The AI Spotlight Follow
GPT-6.2 Astra Rebuilt a House in Blender Better Than Its Rivals. Then It Failed All Six Brain Scans
Model Wars

GPT-6.2 Astra Rebuilt a House in Blender Better Than Its Rivals. Then It Failed All Six Brain Scans

OpenAI's new GPT-6.2 Astra and Sol launched on November 17 with a 20% price cut and a wave of harsh early reviews. We walked through a full hands-on gauntlet, from a two-piano duel and an Unreal Engine sword fight to a hidden frog and six brain scans, to see where the new model actually lands.

Advertisement

OpenAI shipped a new flagship on Tuesday, November 17, 2026, and within hours some of the most-followed AI testers on the internet were calling it a step backwards.

One called the new models "essentially unusable." Another called GPT-6.2 Astra "the worst model OpenAI has ever released." OpenAI, meanwhile, called it "a new standard for professional work."

Both can't be fully true. So the useful question is simple: what happens when you actually hand GPT-6.2 Astra hard work and watch?

One tester did exactly that, putting Astra and its cheaper sibling through ten demanding challenges, from an animated biology lesson to a sword fight in Unreal Engine to a set of real brain scans. The results are the most honest picture we've seen of the model so far. It's uneven. In places it's genuinely the best on the market. In others it isn't close.

Advertisement

Here's the full run, test by test, plus the specs, the price and what it means for you.

First, what OpenAI actually launched

There are two new models, not one.

  • GPT-6.2 Astra is the frontier model, built for your hardest projects: complex reasoning, coding, computer use and polished professional documents.
  • GPT-6.2 Sol is the efficient everyday model, aimed at routine coding, writing and multi-step agent workflows. OpenAI says many of its own employees use it as their default.

Both arrived on November 17 in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.2-astra and gpt-6.2-sol. Free users don't get them yet.

The pricing is the headline OpenAI wants you to remember:

  • Astra: $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens. That's a 20% cut on standard tokens versus the original GPT-6 Astra ($10 and $50) and a 60% cut on cached input.
  • Sol: $2 input, $0.10 cached and $10 output, the same as GPT-6.1 Sol.

OpenAI also says both models get more done with fewer output tokens, so the real cost per task should fall further than the sticker price suggests.

And the timing matters. The launch came about a week after Anthropic released Claude Fable 5.2 at the very same $8/$40 rate, and shortly after Google opened up Gemini 4 Carbon at introductory rates that match Sol. Astra is walking into the most crowded top tier the industry has ever had.

How the tests were run

A quick note on method, because it shapes every result below.

Frontier models in 2026 are built to work as agents, not just chatbots. So the best place to use Astra isn't a plain chat window. It's an agent "harness" that can open apps, read and write local files and run several workers in parallel. Here, that harness was OpenAI's desktop app, the tool formerly known as Codex, which can juggle multiple agents across multiple projects.

Several tests also used a critic loop: a second agent grades the work out of 10 and, if the score falls short, sends the main agent back to revise. It's a common trick for squeezing better output out of any model, and it gives a useful number to compare.

Tip: in that app, /status shows usage for a single chat, while /usage opens your account's weekly limit.

Test 1: an animated tour through a cell

The prompt: create a visually stunning, accurate educational animation that flies through an animal cell and explains every organelle, coded entirely from scratch. No external animation tools, no video generators. Because Astra can't generate voices itself, it was told to use Google's Gemini text-to-speech for narration, with the documentation pasted in.

Time: about 16 minutes.

What came out: a narrated fly-through that hit an impressive amount of real biology in order:

  1. The plasma membrane's two layers of phospholipids, with channel proteins letting specific ions into the cytosol.
  2. DNA wrapped around proteins as chromatin in the nucleus, genes copied into messenger RNA, and mRNA leaving through nuclear pores.
  3. The nucleolus assembling ribosomal subunits, and two subunits clamping around mRNA to build a protein.
  4. A secretory protein entering the rough endoplasmic reticulum, the smooth ER making lipids and storing calcium.
  5. A vesicle carrying cargo to the Golgi for modification and sorting.
  6. Motor proteins hauling vesicles along microtubules from the centrosome, powered by ATP.
  7. Mitochondria using a proton gradient to drive ATP synthase.
  8. Endosomes, lysosomes breaking down cargo in an acidic chamber, and peroxisomes handling fatty acids while catalase clears hydrogen peroxide.

The verdict: the script was solid. The animation wasn't. Shapes overlapped and looked incoherent, the ribosome sat static instead of moving along the mRNA, and the cargo vesicle just slid across the screen. For a frontier model, the tester rated it "not too impressive."

