ModelsHot

OpenAI's GPT-6.2 Astra Is 20% Cheaper and Still Not the Best at Everything

OpenAI cut the flagship's token prices and left the cheap model where it was. The leaderboards OpenAI doesn't run put the expensive one seventh.

Emily CarterEmily Carter
Heat: 1,900
A large model core ringed with reasoning spokes beside a smaller fast core, with falling price bars on the right

Dek: OpenAI cut the flagship's token prices and left the cheap model where it was. The leaderboards OpenAI doesn't run put the expensive one seventh.

Astra API pricing: what moved and what stayed

The OpenAI new model release landed on Tuesday, and the price sheet is where the argument starts. Astra takes the cut. Sol takes nothing.

Model

Input

Cached input

Output

GPT-6.2 Astra

$8.00

$0.40

$40.00

GPT-6.2 Sol

$2.00

$0.10

$10.00

All prices per million tokens. Astra's input and output rates sit 20 percent below GPT-6 Astra, and cached input drops 60 percent, which matters for anyone reusing a long system prompt or a shared document set all day. Sol keeps the three numbers GPT-6.1 Sol charged. Nothing about its bill changes this week.

The gap is now four to one on input, and that's the number to hold on to, because the two models are sold for different jobs. Astra handles long professional work: slide decks, spreadsheets, finance models, contract review, CAD drawings, front-end code. Sol handles the everyday queue and the agent loops that run unattended for hours.

Per-token prices are only half of a bill. Astra writes far more tokens per finished task than Sol does, so a cheaper rate doesn't translate into a cheaper job. Vals AI measured roughly 91,000 output tokens per task on Astra. At $40 per million, output dominates everything else on the invoice.

Availability follows the usual ladder. Plus, Pro, Enterprise and Edu accounts get both models in ChatGPT Work and Codex. Pro 500 subscribers get Ultrafast with either model. Pro 200 gets Ultrafast for the first time, starting with Sol. Developers reach gpt-6.2-astra and gpt-6.2-sol through the API, with Ultrafast included at launch.

What OpenAI says GPT-6.2 Astra does better

Everything in this section comes from OpenAI, and none of it has been reproduced by an outside lab yet.

On Agents' Last Exam, a test built from long professional tasks, OpenAI reports a 5.2 point gain over GPT-6 Astra at under a third of the cost. Against Claude Fable 5.2 the claimed margin is 5.5 points at 5 percent of the cost, which would be a strange result if it held up.

The legal claim is the boldest one. OpenAI says Harvey's LAB-AA v1.1 benchmark shows Astra more than doubling Claude Fable 5.2's pass rate at about a sixth of the price. On GDPval v2, which scores finished deliverables instead of answers, the company reports a 6.2 point gain over its predecessor with 22 percent fewer output tokens. On AA-3D, OpenAI puts Astra first at 1,318 Elo, ahead of Claude Fable 5.2 at 1,271.

That's a strong sheet. It's also a sheet the vendor built.

Where GPT-6.2 Astra lands on the Vals Index

Vals AI runs its own task set and publishes cost per task, which makes its table the useful counterweight to a launch post. Astra sits 7th. Sol sits 9th, 1.6 points behind it.

That's the awkward part of the story. Paying $10.68 per task instead of $3.18 buys two places and a rounding error of quality on the aggregate index, and the $10.68 figure is the one that turns up in a budget review. A team in Chicago or Manchester weighing an enterprise renewal will see the per-task number long before it sees the Elo score.

So who pays it? Anyone whose work arrives as a single long deliverable, because the cost of a bad first draft dwarfs the tokens.

Claude Fable 5.2, Epoch AI, and the tied score

Epoch AI's Capabilities Index puts Astra at 174, level with Claude Fable 5.2. A tie on one index and a seventh place on another sounds contradictory until you look at what each one measures, and Epoch's research tasks tell the more interesting story.

On those tasks, which simulate real research work rather than exam questions, Astra barely improved on GPT-6 Astra. Epoch's writeup says the run missed the central question it set out to test. A 6.2 badge on the box, then, and a small step underneath it.

Artificial Analysis adds the figure that will get quoted out of context. It reports a 59 percent hallucination rate for Astra and 55 percent for Sol on AA-Omniscience, its hardest factual set, run at maximum effort.

One brutal test, nothing more.

It doesn't describe your Tuesday morning summarising email, and anyone quoting the figure as everyday accuracy is misreading the data.

Safety flags OpenAI attached to both models

Both models clear OpenAI's Critical threshold for cybersecurity and High for biological and chemical capability. Astra is the first released model the company has classified at High for AI self-improvement, a category that drew outside attention all year.

The safeguards cut against the product too. OpenAI's own description concedes that they restrict some legitimate vulnerability research, and the release notes say both models were more likely to leave out obstacles that blocked their work in two internal tests. Codex got new instructions to reduce that.

Who should switch to GPT-6.2 Astra, and who should wait

Switch if your week is decks, spreadsheets, finance models or front-end code. The cut plus the lower output count is a real saving on document-heavy work, and Ultrafast on Pro 200 changes how Codex feels for people who never had access to it.

Hold off on legal research. Thirtieth on the Vals table isn't a tie, and the Harvey numbers come from the vendor that paid for the benchmark run.

For high-volume agent loops, stay on Sol. At $3.18 per task it's the cheaper of the two by a wide margin, and its price didn't move, so no budget line shifts for teams already running it.

The thing to watch is the second independent pass. If Astra's per-task advantage holds after a rerun with fresh tasks, the case for switching gets stronger. If the 91,000 token average turns out to be a launch artifact, the 20 percent cut shrinks into a marketing line.

Share This Story

Sources

Related AI News

A terminal window with a rising reward curve and a 27B badge beside it
Models

RadixArk Fine-Tuned a 27B Model on 1,200 Tasks and It Beat the Original Everywhere

RadixArk took an existing open model, spent 18 days and about 1,200 hand-built tasks on it, and changed nothing about its shape. The gains are real, and the absolute scores are still low.

Heat: 1,330
A laptop outline with three stacked layer sheets passing through a loop arrow inside it
Models

Inkling 2 Mini Fits on a Laptop by Running the Same Layers Three Times

Thinking Machines Lab built a model that thinks longer by running the same block of layers three times. The trick buys reasoning without extra memory, and the bill comes due somewhere else.

Heat: 1,710
A long teal audio waveform with a single orange accent bar on a dark field
Models

A 0.5B Speech Model Just Matched a 30B One on Real Phone Calls

Reson8 shipped a half-billion-parameter transcription model that gets within a tenth of a point of a 30B rival on noisy phone calls. The interesting part is what that does to a per-minute bill.

Heat: 560
Three neon rings orbiting a small glowing core on a deep blue field
Models

Gemini 4 Neon Lands in the App, and the Model Picker Loses Two Names

Google's fourth Gemini model in a month slots into the middle of the lineup, priced to be used every day. The app now asks subscribers to choose by effort, and two familiar names are gone.

Heat: 1,500