Physix Frontier · News Briefing Card (Hacker News · Nov 4, 2026)

Nvidia study: AI agents keep accuracy after context compression

KEY FACTS

Nvidia researchers released a preprint testing how models perform agent tasks under short contexts.
The tests covered 13 benchmarks, including two closed-source models, Claude Opus 4.5 and GPT 5.6 Sol.
On long-horizon tasks, the two models' accuracy shifted between -3.7 and +5.6 percentage points.
On long-input tasks, Opus 4.5's recall on RULER rose from 62.5% to 95.7%.
Open-source models generally lost SWE-bench Verified accuracy under the shortest context settings.

KEY DATA

3.6x-7x256K context cost per token multiple (vs.

62.5% to 95.7%Opus 4.5 RULER recall

53.0% to 3.6%Qwen 3 Coder Next SW

$420.00 to $295.75Opus 4.5 SWE-bench cost

PHYSIX OBSERVATION

The long-context race may be heading in the wrong direction. Nvidia's data shows that agents can preserve accuracy through compressed summaries and selective reading, while also saving substantial compute. But open-source models collapse under extreme compression, which shows that the compression strategy itself is a technical barrier. If vendors follow suit, API pricing and hardware demand will both be revalued.

Source: Hacker News report