Test 2: a live two-piano duel

OpenAI says GPT-6.2 is especially strong at computer use, operating real software through its interface. So this test raised the bar on a classic.

Top models, including the original Astra and Claude, can already compose a piano solo and play it live on a browser keyboard. This time Astra had to:

  • Open two separate browsers, each with a different online piano.
  • Compose a fast, roughly one-minute piano duel in the style of Chopin for two players.
  • Perform it live, one player per piano, keeping both parts in sync.

It was set to maximum reasoning and thought for about 20 minutes.

The verdict: coordination worked. The two pianos played in time, which is genuinely hard for an agent driving two interfaces at once. The music itself was very basic and nowhere near a professional piece. A win for computer use, a miss for composition.

Test 3: build a 3D sword fight from two photos

This was the biggest step up in difficulty.

In earlier rounds of testing other models, the agent got a ready-made 3D character. This time Astra got two reference photos: a woman with a sword and a ninja. It had to:

  • Create both characters in 3D, as faithful and realistic as possible, including hair and clothing movement.
  • Add rigging, the digital bones and joints that let a character move.
  • Pull suitable run, dodge, jump and slash animations from Mixamo, Adobe's free library of character animations.
  • Build an imperial Chinese palace interior, generate every asset in Blender and render the game in Unreal Engine.
  • Run the critic loop, revising until the visual quality scored at least 8.5/10, for up to four rounds.

Round one: Astra worked for 1 hour 2 minutes, went through all four rounds and finished at 7.6/10, below target. Its own screenshots showed a basic-looking game with stiff animations.

Round two: one follow-up prompt asked for faster, more epic, AAA-style moves with visual effects and motion blur. Astra worked for roughly another hour.

The result: a playable Unreal project. The palace read as an imperial Chinese palace but looked plain and could use far better lighting. The heroine's anatomy was mostly right, apart from a wrist that looked dislocated, and she resembled the reference photo. The ninja looked good. She could run, jump and slash, and repeated slashes chained into a combo.

The verdict: not bad for something coded from scratch, but nothing like a AAA game. The bigger disappointment was relative. The tester had liked the original Astra and saw no noticeable improvement here.

Test 4: write, sing and produce a song

The brief: an upbeat, futuristic song with English lyrics and female vocals about "accelerating into the singularity." Vocals were to be made in ACE Studio, an AI vocal-synthesis app where you set melody and lyrics. The instrumental was to be built in the tester's Waveform digital audio workstation using any installed plugins or samples, with variations, effects and professional mixing and mastering.

First pass: 52 minutes of work and, in the tester's words, a result that sounded "absolutely awful."

Second pass: a prompt asking for better samples and instruments, a better mix, louder vocals and vocals better matched to the music. About 15 more minutes.

What it built: inside ACE Studio, a lead vocal with melody and lyrics plus added harmony vocals, which honestly sounded decent in isolation. In the DAW, exported vocals, percussion, synths and assorted effects. But the arrangement looked simpler than what Gemini Carbon or Claude Fable had produced on the same kind of task, and the finished song was rated one of the worst the tester had ever heard, with garbled, half-coherent lyrics.

The verdict: a clear fail on music.

Test 5: rebuild a real house in Blender

Here's where Astra surprised everyone.

The task: take a real vacation rental listing and reconstruct the entire building, inside and out, in Blender, then fly through it like a virtual tour and render it. The listing photos showed:

  • A living room with a sofa facing the pool, a dining table to one side and a painting on the wall.
  • A bedroom and a bathroom, each with only a single photo.
  • The exterior, with a circular pool, a rectangular pool and an outdoor sofa and table, plus a pool wall that slants.

First pass: 43 minutes, with thin detail. After a prompt to add detail and stay faithful to the photos, Astra worked for more than two hours.

The result: the exterior was wrong. But nearly everything else impressed:

  • Pillows that looked genuinely realistic, better than the tester had gotten from Gemini Carbon or Claude Fable.
  • Convincing objects on the dining table.
  • The wall painting reproduced correctly.
  • A bedroom rendered very accurately, down to lights, pillows and blanket.
  • A bathroom that even included robes hanging on the wall, a detail neither Fable nor Carbon had managed.

The one complaint was the camera. The "virtual tour" slowly panned and orbited instead of sweeping through like a drone video.

The verdict: the standout result of the whole run. This lines up with OpenAI's own claim that Astra builds better front ends, "particularly for websites, gaming, and 3D design," and with early testers who praised its 3D and visual work.

Test 6: a bullet through a water balloon

The task: a 3D simulation of a bullet piercing a water balloon, the balloon popping, water bursting out, deforming and falling to the ground. No Three.js and no external web libraries: everything coded from scratch, with a critic loop requiring a score of 8 or more over up to three rounds.

