Medium impact

Show HN: Portal2Bench – Can AI agents solve Portal 2?

What happened

Portal2Bench is a benchmark that tests AI agents on Portal 2 test chambers. The leaderboard shows GPT‑6.1 Sol and GPT‑6 Astra solving 20 of 26 chambers, GPT‑6 Terra 19, Opus 5.5 11, Sonnet 5.5 6, Fable 5.1 3 of 15, and GPT‑6 Luna 0. Each run is limited to 300 plans and $30, with costs ranging from $0.74 to $13.12 per chamber. The agents can only send keys, turn the view, and receive screenshots; they have no shell or web access.

Why it matters

The benchmark demonstrates how far current large‑language‑model agents can go in a complex puzzle game, showing both progress and remaining limits. It may inform developers about the practical cost and performance of AI agents in interactive environments.

Reality check

An automated, sceptical read of the source

Mostly marketingIncrementalNotableGenuine breakthrough

Incremental. The results show modest improvements over previous attempts but still fall short of full game mastery, indicating incremental progress.

Who it affectsResearchersDevelopers

What to watchFuture updates may include more chambers and higher‑effort runs to push the agents further.

Do you agree with this rating?Agree Disagree

More AI news

Every Way to Get 192GB for Local AI (Without a $9,500 Mac) [video]Industry news · 14 Nov 2026DeepSeek V4 Pro 0813DeepSeek · 14 Nov 2026DeepSeek V4 Pro 0813DeepSeek · 14 Nov 2026Anthropic Upgrades Claude Code Projects to Run Parallel AI Workers Hands-FreeAnthropic · 14 Nov 2026Claude designed FPGA scaler board for 6th-gen consolesAnthropic · 13 Nov 2026

This page was collected from kingstonyip.com and summarised automatically by AIPulsee's pipeline. See how it runs.