Epoch AI updates EBR-bench after GPT-6 Astra exploits a powerful card

Epoch AI updates EBR-bench after GPT-6 Astra exploits a powerful card

Epoch AI published this update on October 7, 2026. It says GPT-6 Astra achieved a perfect score on EBR-bench by using a card that can bypass the game’s usual limit on turns. Epoch AI says the strategy is legitimate, but when fully exploited it sidesteps the tactical play the benchmark was designed to measure. The organization has therefore banned the card in the default experiment, while continuing separate experiments that allow it. With the card banned, Astra’s highest score came close to the benchmark maximum, but Epoch AI says its tactical play remains weaker than that of the top human players and shows no sign of improving at minimizing fatigue over time. Epoch AI also reports that multi-agent setups—where copies of a model divide work and consult one another—encouraged more varied deck choices in most tested models, without significantly changing overall scores. The update’s broader point is that a striking benchmark score can reflect discovery of a shortcut rather than strength across the capabilities the test aims to measure. Epoch AI expects EBR-bench to reach saturation within a few months and hopes to challenge models with more complex games. The update does not establish how models will perform as EBR-bench’s default rules change.

Publication: 92f8122e-a3ad-6465-0af2-d3563e855059

attribution — pass: The piece is a single Epoch AI publication, and the reading consistently attributes the findings, the card ban and the forecast to Epoch AI rather than stating them as fact.

EBR-bench is a benchmark that we designed to test models’ ability to learn from experience by playing the board game Earthborne Rangers.

summary — uncertain: The summary never says EBR-bench is based on the board game Earthborne Rangers or that it tests learning from experience on long-horizon tasks. It also leaves 'fatigue', 'deck', 'saturation' and 'multi-agent' context thin. Without this, a general reader cannot tell what the card or the score means. It also drops Epoch's points that Astra appeared superhuman at discovering the exploit and still made roughly a 50% jump under the ban. Smallest repair: add one sentence explaining the game-based design and fatigue, and restore the 50% jump and the 'valuable information' framing.

Compared to the strongest previous models, this is still roughly a 50% jump in average performance, showing that Astra is not reliant on the banned card for strong performance.

date — pass: The reading gives October 7, 2026, which matches originalPublished as a UTC day. The saturation expectation is reported as Epoch's forecast, not an outcome. The reading includes no hindsight framing.

We therefore expect EBR-bench to be fully saturated within a few months.

support — uncertain: The reading states as 'the update’s broader point' that a striking score can reflect a shortcut rather than broad strength. That is an unlabeled interpretation. The source instead stresses that the exploit is 'still a valuable piece of information' and that Astra is not reliant on the card. The claim 4 excerpt also does not cover the 'more complex games' part; that text appears elsewhere in the source. Smallest repair: label the line as Escape Meridian interpretation, or rephrase it in Epoch's own terms.

Allowing the card by default would reduce EBR-bench to a simple measurement of whether a model can figure out this one single strategy. To be clear, this is still a valuable piece of information.

media — pass: No media was selected, and no verified media candidates were supplied. Nothing was invented.