Time: 41 minutes.

The verdict: fairly realistic physics and mechanics from several angles. Decent overall, though the tester slightly preferred Claude Fable 5.2's version for realism.

Test 7: find the hidden frog

A favorite visual test: a photo of leaf litter with a camouflaged frog hidden somewhere in it. The question was simply whether there's any animal in the image and, if so, to circle it.

Astra worked for about three minutes, seemingly scanning the image in a grid. It answered that there "may be a small frog or toad tucked beneath a curled leaf" and circled a spot.

There was no frog there. Fail. In this tester's runs, only Gemini Carbon and a DeepSeek model have found it.

Test 8: six brain scans

The task: six brain scans, each containing a type of brain tumor. Identify the type in each, if any.

Astra thought for just over a minute and then went zero for six:

  • Top left: it hedged between a pituitary macroadenoma and a suprasellar meningioma. Treating its first suggestion as the answer, wrong.
  • Top middle: "no tumor." Wrong, every image had one.
  • Top right: meningioma. Wrong.
  • Bottom left: nothing identified. Wrong.
  • Bottom middle: craniopharyngioma as first choice. Wrong.
  • Bottom right: wrong.

The verdict: a stark reminder that a general chatbot is not a radiologist. More on that below.

Test 9: deep research on Alzheimer's

The task: analyze the key protein mechanisms in Alzheimer's disease, contrast the monoclonal antibody therapies that target each one, and critically appraise cognitive and imaging results from recent phase 3 trials.

There's no single right answer to a task like this, and every frontier model can now search and synthesize scientific literature reasonably well. But Astra's report stood out for format: a clear flowchart and a detailed but compact table of trials and outcomes.

The tester found it more thorough than Claude's answers, which in their experience tend to hold back on medical research questions.

Test 10: fresh ideas for plastic waste

The prompt: propose three novel solutions to plastic waste at scale. Each had to be highly scalable, cost-competitive with current methods or with making virgin plastic, and work on mixed, contaminated or hard-to-recycle plastic.

Astra's three ideas:

  1. Mold mixed plastics into replaceable cores for reusable freight pallets, turning messy waste into a durable industrial part.
  2. Recover clear polyethylene from printed and adhesive-covered film, a notoriously hard packaging stream.
  3. Feed worn polyester textiles and opaque packaging into a shared monomer plant, breaking them back down into building blocks.

Grounded, industrial ideas you could actually take to an engineer, rather than science fiction.

What the benchmarks say

On paper, Astra looks far better than in many of those tests. The charts shown at launch and on third-party leaderboards tell a consistent story: near-frontier scores at a much lower cost per task.

  • Agents' Last Exam, which measures long, professional workflows: Astra scored above Claude Fable 5.2 at a much lower cost.
  • Terminal-Bench Science, for scientific research workflows: ahead of Fable 5.2 and Claude Opus, again cheaper.
  • Coding: behind Claude Fable and Opus 5.5, though much cheaper.
  • Legal work and GDPval, which covers knowledge work from finance, data science, biotech and marketing: the GPT models led Claude at a fraction of the cost.
  • Artificial Analysis Intelligence Index: second, behind Fable 5.2 and ahead of Gemini Carbon and Opus 5.5. Based on hands-on results, the tester thought that placement was too generous and trusts the index less as its definitions keep changing.
  • Speed: time per task was far lower than Claude's models or Gemini Carbon, even at maximum reasoning.
  • Cost per task: under half of Opus 5.5 and nearly seven times cheaper than Fable 5.2.
  • Omniscience accuracy, how often answers are factually right: second, ahead of Opus 5.5 and Carbon.
  • Hallucination rate, how often it makes something up instead of admitting it doesn't know: roughly 60%, among the worst on the chart.
  • Vals Index, across finance, coding, legal and more: seventh, behind even the smaller Claude Sonnet and Gemini Hydrogen.
  • KernelBench, very hard low-level GPU code: fifth, well behind Opus 5.5.
  • ARC-AGI-3, which drops an AI into an unfamiliar game without telling it the rules: largely saturated, with Gemini 4 Carbon near 98% and the original Astra near 100%. GPT-6.2 Sol posted a comparable score at far lower cost.
  • Epoch Capabilities Index and LiveBench: Astra tied with Claude Fable for first, while costing much less. The split, by those measures, is that Astra leads on coding, browser use and data analysis, while Fable leads on agentic work, design and research.

The tester's summary: OpenAI wasn't aiming for the smartest model this time. It was aiming for the most cost-efficient near-frontier model, and by the numbers, it may have built it.

Why the reviews are so split

The hands-on results line up closely with the launch-day backlash.

