AI news digest — October 29, 2026
4 items, each with its source.
OpenAI gives ChatGPT Pro and Business subscribers up to $500 monthly API credits
OpenAI announced a new benefit granting eligible ChatGPT Pro and Business users up to $500 per month in free platform API credits. The allowance matches subscription tiers and covers model calls across the Responses, Agents, Realtime, and Images APIs. Codex usage executed through standalone API keys is excluded from the credit program.
Why it matters. Developers running agent harnesses and production prototypes can subsidize background inference costs directly through their standard ChatGPT subscription fees.
explainx.aiMeta introduces SixthView benchmark to evaluate 3D spatial reasoning in multimodal models
Meta Research released SixthView, an evaluation benchmark featuring 604 interlocking-cube puzzles that require vision-language models to predict a hidden sixth face from five visible perspectives. The dataset includes 11,634 images presented across four distinct input modalities ranging from pure spatial coordinates to single composite images. Scoring relies on exact-match reconstruction grids without evaluating free-text reasoning.
Why it matters. Multimodal benchmark suites can now isolate whether 3D spatial reasoning errors stem from image feature extraction or underlying volumetric inference failures.
explainx.aiSamsung reports record quarterly operating profit as HBM4 memory sales triple
Samsung Electronics posted record third-quarter 2026 operating profits of 107.4 trillion won, driven by a surge in demand for artificial intelligence hardware components. High-bandwidth memory sales more than tripled quarter-over-quarter, with HBM4 chips growing to represent 60 percent of the division's total HBM volume. The company also confirmed that its entire 2027 HBM allocation is fully committed under fixed contracts.
Why it matters. Persistent multi-year memory supply commitments will keep high-end accelerator hardware prices elevated and capacity constrained for frontier foundation model training.
explainx.aiStartup Particle launches Particle-1 model claiming 45 percent fewer output tokens
Startup Particle launched Particle-1, a coding-focused post-train of DeepSeek V4.1 Flash built using reinforcement learning on synthetic data. The model optimizes generation efficiency to solve programming and data engineering tasks while utilizing significantly fewer reasoning and tool-calling tokens. Particle claims reductions of 28 to 52 percent in total output tokens across multiple benchmark suites without sacrificing task pass rates.
Why it matters. Compressing chain-of-thought token overhead cuts per-task inference latency and hosting expenses for multi-turn autonomous programming agents.
explainx.aiThe stories that matter for people who build with AI, with the source for every claim. Free, no spam.
Feed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.