Anthropic says Claude Sonnet 5.5, the second model in the Claude 5.5 family, generates output more than 30% faster than Sonnet 5 and costs up to 30% less per task for most work, at the same $2 per million input tokens and $10 per million output tokens. Anthropic says it scores 70.6% on Terminal-Bench 4.0, versus 10.3% for Sonnet 5, and lands two points below Opus 5.5 on GDPval-AA. It is available now, including on AWS, Google Cloud and Microsoft Azure and on the Claude Platform as claude-sonnet-5-5, and Claude Haiku 5.5 is due in the coming weeks.
AMD says it has signed a definitive agreement to acquire World Labs, the San Francisco lab led by Fei-Fei Li that builds spatial-intelligence models that generate, reconstruct and simulate interactive 3D environments from text, images and video. The all-stock deal is valued at about $8.2 billion and is expected to close by the end of 2026, subject to regulatory approvals. After closing, Li will join AMD as executive vice president and chief scientist, reporting to CEO Lisa Su.
Shopify's developer changelog says browser agents can now read and update the active checkout in a buyer's browser session with WebMCP tools, including get_checkout, update_checkout and complete_checkout, which submits the order after the buyer confirms. When the buyer's input is needed, such as for 3D Secure authentication, the tools hand control back to the buyer. With the storefront and cart tools already live, agents can now cover the full path from product discovery to order confirmation, and Shopify says the tools need no new API or merchant configuration.
SiMa.ai, which makes chips and software for physical AI in humanoid robots, cars and drones, says Fidelity Management & Research Company and Amplify co-led the round, bringing its total funding to $500 million. It says the money will scale Palette Neat, its agentic software environment for physical AI, and fund next-generation hardware slated for the first half of 2028 that it says will deliver 1,000 dense TOPS.
OpenAI's ChatGPT release notes say users can now select a chart, metric or record in the Health tab to get personalized explanations and insights based on their connected health information. Health is available to logged-in Free, Go, Plus and Pro users in the U.S. aged 18 or older on web and iOS, and is rolling out gradually.
OfficialTalkToMe, a macOS menu-bar app, lets your coding agent call you for a voice conversationSep 28
TalkToMe, an open-source app posted to Hacker News by its developer, lets an agent already running in Codex, Claude Code, Hermes Agent, OpenClaw or another host ring you with a talktome call command, then passes what you say into that session and reads the replies aloud. The app does not supply a language model; speech runs locally with Whisper and Kokoro or through ElevenLabs, and the disk image is for Apple silicon Macs.
OfficialiReadAll launches a free AI reader that picks a narrator voice to fit your textSep 28
iReadAll reads articles, PDFs, EPUBs and other documents aloud with synchronized highlighting, and says it prepares an AI narrator based on each source's tone and language. A free account includes 200,000 characters a month with saved sources and reading progress, and a guest preview allows 20,000 characters without an account.
OfficialCorral, a Linux CLI, makes sure no process started by an AI agent's command is left runningSep 28
Corral runs a command with a time limit and, before it returns, verifies that every process the command started has exited, returning exit code 120 if it cannot prove that. Its developer built it for AI coding agents and CI jobs that leave behind background jobs, daemons or processes holding output open; it uses cgroup v2's cgroup.kill when it can and otherwise acts as a child subreaper. Version 0.1.0 needs Linux 5.11 or later and is not a security sandbox.
OfficialJylus turns logs and changing data into proof-linked evidence packs for AI systemsSep 28
Jylus ingests logs, OpenTelemetry, Docker events, telemetry or JSON and returns a Context Pack of facts, timeline, conflicts and missing evidence tied to proof IDs through its /api/v1/analyze endpoint. The company says a trial needs no account and runs in a one-time isolated tenant that is then deleted, and it lists 1 GB of free storage in a Sydney data region. Benchmark figures on its site, such as 50.695% NDCG@10 on QASPER, are its own measurements.
OfficialMuseTogether lets different people's Meta Muse agents message each other to coordinate plansSep 28
After you paste a setup note into your Meta Muse chat, MuseTogether lets your Muse exchange messages with friends' Muses to find a time and place, then brings you a plan to approve. Both people need their own Muse, and the site says MuseTogether only delivers the messages and cannot read them. Its developer says it is free for now, with no paid plan.
Reported by IT HomeKuaishou's Kling 4.0 video model is due in October, with Flash in early access, per IT HomeSep 28
IT Home reports that Kuaishou's Kling AI says Kling 4.0 will officially launch in October, and that Kling 4.0 Flash opened for limited early access on September 28 to Black Gold annual members. Kling 4.0 can generate videos up to 30 seconds long in one pass and combine up to 10 images, 5 videos and 7 subjects as references in a single task. 4K and 1080p 10-bit HDR output is listed as coming soon.
Reported by ReutersOpenAI scraps October release of GPT-6.1 Astra after internal safety tests, OpenAI confirmsSep 28
OpenAI has scrapped GPT-6.1 Astra, which was planned for an October debut in ChatGPT and Codex, after internal testing found it did not meet the company's safety and alignment standards, Reuters reports. The Wall Street Journal first reported the decision, and OpenAI confirmed it. Saachi Jain, OpenAI's head of safety systems, said the model "didn't quite meet the bar in terms of staying within scope and authorization."
OfficialRange opens container images and Hugging Face model repos without downloading them firstSep 28
Range, an open-source (Apache-2.0) personal project by AWS engineer Andrey Grehov, opens a shell in a container image, Hugging Face repository or S3 or HTTP environment and transfers only the bytes a program actually reads. In his EC2 tests, a Python script read one tensor of the 1.03 TB Kimi K2 in 3.4 seconds while moving 9.5 MB of the model, and a llama.cpp chat demo answered in 6.7 seconds versus 18.3 seconds for docker pull plus hf download, with the image indexed once. It runs on Linux and on macOS through a small Lima VM; the figures are the developer's own.
Reported by TechCrunchModal Labs nears a $750 million round led by Accel at a $15.75 billion valuation, per TechCrunchSep 28
TechCrunch, citing a source with knowledge of the funding, reports that AI inference provider Modal Labs is nearing a $750 million round led by Accel at a $15.75 billion valuation that includes the investment. That would more than triple the $4.65 billion valuation Modal reached with its $355 million round four months ago. Modal Labs declined to comment.
OfficialAutoTrust AI releases JEV-27B, an open-weights decision model for self-hosted AI agentsSep 28
AutoTrust AI's JEV-27B, released under Apache-2.0, adds a 108.9M-parameter decision block to a frozen Qwen3.8-27B so it can answer yes/no, multiple-choice and 0-5 rating questions in one forward pass with a calibrated probability for each option, while leaving the base model's generation path unchanged. The company says the block trained in about 9.2 hours on one NVIDIA B200, and that JEV-27B averaged 84.07% on six public decision benchmarks versus 83.85% for TypeSafe's hosted Jev 1.13, the model it was distilled from; AutoTrust ran both comparisons itself.
Reported by 9to5GoogleGoogle will turn Gemini Gems into skills starting November 17, per 9to5GoogleSep 27
9to5Google reports that a notice in the Gemini app's Gems manager says Google will begin automatically migrating Gems to skills on November 17, 2026, and that Gems can be used until they migrate. Skills, introduced with Gemini Spark, can be invoked with a / in the prompt box, and several can be used at once. TechCrunch reports that users won't have to do anything to make the transition.
Daily AI news, sourced