RELEASE REPORT · NOVEMBER 17, 2026

GPT-6.2 Astra: release report

OpenAI released GPT-6.2 Astra on November 17, 2026. Specs, price, where it ranks, how it plays chess, how it solves sudoku, how it plays Mines, how it solves nonograms, how it plays FreeCell, how it plays Starjar, how it plays Nations Royale, how it commands a Wildhaven raid, how it plays Hustle Sim, how it plays Elemental Armory, how it drives Micro Mouse and how it plays Ravelin, with more added as evidence arrives.

DATA AS OFNov 18, 2026
SOURCELab announcement
PUBLISHEDNov 17, 2026

The release

OpenAI released GPT-6.2 Astra on November 17, 2026. Read the announcement.

  • Takes unknown. Returns unknown.
  • Not a reasoning model.
  • Closed weights.

OpenAI's release before this was GPT-6 Cyber on November 5, 2026, 12 days earlier. Over the past year it has shipped a model every 11 days on median (25 releases).

Price

$8.00 per million input tokens and $40.00 per million output tokens on OpenAI's API. Cached input is $0.40. GPT-6 Astra, the model it follows, costs $10.00 and $50.00.

Where it ranks

Added November 18, 2026.

Epoch AI's Capabilities Index puts it at 174.2 (95% interval 170.3 to 178.6), number 1 of 257 models Epoch has scored. That is the frontier: no model scores higher.

See it on Rankings.

Our chess benchmark

Added November 18, 2026.

Rated 1,751 (95% interval 1,538 to 1,964) over 18 games against calibrated Stockfish, number 1 of 16 models we've tested. Standard chess 1,879, Chess960 1,586.

In standard chess: 91.9 accuracy in open positions, 3.0 blunders per 100 moves, 1.6% of answers were illegal moves.

Each game took 13.5 minutes of model time. At the lab's list price its tokens would cost $1.17 a game.

Every game, move by move.

Our sudoku benchmark

Added November 18, 2026.

Scored 100 out of 100 (95% interval 91 to 100), number 1 of 11 models we've tested. It solved 42 of 42 grids and filled 100.0% of blank cells correctly.

By tier, from singles to grids no human technique finishes: Cozy 6 of 6, Gentle 6 of 6, Steady 6 of 6, Tricky 6 of 6, Devious 6 of 6, Expert 6 of 6, Beyond 6 of 6.

It said it was 100% sure on average.

Each grid took 178 seconds and 3,538 output tokens on average. At the lab's list price its tokens would cost $0.20 a grid.

Every answer, cell by cell.

Our Mines benchmark

Added November 18, 2026.

Scored 83 out of 100 (95% interval 68 to 98), number 1 of 8 models we've tested. It won 20 of 24 no-guess boards and cleared 88% of the safe cells on average.

By size: Beginner 5 of 6, Intermediate 5 of 6, Expert 5 of 6, Master 5 of 6.

It left 3 flags on safe cells.

Each board took 540 seconds of model time on average. At the lab's list price its tokens would cost $1.17 a board.

Every board, turn by turn.

Our Nonograms benchmark

Added November 18, 2026.

Scored 69 out of 100 (95% interval 58 to 80), number 1 of 8 models we've tested. It solved 25 of 36 puzzles, got 69.3% of cells right and satisfied 61.8% of row and column clues.

By tier: 5 by 5 6 of 6, 10 by 10 6 of 6, 15 by 15 6 of 6, 20 by 20 5 of 6, 25 by 25 1 of 6, Beyond 1 of 6.

It said it was 69% sure on average.

Each puzzle took 230 seconds of model time on average. At the lab's list price its tokens would cost $0.251 a puzzle.

Every answer, cell by cell.

Our FreeCell benchmark

Added November 18, 2026.

Scored 33 out of 100 (95% interval 20 to 46), number 3 of 8 models we've tested. It won 10 of 30 deals and got 22 of 52 cards home on average.

By tier: Easy 5 of 6, Four cells 2 of 6, Three cells 3 of 6, Two cells 0 of 6, Beyond 0 of 6.

Planning the whole game in one reply, it won 4 of 6.

The referee refused 0.8 of its moves per game.

