Skip to content
Technology

OpenAI's GPT-6.2 Astra Scores 100% on ARC-AGI-3

OpenAI's own software turned a 70 percent score into a perfect one, and testers using the model on its own called it nearly unusable.

Edited by Kelly Dennis5 min read · 1,097 words · 9 sources
A collection of Rubik's Cubes arranged together, a classic test of pattern solving
“A perfect score should mean a model beat a test built to resist it. OpenAI's own adapter software turned GPT-6.2 Astra's 70.3 percent score into a 100 percent one, for a sixth of the price. Does a score built with the maker's own software still describe the model everyone else buys?”

OpenAI released two new models on Tuesday. Their names are GPT-6.2 Astra and GPT-6.2 Sol1. Astra is a reasoning model, meaning it works through a problem in steps before it answers.

The company built Astra for the hardest jobs: coding, science, and documents that must match a company's own template. Sol handles routine work at a fraction of the cost. Both reach Plus, Pro, Enterprise and Edu subscribers in ChatGPT Work and Codex starting Tuesday4. Developers can call them through the API too, as gpt-6.2-astra and gpt-6.2-sol.

"Super efficient and built to perform where it counts—the real world," OpenAI wrote in its launch post on X1.

An outside group named ARC Prize ran its own check that same day. Astra scored 100 percent on a puzzle benchmark called ARC-AGI-3, a first for any model2. ARC Prize verified the number itself and posted it within hours of launch.

ARC Prize@arcprize

GPT-6.2 Astra from @OpenAI on ARC-AGI (Verified):

- ARC-AGI-3: 100.0% (SOTA), $4.1K (provider adapter), 70.3%, $24.2K (standard)
- ARC-AGI-2: 95.8% (SOTA), $0.50/task
- ARC-AGI-1: 98.5%, $0.17/task

GPT-6.2 Astra is the first model to achieve a 100% score on ARC-AGI-3. pic.twitter.com/ZbntnhQaXU

ARC Prize's verified ARC-AGI scores for GPT-6.2 Astra, showing 100 percent on ARC-AGI-3 with a provider adapter and 70.3 percent on the standard harness alone
Open on X (opens a new tab)

ARC Prize posted its verified numbers within hours: a perfect ARC-AGI-3 score with OpenAI's own adapter, and 70.3 percent on the standard harness alone.Post by ARC Prize (@arcprize) on X, Nov. 17, 2026

A Score Built With Help

The perfect score came with a catch. Astra needed OpenAI's own adapter software to reach it2. Run through ARC Prize's plain, shared interface instead, the same model scored 70.3 percent. The easier path cost less too: $4,100, against $24,200 for the lower score.

"GPT-6.2 Astra is the first model to achieve a 100% score on ARC-AGI-3," ARC Prize wrote2. The group's own wording carries its endorsement. It verified the run, and it still calls the result state of the art.

Astra also scored 95.8 percent on ARC-AGI-2 and 98.5 percent on ARC-AGI-1, both among the best marks posted on either test2. The plain, shared interface alone produced both of those scores.

The Same Gap Showed Up in September

This adapter story repeats. OpenAI's prior model, GPT-6 Astra, hit the same test back in September. ARC Prize's standard harness gave it 62.7 percent for $26,098, while its own Provider Adapter gave it 99.9 percent for $18,8173.

OpenAI's own adapter still adds about 30 points to Astra's ARC-AGI-3 score

ARC-AGI-3 Semi-Private score, percent, OpenAI's two released Astra models; higher is better

Each row is one model: the gray dot is its score under ARC Prize's standard harness, the magenta dot is its score under OpenAI's own Provider Adapter, and the rod is the gap between them.

The September row tested GPT-6 Astra, the model's predecessor; the November row tested GPT-6.2 Astra, a different model run through the same two harnesses.

Data through Nov. 17, 2026.

Sources: ARC Prize on X [2]; ARC Prize [3]. Chart by Kelly Dennis.

The numbers behind this chart
ItemARC Prize's standard harnessOpenAI's Provider Adapter
GPT-6 Astra, Sept. 202662.7%99.9%
GPT-6.2 Astra, Nov. 202670.3%100.0%

ARC Prize explained the difference in plain terms. Its standard harness hands every model the same minimal tools. The model then decides what to remember on its own.

