2026-07-31

1. A benchmark number is a claim about a harness, and today it moved 4.9×

OpenAI published that enabling two Responses API settings — retained reasoning (keep the chain of thought between steps) and compaction (summarize old context instead of truncating it) — took GPT-5.6 Sol from 7.8% to 38.3% on the ARC-AGI-3 public task set, while cutting output tokens 6×. Same weights, same tasks, a 4.9× spread from scaffolding.

38.3% is above the 30.2% that Claude Opus 5 set as ARC Prize-verified SOTA on July 27. But nothing published says whether Opus 5 was measured with any equivalent state-preserving configuration, and ARC Prize's official scores deliberately use one standardized harness so labs stay comparable. So the two numbers were not produced the same way and do not rank the models. ARC Prize's reported position is that general settings available to all users are fair game if properly reported.

There is also a second, quieter discrepancy: the "official harness" figure appears as 7.8% (ARC Prize verified) and as 13.3% (coverage of OpenAI's own re-run). Nothing read reconciles them, so both are recorded.

Why it matters: on long-horizon agentic benchmarks the harness now moves the score more than a model generation does — which means every published agentic number without its configuration attached is unfalsifiable, and the direction evals are heading is exactly the direction that makes this worse.

→ Eval Harness Configuration (new page), GPT-5.6 Sol (and Terra, Luna), Claude Opus 5 → Source · The Decoder · ARC Prize results

2. OpenAI cuts Luna 80% and says the model paid for it

Luna drops from $1/$6 to $0.20/$1.20 per million tokens, Terra from $2.50/$15 to $2/$12, Sol unchanged with "a faster option" added. The lower rates also change how usage is counted in Codex and ChatGPT Work.

The stated cause is the part worth reading twice. OpenAI says it applied GPT-5.6 Sol to its own serving stack after GA: 20% lower serving costs from production GPU kernels Sol rewrote in Triton and Gluon from inside Codex, and 15%+ better token-generation efficiency from a speculative-decoding draft model Sol redesigned across "hundreds of autonomous experiments". The open-source FpSan sanitizer was used to verify the kernels the model wrote.

Why it matters: OpenAI endorsed the "Pacing the Frontier" statement on July 28 — a statement specifically about pacing automated AI development. Two days later it announced that its model improved its own infrastructure, framed as a cost saving and passed to customers as a price cut. Both things can be true and neither is hidden; the point is that the mechanism the statement asks labs to build a brake for is now a line item in a pricing announcement.

→ GPT-5.6 Sol (and Terra, Luna), OpenAI, Frontier Pacing → Source · CNBC · @OpenAI

3. DeepMind ships the robot brain publicly and keeps the body private

Three models across two announcements: a VLA controlling a full humanoid — legs, torso, arms and multi-finger hands — under one learned policy (July 28), an embodied-reasoning VLM (July 30), and an on-device VLA. Only the reasoning model is public, via the Gemini API as gemini-robotics-er-2-preview; both acting models are early-access partnership.

The dexterity table is the honest state of the field. On the 22-DoF SharpaWave hand: 92% unscrewing a light bulb, 36% screwing one back in, 44% tying a trash bag, 40% sealing a ziplock, 32% dustpan. Whole-body manipulation on the Apollo 2 humanoid runs 45.7% (floor pick) to 76.3% (shelf pick). ER 2 reports 91.3% on moment finding at 0.96 s mean absolute distance. On-Device 2 adapts to a new robot embodiment in hours from fewer than 200 demonstrations.

Shipping alongside: ASIMOV-Agentic, an open benchmark for whether the reasoning layer refuses dangerous commands from the acting layer, judges physical feasibility before attempting, and asks for human help when uncertain.

Why it matters: insertion and deformable objects are still unsolved — removing a bulb is near-solved, putting one back is a third as reliable. And the release shape is a policy statement: the layer that plans is a public API, the layer that moves is not.

→ Gemini Robotics 2 (new page), Gemini Robotics ER 2 (new page), Embodied Agents, Google DeepMind → Source · DeepMind · Engadget