Each game took 404 seconds of model time on average. At the lab's list price its tokens would cost $0.619 a game.

Every game, move by move.

Our Starjar benchmark

Scored 99 against the reference bot's 100 (95% interval 85 to 113), number 6 of 9 models we've tested. It made 1,164 points a game, beat the bot on 1 of 6 jars and never overflowed the jar.

By tier: Daily opening 95, Fresh jar 89, Midgame 114.

Each drop took 12 seconds of model time on average. At the lab's list price its tokens would cost $1.60 a game.

Every jar, drop by drop.

Our Nations Royale benchmark

Scored 83 out of 100 (95% interval 65 to 100), number 4 of 9 models we've tested. It won 10 of 12 wars against the game's 47 commanders and held 70% of the world at the end on average. Of the rest, none were still undecided at 15 minutes and 2 were lost.

By tier: Easy 3 of 3, Medium 3 of 3, Hard 3 of 3, Extreme 1 of 3.

War by war it scored 33 points above the game's reference player and 52 points above the same player held to its own clock.

The game refused 0.1 of its orders per decision on average.

Each decision took 14 seconds of model time on average. At the lab's list price its tokens would cost $1.06 a war.

Every war, decision by decision.

Our Wildhaven benchmark

Added November 18, 2026.

Scored 86% destruction (95% interval 80 to 91), number 1 of 8 models we've tested. It earned 2.4 stars a raid, took all three stars in 20 of 48 raids and brought the Townheart down in 48.

By tier: Hunt 97, Reckoning 85, Outmatched 75.

Raid by raid it destroyed 28 points above the game's own raiders and 1 points below a search that tries 64 plans in the simulator.

Each request took 35 seconds of model time on average. At the lab's list price its tokens would cost $0.101 a raid.

Every raid, look by look.

Our Hustle Sim benchmark

Scored 80 against the game's best line's 100 (95% interval 53 to 107), number 1 of 12 models we've tested. It reached the goal before its replies ran out in 6 of 6 runs, 1.3 times the best line's game time on average.

By tier: Fresh start 68, Takeover 92, Empire 80.

The game refused 0.0 of its purchases per decision on average.

Each decision took 16 seconds of model time on average. At the lab's list price its tokens would cost $0.934 a run.

Every run, decision by decision.

Our Elemental Armory benchmark

Added November 18, 2026.

Scored 59 out of 100 (95% interval 51 to 66), number 2 of 8 models we've tested. It won 88 of 150 Bounce Duels matches against the game's rival and 55% of the rounds.

By level: Level 1 31 of 50, Level 5 29 of 50, Level 8 28 of 50.

For scale, the rival playing its seat scores 49 and a player who measured every card scores 68.

Its cards averaged 2.8 stars out of 3.

Each decision took 23 seconds of model time on average. At the lab's list price its tokens would cost $0.120 a match.

Every match, round by round.

Our Micro Mouse benchmark

Added November 18, 2026.

Scored 79 against the median person at 100 (95% interval 60 to 98), number 1 of 8 models we've tested. It was faster than the median person on 0 of the 6 daily mazes people raced, and the game's bot scores 94 on them.

On the campaign mazes it scored 84 against the game's bot at 100, and it reached the dock in 14 of its 14 runs.

By tier: Daily mazes 79, Workbench 83, Cold Storage 84, Fast Lane 80, Night Shift 87.

The game's test pilot, which steers every step along a route it is given, scores 48 on the daily mazes.

Each reply took 11 seconds of model time on average. At the lab's list price its tokens would cost $0.554 a run.

Every run, reply by reply.

Our Ravelin benchmark

Scored 89 against par at 100 (95% interval 41 to 137), number 1 of 9 models we've tested. It held 4 of 6 boards and beat par on 4, losing 1.3 lives a battle; the game's own bot scores 76 on the same boards.

By tier: Daily siege 130, Campaign 124, Iron 13.

Each decision took 14 seconds of model time on average. At the lab's list price its tokens would cost $0.323 a battle.

Every battle, decision by decision.

Each section was written on the date shown and is not revised afterwards. Specifications from Models.dev, prices from provider listings, capability scores from Epoch AI under CC BY 4.0, chess, sudoku and Crest Arcade game results from Glimmer Index Bench.