Billing data from 70,000 companies shows the money flowing to Opus 4.8 and Sonnet, not to the newest and most capable model on the shelf.
Drew Breunig argues Fable broke the habit of waiting for a cheaper model to fix your harness, forcing real routing decisions.
The Linux commit message says the model kept declaring the bug unsolvable, then kept working when pushed.
Willison reframes the agent review problem as instructing confidently and verifying confidently, by whichever method actually works.
Chat captures intent but buries the work; a canvas makes a long agent run visible, steerable and cheaper to correct.
Authentication, Actions, APIs, pull requests, issues and Copilot all went down together, putting the assistant on the critical path.
Anthropic v1.0.0 follows OpenAI v3.0.0 from httpx to httpx2, breaking transitive dependencies across the plugin ecosystem.
Clips of up to 30 seconds generated from documents, slide decks and web pages, after a public beta running since August 6.
The framework calls assisted work copyrightable and synthetic performances not, and asks the industry to stop using one word for both.
Direction, voices and live-action stayed human across a three-month build; Guinness declined the record for lack of a category.
Ten finalists present in Los Angeles on September 25; the winner develops the feature with Range Media against a budget reaching $15M.
Interactive 3D scene building paired with 2D generative models, pitched by fxguide as the moment AI filmmaking becomes directable.
WorldClaw returns instance-level meshes rather than a fused render, though no code, weights or demo have been published yet.
Ultimate EFS bundles shared storage, media management and FLOW AI in one box, with transcription now across 13 languages.
Hyperion 2.5 targets banding and crushed blacks in generative and archival footage, shipping in Astra ahead of Topaz Video.
Tedial will let external assistants search repositories, pull metadata and trigger workflows from outside the platform.
Automatic segment extraction lands beside generated ad spots, next-generation captioning and a European debut for voice translation.
AssemblyAI, Deepgram, ElevenLabs and Whisper become selectable per job, alongside IMSCR support and batch burn-in.
Ingest, real-time analysis, metadata, clipping and measurement in a single cloud workflow, with human-in-the-loop options.