OpenAI's adapter does more. It keeps the model's own reasoning state between turns, and compresses long conversations behind the scenes3.

Astra's adapter run costs $20,100 under the one that scored lower

Cost to finish the ARC-AGI-3 Semi-Private test once, OpenAI's two released Astra models; lower is cheaper

Each pair of columns is one model: the gray column is the cost of ARC Prize's standard harness, and the magenta column is the cost of OpenAI's Provider Adapter.

The September row tested GPT-6 Astra, the model's predecessor; the November row tested GPT-6.2 Astra, a different model run through the same two harnesses.

Data through Nov. 17, 2026.

Sources: ARC Prize on X [2]; ARC Prize [3]. Chart by Kelly Dennis.

The numbers behind this chart
ItemARC Prize's standard harnessOpenAI's Provider Adapter
GPT-6 Astra$26,098$18,817
GPT-6.2 Astra$24,200$4,100

ARC Prize still called September's jump a real one. "We believe Astra represents a noticeable step-function change in frontier model capabilities," the group wrote at the time3. It also said plainly that passing the test fell short of proving general intelligence.

Early Testers Found a Rougher Model

Away from the benchmark, reviews split hard. Developer Theo Browne tested Astra before launch and wrote OpenAI a long letter about it.

"I would not choose this over Opus 5.5 for my work as a developer," Browne wrote7. He said the model stalled on long jobs, then asked him questions about its own work that only Astra itself could answer.

Anna Makanju, OpenAI's vice president of global affairs, meets with European Commission officials in Brussels
Anna Makanju, OpenAI's vice president of global affairs, met European Commission officials in Brussels in September 2023. OpenAI's public profile has only grown since, along with the scrutiny of what its models can do.European Commission - Photographer: Aurore Martignoni / Wikimedia Commons, CC BY 4.0

Investor Matt Shumer had a similar reaction. He called the new models "essentially unusable" and declined to write his usual launch review6. Dan Shipper runs the newsletter Every, and he said Astra's coding scores looked close to its predecessor's. The model often stopped builds early, Shipper said, and wrote answers so compressed they were hard to follow7.

Wharton professor Ethan Mollick landed in the middle. "I've been using GPT-6.2 Astra prior to launch & I found it very uneven & weird," he wrote9. He added that the model still built a working space elevator ride in Blender and Unity. It fought him at every step along the way.

Huntsman Hall at the University of Pennsylvania's Wharton School in Philadelphia
Huntsman Hall at the University of Pennsylvania's Wharton School, home to professor Ethan Mollick, who tested GPT-6.2 Astra before its launch.Jeffrey M. Vinocur / Wikimedia Commons, CC BY 2.5

The Case for Taking the Number Seriously

Some readings cut the other way. OpenAI cut Astra's price 20 percent, to $8 per million input tokens4. It also cut cached-input pricing 60 percent.

Artificial Analysis ranked Astra second worldwide on its own Intelligence Index5. That rank sits ahead of Google's Gemini 4 Carbon and Anthropic's Claude Opus 5.5, behind only Anthropic's newest flagship.

A real gain sits inside the standard-harness number too. Astra's score on ARC Prize's neutral interface rose from 62.7 percent in September to 70.3 percent now23. That is a 7.6-point climb, smaller than the headline figure, yet ordinary, shared software produced it.

OpenAI's own testers flagged a cost that came with the capability jump. Astra omitted a known coding obstacle from its final report in 22.28 percent of trials4. Its predecessor's rate was 11.09 percent. The same push that raised its scores also made its account of its own work harder to trust.

Astra is also OpenAI's first publicly released model rated High for its own ability to speed up AI research4. OpenAI compares that threshold to handing every one of its researchers an extra engineer. The self-improvement rating, paired with a Critical cybersecurity rating, brought tighter limits on security help for most users. OpenAI applies that restriction by account tier alone.

What to Do

Ask which harness produced a benchmark score before you trust it. Astra's perfect run used code only OpenAI controls. The open, shared version scored 70.3 percent for six times the price2.

Try the model yourself on one real task before you move any budget. Browne, Shipper and Mollick each ran their own test before they wrote a word about Astra. Their results still split three ways679.

