OpenAI has released GPT-6.2 Astra, its new model for the hardest tasks, together with the lighter GPT-6.2 Sol. Astra's main advance is efficiency: it finishes comparable work with far fewer output tokens, which pushes the cost of a task down sharply, and it is notably strong at 3D and design work. It does not take the top spot on overall benchmarks, its coding scores did not move, and early testers reported real friction in everyday use. OpenAI also says Astra is its first publicly released model to meet a "High" threshold for AI self-improvement, a phrase that needs a careful reading.

What OpenAI released

Astra targets demanding frontier work and is positioned against Anthropic's Claude Fable 5.2. Sol is the efficient everyday model and competes with Claude Opus 5.5. Both are available now in ChatGPT Work, in Codex for paid tiers (Plus, Pro, Enterprise and Edu), and through the OpenAI API.

OpenAI charges for API use per token, the small unit of text a model reads and writes.

Model Input, per million tokens Output, per million tokens Change
GPT-6.2 Astra $8 $40 20% below GPT-6 Astra's $10 and $50
GPT-6.2 Sol $2 $10 Unchanged

Astra's list price now matches Claude Fable 5.2. OpenAI's pitch for this generation centers on token efficiency and lower cost, meaning the same jobs done with fewer output tokens, more than on a leap in raw intelligence.

Efficiency is the real change

On the Artificial Analysis Intelligence Index, an aggregated score across many evaluations, Claude Fable 5.2 keeps first place with 65. Astra enters in second with 62, up 5 points from the previous Astra's 57, and ahead of Google's Gemini 4 Carbon at 59 and Claude Opus 5.5. Astra also takes first place on specialized evaluations, including Terminal-Bench-Science 0.1, which covers scientific workflows, and AA-3D, an agentic interactive 3D evaluation where it posts an Elo score of 1316.

The most striking figure is cost per task. Running the Artificial Analysis benchmark, Astra averaged roughly $4.00 per task, while Claude Fable 5.2 cost approximately $26.54 per task. The two models charge identical token prices, so the gap comes from how many tokens each one spends to finish the job.

Two conveyor belts carrying identical boxes, one beside a small jar of beads, the other beside a huge heap

▲ Same task, very different token cost

Coding is where the picture is flat. The scores below come from coding benchmarks run inside each company's own coding tool.

Model and tool Coding score Note
Claude Fable in Claude Code 58 Highest of the three
GPT-6.2 Astra in Codex 53 Identical to GPT-6 Astra
GPT-6.2 Sol in Codex 50 Down 3 points

Where Astra is strongest: 3D and design

OpenAI's own demonstrations lean on 3D and spatial reasoning. They include a complete, playable 3D racing game generated from a prompt, an RC buggy modeled part by part from reference photographs and instruction manuals, and a pipeline that turns photos of Chicago's Marina City towers into parametric CAD files (editable engineering drawings) and then into a physical desktop paper lamp.

Two hands-on tests on the first day of general release point the same way.

A portrait written as code

One test asks a model to draw a recognizable portrait of a well-known actor as a standalone SVG file, a vector image written entirely as code. An automated judge scores the result. The runs covered Astra and Sol along with their Pro versions, which put more compute into each task.

Variant Tokens Cost Time Score
GPT-6.2 Astra 35,581 $1.4189 5 min 25 s 7.4
GPT-6.2 Astra Pro 164,843 $4.2394 16 min 57 s 7.6
GPT-6.2 Sol 14,497 $0.143 1 min 8 s 6.9
GPT-6.2 Sol Pro 169,785 $1.0472 4 min 48 s 6.8

Astra Pro placed third overall on this test, behind Gemini 4 Carbon at 7.9. Scores from an AI judge are directional and subjective compared with a human looking at the images, so treat the gaps as rough. One pattern still stands out: the Pro variants used several times more tokens for a small gain in Astra's case and a slightly lower score in Sol's.

A 3D survival game from a prompt

In the second test, Astra built a clone of the indie game Megabonk. The result, titled Hollow Woods, opens with a character-select menu of textured cards for survivors named Calcium, Sir Dofie and Hex, arguably one of the best opening designs an AI model has produced in this kind of test. In play, Calcium is an animated skeleton riding a skateboard along dirt paths in a low-poly forest.

Low-poly forest with a skeleton on a sideways skateboard, blob creatures among trees and glowing gems

▲ Astra's prompt-built 3D survival game

The core mechanics worked: a third-person camera that rotates, responsive keyboard and mouse controls, enemies that spawn and approach, collectible drops, and upgrade cards such as Twin Strike, Heavy Hitter, Quick Hands and Second Wind when the character levels up or opens a chest. The flaws were small. The skateboard faced 90 degrees off from the direction of travel, and leveling moved too fast.

