Sign InOpen Brain
OpenAIOfficial ReleaseOfficial Source

How enabling two settings tripled our scores on the ARC-AGI-3 benchmark

Two API settings—reasoning retention and compaction—reportedly tripled GPT-5.6’s ARC-AGI-3 score. Agent evals should treat runtime configuration as part of the tested system.

OpenAI · Jul 29, 2026
Open Source Open MarkdownOpen JSON
Source Summary

OpenAI says enabling **reasoning retention** and **compaction** produced **3× ARC-AGI-3 scores** for GPT-5.6 while also improving efficiency.

Practical Implication

Record these settings alongside the model name in agent evaluations. Configuration can materially affect results, so defaults and explicit settings should not be compared as equivalent systems.

Agent-Ready Context
OpenAI says enabling **reasoning retention** and **compaction** produced **3× ARC-AGI-3 scores** for GPT-5.6 while also improving efficiency.

Record these settings alongside the model name in agent evaluations. Configuration can materially affect results, so defaults and explicit settings should not be compared as equivalent systems.

The supplied material provides no absolute scores, token usage, latency, or experimental detail, leaving the size and generality of the efficiency gain unclear.
Connected Context · Feed7 Judgment

This makes inference configuration part of the evaluated system: the same named model can produce materially different benchmark results when reasoning retention and compaction change. It reinforces prior warnings that infrastructure and inference protocols can move scores independently of model identity, while the absent absolute scores, compute accounting, and experimental detail prevent judging effect size, efficiency, or generality beyond the reported setup.

Quantifying infrastructure noise in agentic coding evalsBoth show that benchmark scores depend materially on the surrounding system, not only the model; together they require reporting inference settings as well as container resources before comparing results.Test-Time Scaling in Reasoning LLMs: Inference Regimes, Evaluation, and ReproducibilityThe reported threefold change is a concrete example of why the paper calls for full inference-protocol disclosure, compute accounting, uncertainty, and replay artifacts.Sample More, Reflect Less: Self-Refine and Reflexion Lose to Repeated Sampling at Equal Token Cost, from 1.5B to 7BThe candidate’s token-matched comparison requirement highlights a missing control here: without token usage or equivalent compute accounting, the claimed efficiency improvement cannot be evaluated against the score gain.
Context Map
benchmark#agent-evals#benchmark-integrity#context-caching
Uncertainty
The supplied material provides no absolute scores, token usage, latency, or experimental detail, leaving the size and generality of the efficiency gain unclear.