According to VentureBeat, investor Matt Shumer said he gave Astra work as a sub-agent and the Claude models supervising it ended up redoing the work. Developer Theo Browne said Astra spent seven days on his experimental TypeScript-compiler rewrite without landing an improvement, produced a version that ran 56% slower, then spent four more days without landing a change. Every CEO Dan Shipper said it often stopped early on end-to-end builds and wrote so briefly that some explanations were hard to follow. Wharton's Ethan Mollick called it "very uneven & weird."

Yet the same critics found real strengths. Browne credited Astra with catching code-review issues other models missed and praised its 3D, game development and computer-use abilities. Mollick got it to build a playable space elevator in Blender and Unity. Other testers praised its animated visuals, motion graphics and video editing.

That's the same shape as the ten tests above: strong on spatial, 3D and computer-use work; weak on long-haul autonomy, creative taste and knowing what it doesn't know.

What this means for you

If you pay for ChatGPT, you already have both models in Work and Codex. The default on consumer plans is Sol at light reasoning effort. Open the advanced settings to choose Astra and a higher effort for hard jobs.

If you build on the API, the price cut is real, and OpenAI's own changelog adds two-hour and eight-hour background windows that cut token rates by 30% and 40% for work that can wait. Note the migration changes: tool calling requires the Responses API, and GPT-6.2 changes how reasoning is preserved when you edit earlier parts of a conversation.

If you do 3D, game prototyping, product mockups or interface work, Astra deserves a serious trial. The Blender house is the kind of result that saves days.

Practical takeaways

  • Use a critic loop. A second agent grading the output pushed several results up. It's cheap insurance with any model.
  • Budget for a second pass. Almost every good result here came after one blunt follow-up prompt. Plan for it.
  • Ask for uncertainty. With a hallucination rate around 60% on the "don't know" test, tell Astra to say when it isn't sure and to cite sources you can open.
  • Match the model to the job. Astra for 3D, computer use, data analysis and documents; consider Claude for long agentic coding and design; Gemini Carbon for visual search tasks like the frog.
  • Count finished work, not tokens. A cheap model that has to be sent back three times isn't cheap.

Counterpoints and risks

  • These are single runs. Each test was one attempt by one tester with one set of prompts. Models vary run to run, and a different harness or prompt could flip several results.
  • Benchmarks disagree. First place on one index and seventh on another is a warning to test on your own work before you switch.
  • Medical use is a hard no. Zero for six on tumor types isn't a quirk. No general-purpose chatbot should be used to read medical imaging. That's a job for trained clinicians and regulated, validated tools.
  • Early reviews can age. OpenAI fixed some of Browne's complaints before launch, and model behavior often shifts in the weeks after a release.

FAQ

What is the difference between GPT-6.2 Astra and GPT-6.2 Sol?

Astra is OpenAI's frontier model for the hardest reasoning, coding, computer-use and document work. Sol is the cheaper, faster everyday model for routine coding, writing and multi-step agent tasks. At standard API rates, Astra costs four times as much as Sol.

How much does GPT-6.2 Astra cost?

$8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens at standard rates, a 20% cut on standard tokens versus GPT-6 Astra.

Is GPT-6.2 Astra better than Claude Fable 5.2?

It depends on the job. Leaderboards put them roughly level or Astra slightly behind on overall intelligence, with Astra far cheaper per task and stronger on browser use and data analysis. Fable tends to win on long agentic work, design and research.

Why do some reviewers hate it?

The main complaints are that it stalls or stops early on long projects, writes too tersely, has a noticeable "house style" and is hard to steer. The praise centers on 3D, games, visuals and computer use.

The bottom line

GPT-6.2 Astra isn't the leap its launch page promises, and it isn't the disaster its harshest critics describe. It's a cheaper, faster, very uneven model with a few genuinely best-in-class tricks, especially 3D scene building and operating software, and a few dangerous blind spots, especially when it doesn't know an answer.

Treat it like a talented specialist, not a general genius. Give it the jobs it's good at, check everything else, and never let it near your medical scans.

Sources

Stanford's 'Next Concept' AI Solved 86 of 100 Hard Sudokus Where a Normal Model Solved One. Did It Really Break LLMs?Up next · 13 min readStanford's 'Next Concept' AI Solved 86 of 100 Hard Sudokus Where a Normal Model Solved One. Did It Really Break LLMs?A Stanford and Together AI paper says language models don't have to write one word at a time. Its Next-Concept Modeling sketches a whole answer, then sharpens it in a dozen parallel steps, running up to 31 times faster than diffusion rivals. The internet declared ChatGPT dead. The tables tell a more careful, and more interesting, story.Keep reading →