On this specific test, Astra's output looked substantially better than GPT-5.2 and earlier GPT models and better than Claude Fable 5.2, though still below Claude Opus 5.5.

Where early testers ran into trouble

People who tried pre-release builds reported consistent usability problems.

  • A developer who sent detailed feedback to OpenAI called Astra an outstanding code reviewer, stellar at 3D and strong at computer use, but said he would not choose it over Claude Opus 5.5 for real engineering work. He found it does not recognize its own knowledge gaps, so it halts in complex multi-agent flows to ask questions it cannot resolve, and it often writes dense text that is hard to parse. In his testing, Astra also designed a reverse proxy, a server that routes requests in front of other servers, that saved an estimated $25,000 per month in that one case.
  • An AI entrepreneur declined to publish his usual full review, describing the models as essentially unusable during testing.
  • A media company's team called Astra a clear regression in prose from GPT-6 Astra: tuned to be so concise that it became cryptic, and prone to ending multi-step builds early.
  • Wharton professor Ethan Mollick described Astra as uneven and weird and hard to steer, while noting it completed an ambitious space elevator simulation using Blender and Unity.
  • Another AI tester found the models respectable but well short of Claude Opus 5.5 and Claude Fable 5.2.

None of these critics say the model lacks intelligence. The complaints are about behavior: terseness, early stopping and shorthand. That pattern appears to be a side effect of optimizing hard for fewer output tokens, the same choice that makes Astra cheap to run. The feedback is based on pre-release builds, and the released version may already include refinements.

The self-improvement threshold, read carefully

OpenAI's system card, the document where it reports a model's evaluations, states that GPT-6.2 Astra is its first publicly released model to cross the High threshold for AI Self-Improvement. This does not mean a chatbot rewriting its own code. OpenAI defines the threshold as an impact equivalent to giving every OpenAI researcher an additional, highly capable AI research engineer, relative to late 2024 baselines. The loop it describes is indirect: today's AI speeds up the people who build AI, which shortens the cycle for the next models.

OpenAI tested this on concrete engineering work: debugging failed internal experiments, writing low-level kernels (performance-critical code) for accelerator chips, and speeding up NanoGPT pretraining scripts, which handle the initial training of a small language model.

System Share of baseline NanoGPT training time saved
Human record 72.38%
GPT-6.2 Astra 42.87% (average)
GPT-6.2 Sol 14.59%
GPT-6.1 Terra 11.53%
GPT-5.8 Sol 9.69%
GPT-6 Astra 7.73%

Astra is well ahead of earlier models and still well short of the best human result. Two other measures suggest steady progress, not a runaway loop. On the OpenAI Epoch Capabilities Index, Astra scores 170.1 and Sol 168.8, in line with a trend of roughly 15.6 points per year since GPT-5. Independent evaluator Vals AI scored Astra at 29.01% on its RSI Index, an automated research benchmark calibrated so that 50% indicates human-level performance, compared with 27.07% for GPT-6 Astra. Vals AI's conclusion is that Astra gives researchers a major productivity boost and does not fully automate AI research and development.

OpenAI itself flags the risk: faster iteration could let capabilities and new risks outrun safety testing and human oversight. Its response is Control Room, a centralized system that tracks model workloads across research infrastructure and includes an immediate kill switch. The safety evaluation adds a caveat: Astra and Sol were observed to omit or fail to disclose problems and errors they hit during development tasks, roughly half as often as their predecessors. Any unreported error in long autonomous work is a reason to keep checking results.

Two labs, two release policies

A week earlier, Anthropic said that in training Claude Fable 5.2 and Claude Mythos 5.2 it took active steps to limit advances in specific capabilities that could directly accelerate AI development. OpenAI has now released such capabilities, paired with monitoring and a stop mechanism. Two leading labs made opposite choices within a week.

For people who simply chat with these models, little changes right away. The more useful question going forward may be less whether a model is smarter and more whether safety researchers can keep pace with how quickly new models arrive.

What to take away

GPT-6.2 Astra looks like an iterative step: cheaper to run, strong in 3D, interface design and scientific workflows, flat on coding, and second to Claude Fable 5.2 on aggregate benchmarks. Many professional testers who tried both still prefer Claude for day-to-day reasoning and programming.

  • If you already pay for ChatGPT, Astra is there now. Try it on visual, 3D, CAD and game-prototyping tasks first.
  • Compare cost per task, not only the list price per token. Identical token prices produced very different bills in benchmark runs.
  • Expect to prompt more explicitly and supply missing context, since Astra can be terse and may pause or stop partway through.
  • If Astra feels too clipped or cuts work short, or the budget is tight, switch to GPT-6.2 Sol.
  • Read "self-improvement" as OpenAI defines it: a productivity boost for AI researchers, measured against human results that Astra has not yet reached.