OpenAI has launched GPT-6.2 Astra, a frontier-class model

1

2

. Per Wccftech, it is the first model to score 100% on ARC-AGI-3, a benchmark of interactive game-like environments from François Chollet's ARC Prize team, built to test fluid intelligence

2

. User reactions in the first days have been mixed to negative

2

.

Source: Geeky Gadgets

Source: Geeky Gadgets

The rest of the scoreboard

The benchmark win is the headline; the other charts are the problem. The model shows no improvement on the Artificial Analysis Coding Agent Index, matching GPT-6 Astra's score of 53

1

2

. Geeky Gadgets reads that as a slight gain on scores rounded to whole numbers

1

. Vals AI ranks it seventh on its index, trailing several Anthropic and Google models, per Wccftech

2

.

The pricing is where OpenAI pushes hardest. Astra costs $8 per million input tokens and $40 per million output tokens, about 20% below its predecessor GPT-6 Astra

2

. On cost per task the gap is the pitch: about $4 per Artificial Analysis Intelligence Index task against Fable's $26.50, roughly 8x cheaper on Wccftech's math

2

. Geeky Gadgets puts the gap at about seven times on a different benchmark

1

. On intelligence the model still trails: Astra scores 62 on that index against Fable 5.2's 65, about 3 points behind (Geeky Gadgets' figures run 60.2 and 64.6)

1

2

.

What testers found

Dan Shipper reported that his team, after weeks with the models, found the release disappointing next to Fable 5.2

1

. Astra often stopped too early on real coding tasks, wrote answers so brief they could be hard to follow, and showed good default design instincts with little variety

1

. In one run it spent 30 minutes on a task and came back with a too-short answer instead of listing seven items

1

. Users have reported it pausing long tasks to ask odd questions, producing output that could not be parsed, and getting stuck in loops

2

. The criticism has sharpened with the wait, too; the models took a long time to arrive, and reviewers judged them against that wait

1

.

There are real wins. Per Wccftech, Astra is strong on science and knowledge benchmarks and leads in 3D and robotics

2

. It leads Agent's Last Exam at about $2 per task against Fable's $60, and its extra-high reasoning setting beats its max setting on some benchmarks

1

. Beyond that the picture varies by test: Fable 5.2 leads Frontier Code and Gemini 4 Carbon tops Automation Bench

1

. Astra's lead on LiveVal comes on a chart missing two rivals

1

.

The self-improvement bet and OpenAI's answer

Per the company's own rating, Astra is the first model OpenAI has publicly released rated 'High' for AI self-improvement

1

2

. Its system card presents the model as an early milestone on the road to fully automated self-improvement, and the model writes faster kernels for OpenAI's own chip

1

2

. Full recursive self-improvement is still only expected by July 2027, per Vals AI

1

. The two labs also sit partly opposite on protecting their models. OpenAI says Astra and its sibling Sol leak their underlying reasoning about 1% of the time

2

, while Anthropic keeps its models off machine-learning tasks to discourage distillation

1

. Per creator David Ondrej, Anthropic has surpassed OpenAI in revenue and valuation, a reversal of last year's positions

1

.

Per Tibo, an OpenAI figure, the divide came down to timing

2

. As Wccftech put it, a few people got early access a while back, and the team "then made a lot of changes in the" final 3 to 4 days without re-testing

2

. The influencer videos were recorded before those changes landed

2

. "Don't get me wrong, while the models are a hefty jump compared to their GPT predecessors, they are by no means perfect," he wrote

2

. He added that the teams want the models to get better at the taste and niceness of use in future iterations, and that OpenAI hears the feedback

2

.

Source: Wccftech

Source: Wccftech

© 2026 TheOutpost.AI All rights reserved