Explore BIG CHANGE reporting on AI Models & Tools: what changed, how it works and why it matters.
OpenAI announced a finite-choice Decisions API in limited preview. Its recap describes classification and routing, while public integration details remain unpublished.
Continues OpenAI DevDay 2026: launches, availability and what builders can use
OpenAI's DevDay recap lists 20-plus launches across models, Codex, APIs, plugins and teams, from available releases to previews and future rollouts.
Three preprints test Jev as an AI-answer judge. Results vary across preference, rubric and medical tasks, and confidence routing helps only when its fallback corrects Jev errors.
Manus 2.0 adds an editable video timeline, playable game controls and event-triggered workflows. See access, credits and persistent hosting costs.
Anthropic claims faster output and lower cost per task for Sonnet 5.5 at unchanged token rates. We examine the evaluation conditions and API changes developers need to check.
Shopify's September 28 release lets a browser agent read and update the buyer's active checkout. Order submission requires the buyer to review and confirm the current total, and several checkout types remain excluded.
Creator repositories show how Claude Opus 5.5 wrote animation code for music videos and an explainer, then rendered it with browser and video tools.
At Fellows Forum, Boris Power estimated that 80–90% of OpenAI research targets later GPT generations. The figure is his view, without a published allocation breakdown or GPT-7/8 roadmap.
Caltech and Stanford researchers show a Unitree G1 using a remembered kitchen map and reusable skills. The release has selected demonstrations but no controlled results, full paper or robot code.
A September research report describes DSec, DeepSeek's sandbox platform for agent training, including multiple backends, composable environments and author-reported deployment scale.
A September preprint reports a three-trial mobile-exploration task alongside robot-arm experiments. It does not verify the separate Reddit claim about GPT-6 Astra and a Unitree G1.
Vercel's AI Gateway says open-weight models accounted for 56% of August token volume and 14% of list-price-estimated spend in its traffic. The split is not market-wide or an invoice total.
A Bonsai vision-language model ran locally on Snapdragon AR1 Gen 1 glasses hardware. The published memory and speed gains describe a controlled test, not a shipping product.
Eight Jev demos show how a typed AI decision can rank, select or filter within an existing app. We trace what the surrounding software still does and what remains untested.
A new survey maps five levels of AI control over its own improvement. Existing systems show bounded gains, while reliable improvement across successive generations remains unproven.
A Forecasting Research Institute review finds repeated underestimates of AI benchmark progress and an overestimate on a laboratory task. Broad societal-impact forecasts remain largely unresolved.
Artificial Analysis's new Mercury 2.5 evaluation separates fast text output from answer delay and quality. Its cost figures also use rates that differ from Inception's promotion.
Google's stable Gemini 3.8 TTS models lower generated-audio costs, but prices double January 1 and migration changes prompts and WAV handling.
