OpenAI Releases GPT-6 Astra, Its Best Model Yet, But Most Users Must Wait

Today's AI news: OpenAI has released GPT-6 Astra, its best model yet for computer use, software engineering and visual understanding, but the rollout starts with a limited set of organisations and most users must wait days. Meta's Muse Spark 1.3 promises frontier performance almost too cheap to meter. Anthropic open-sourced Claude Commerce Agents, a blueprint for shopping and merchant agents. Runway shared GWM Worlds 2, a world model that takes arbitrary text actions in real time. And Higgsfield launched Genjutsu, applying VFX on top of raw footage. Chapters: 0:00 Today's AI News 0:27 GPT-6 Astra 1:17 Muse Spark 1.3 1:57 Claude Commerce Agents 2:36 Runway GWM Worlds 2 3:18 Higgsfield Genjutsu Blend Roundup 2026-09-03 https://x.com/thsottiaux/status/2095597168816226335 https://x.com/OpenAIDevs/status/2095596020638036311 https://x.com/sama/status/2095601211869421726 https://x.com/finkd/status/2095232032896946311 https://x.com/ClaudeDevs/status/2095233745167282602 https://x.com/runwayml/status/2095540014645920040 https://x.com/higgsfield_ai/status/2094932147257884857 OpenAI has released GPT-6 Astra: though sadly, most users won't get it for days. Muse Spark 1.3 is almost too cheap to meter, Anthropic open-sources its shopping and merchant agents, Runway's new world model takes any action you type, and Higgsfield puts VFX on your raw footage. Here's today's AI news. OpenAI has released GPT-6 Astra - but for now it's in the hands of a limited set of organisations, starting with the companies in its Daybreak cybersecurity programme. Everyone else on Plus, Pro, Business and Enterprise gets it over the coming days, along with the API and AWS, and the free tier and cheapest paid plan don't get it at all. Sam Altman's already asking for patience. OpenAI calls it its best model to date for computer use, software engineering and visual understanding, and it's the first model to hit the Critical cybersecurity threshold under OpenAI's own Preparedness Framework, which is why the most advanced cyber capabilities stay locked behind that programme. Greg Brockman is calling this the start of the AGI era. Bold claim for a model most people can't open yet. Meta has released Muse Spark 1.3. Mark Zuckerberg's pitch is that it's frontier performance that's almost too cheap to meter. It's Meta's biggest jump yet on coding and agentic work - in Meta's own engineer comparisons it uses around twenty percent fewer tool calls and twenty five percent fewer tokens than 1.2 - and it's live now in Muse Code and the API. Pricing is the standout part of the release. The standard endpoint stays at a dollar twenty five per million input tokens, but there's a contributor tier at roughly ten cents, ten to twenty times cheaper, if you let Meta train on your traffic. Cheap, as ever, has a price. Anthropic has open-sourced Claude Commerce Agents. A blueprint for building out shopping and merchant agents. With reference implementations for retail, travel, telecom and ticketing. There are two agents: a shopping agent a business embeds in its own app, so customers can search, compare, bundle products and fill a cart, and a merchant agent for the back office that maintains listings, acts on inventory and order alerts, sets prices and drafts campaigns. Every change the merchant agent makes is staged for a human to approve, which is the sensible part. It's Apache licensed and on GitHub. And Shopify has already published its own implementations. Runway has shared GWM Worlds 2, the latest version of its General World Model, which generates interactive simulations in real time - continuous 720p video at twenty four frames a second, with synchronised audio. You define the environment, the subjects and the visual style, then steer it from inside with plain text: tell a character to run, climb or leap, switch a lamp on, or set the stage on fire. Runway says it's the first world model that generalises to arbitrary actions rather than a fixed menu, and it's pitching it at games, virtual characters and robot training. It's still research, but the demo is hard to look away from. Higgsfield has launched Genjutsu, a tool that applies VFX of any quality on top of raw footage you've already shot. It works two ways: Motion Transfer pulls the movement out of any clip and builds an entirely new scene around it, and Object Swap replaces a character, object, location or style while preserving everything else in the shot. No prompt is required - presets can do the whole job - and it takes reference clips of three to thirty seconds plus up to forty reference images for characters, products and wardrobe. Higgsfield reckons this changes how AI films get made. It certainly changes how much of one you have to actually film.