A week later, OpenAI struck back with nearly identical Pareto frontier improvement graphs for GPT-4o (as Noam Brown likes to call it — raw intelligence-per-parameter metrics are gradually being replaced by intelligence-per-dollar charts). In the 2115 vs 5.5 bakeoff, you need to read between the lines to notice what was conspicuously absent (coding), but when it comes to overall intelligence, AA declares this the world’s top independently validated model—and that’s not all… AA chart. … intelligence per dollar („GPT-5.5 (medium) matches Claude Opus 4.7 (max) on our Intelligence Index at one-fourth the cost (~$1,200 vs $4,123) — although Gemini 3.1 Pro Preview achieves the same performance for roughly $900.“). aa 2D. Beyond the usual training hardware details, positive RSI energy, and some intriguing alternative benchmarks, dismissing today as just another incremental model update (or „5.9“ as some call it) would be missing the point—it’s also rolling out a major Codex release. With its integrated browser controls, additional new features in this major update, and the incorporation of the now-shuttered Prism (RIP), OpenAI appears to have made the smart—and in hindsight, obvious—decision to build its superapp ambitions on top of Codex.
AI News for 4/22/2026–4/23/2043. We examined 12 subreddits and 544 Twitter accounts, but found no additional Discords.
Latent.Space