and you get 6.6x, not 8x. If outlets covering this launch cannot do the arithmetic on their own quoted figures, a buyer should not accept any cost-cut claim at face value, including mine.

Where the skepticism runs out. Even using the more conservative, independently measured number, a roughly 6.6x drop in cost per task for a 3-point loss on an aggregate intelligence score is a real trade, not a rounding error, and it shows up outside OpenAI's own marketing chart. On OpenAI's benchmark data, the pattern holds across most of the tasks it published: Agents' Last Exam v1.1 cost fell from $8.99 to $2.89 per task while accuracy rose 5.2 points; Terminal-Bench-Science 0.1 fell from $23.80 to $7.15; LifeVal fell from $9.04 to $4.23. The one exception in OpenAI's own table is SecureSWE, where Astra's cost per task actually rose, from $3.21 to $5.26. That exception matters more than the headline number. It means "cheaper" is not a property of the model. It is a property of the specific task you are running through it.

What this means for a business buying model access. Stop comparing vendors on the price-per-million-tokens line of their pricing page. Price a representative sample of your own actual tasks through the candidate models and compare the cost to finish each one, because that number can move in the opposite direction from the sticker price, as SecureSWE just demonstrated inside a single release. Route the routine 90 percent of your workload to whichever model clears your accuracy bar at the lowest cost per task, and reserve the expensive model for the work where the accuracy gap is the thing you are actually paying for.

That comparison, done against your own workload rather than a vendor's demo benchmark, is the core of the model and vendor selection work MojoAI does for clients before they sign an API contract. If your team is choosing between model vendors on the strength of a press release, that is worth a second look before the invoice arrives.

Sources

References used in this article. Links also appear alongside the relevant claims.

Let's talk

Let's make it happen.

You don't need a polished brief. A couple of lines about where you want your company to go is plenty, and we'll come back with what we'd tackle first.

We scope the work around your goals and timeline before quoting anything, so you know exactly what you're getting.

LocationBoca Raton, Florida
CoverageSouth Florida + remote nationwide
Status Now accepting clients