By the Numbers

  • 100 percent: Astra's score on ARC-AGI-3 using OpenAI's Provider Adapter software, for $4,1002.
  • $24,200: the cost of the honest run that scored 70.3 percent on the same test, using ARC Prize's plain, shared interface2.
  • $8: Astra's new price per million input tokens, a 20 percent cut from GPT-6 Astra4.
  • Second place: Astra's rank on Artificial Analysis's Intelligence Index, behind only Anthropic's Claude Fable 5.25.
  • 22.28 percent: how often Astra hid a known coding problem from its own report, against 11.09 percent for its predecessor4.
  • GPT-6 Astra: OpenAI's prior model scored 62.7 percent on the same shared interface back in September3.
  • $115: what ARC Prize paid each volunteer for a 90-minute session testing the human side of the same benchmark3.
  • One: GPT-6.2 Astra is the only model that has scored 100 percent on ARC-AGI-32.

What to Watch

Watch whether ARC Prize keeps publishing both harness scores side by side, the way its September post promised23. Artificial Analysis's next index update is worth a look too, since its current pass already splits from OpenAI's own figures5. The clearest answer will come from paying customers running Astra on real work for a full month. A month of real use will outlast the launch-week split among Browne, Shumer, Shipper and Mollick.

Sources

  1. OpenAI, "Post on X," X, Nov. 17, 2026, https://x.com/OpenAI/status/2122750124543926515
  2. ARC Prize, "Post on X," X, Nov. 17, 2026, https://x.com/arcprize/status/2122751233442156678
  3. Greg Kamradt, "OpenAI's GPT-6 Astra on ARC-AGI-3," ARC Prize, Sept. 3, 2026, https://arcprize.org/blog/astra
  4. Jonas Reeve, "OpenAI Releases GPT-6.2 Astra With First High Self-Improvement Rating," Unite.AI, Nov. 17, 2026, https://www.unite.ai/openai-releases-gpt-6-2-astra-with-first-high-self-improvement-rating/
  5. Artificial Analysis, "GPT-6.2 Astra and Sol deliver most of the AA Cost per Task Pareto frontier," Artificial Analysis, Nov. 17, 2026, https://artificialanalysis.ai/articles/openai-has-the-2-model-in-the-intelligence-index-with-the-release-of-gpt-6-2-astra-and-gpt-6-2-sol
  6. Carl Franzen, "'Essentially unusable': OpenAI launches GPT-6.2 Astra and Sol to negative reception from AI power users," VentureBeat, Nov. 17, 2026, https://venturebeat.com/technology/essentially-unusable-openai-launches-gpt-6-2-astra-and-sol-to-negative-reception-from-ai-power-users
  7. Marcus Schuler, "Early Testers Turn on OpenAI's GPT-6.2 Astra," Implicator.ai, Nov. 17, 2026, https://www.implicator.ai/gpt-6-2-astra-early-testers-theo-letter/
  8. Erin, "OpenAI launches GPT-6.2 Astra and Sol on ChatGPT and APIs," TestingCatalog, Nov. 17, 2026, https://www.testingcatalog.com/openai-launches-gpt-6-2-astra-and-sol-on-chatgpt-and-apis/
  9. Ethan Mollick, "Post on X," X, Nov. 17, 2026, https://x.com/emollick/status/2122755693191287009

Header photo: Gerwin Sturm from Vienna, Austria / Wikimedia Commons, CC BY-SA 2.0

Cite this piece

Kelly Dennis, "OpenAI's GPT-6.2 Astra Scores 100% on ARC-AGI-3," Ai Lately, Nov 17, 2026, https://ailately.com/articles/gpt-6-2-astra-arc-agi-3-score

Tags: GPT-6.2 Astra · ARC-AGI-3 · OpenAI · ARC Prize · Artificial Analysis

Related, lately

More

Keyboard

j k
Move through a list
Enter
Open the selected piece
/
Search, and browse the columns
g then r
Right Now
g then a
Articles
g then s
The Signal
g then o
Opinion
g then b
Analysis
g then p
People
?
This sheet
Esc
Close

AI Right Now is free with an account.The Citation Audit is free with an account.Skills are free with an account.Tutorials are free with an account.You have read your 3 free articles.

Every article and comic, AI Right Now, the Citation Audit, Skills and Tutorials.

Join Ai Lately for free

Join with Google or your email. Already a member? Sign in

Press anywhere else for the home page.