|
|
TOOL
MAJOR
2026-09-29
ChatGPT Space and Pages — OpenAI's shared workspace for teams and agents
OpenAI puts documents, files and agents in one shared place inside ChatGPT, aimed straight at Microsoft Office.
What is it?
ChatGPT Space is a shared workspace where teammates, ChatGPT and each person's dot work from the same pile of project knowledge. Inside a Space, Pages is a new kind of document built for people and agents to edit together, with charts, images, checklists and dashboards.
How does it work?
Pages and files live together in a Space like a shared drive — but tasks can be assigned to pages and ChatGPT can build them from the work a team needs done. A Notion-style slash menu and Space-wide context let agents and people work from the same project knowledge.
Why does it matter?
OpenAI now ships its own answer to a word processor and shared drive, putting it in direct competition with Microsoft's Office suite. For teams already using Dots, a Space gives those agents a shared place to read context and hand back finished work.
Who is it for?
Teams on ChatGPT Pro, Business and Enterprise.
|
|
|
|
SECURITY
MAJOR
2026-09-29
Anthropic tests GLM-5.3 — its safeguards fall to simple tricks up to 100%
Anthropic says an open model now matches its restricted cyber model closely, but anyone can remove GLM-5.3's safety limits.
What is it?
Anthropic published a security study of GLM-5.3, Z.ai's open-weight coding model. The report finds GLM-5.3 can build working end-to-end cyber exploits at a rate close to Claude Mythos Preview, Anthropic's restricted cyber model.
How does it work?
On ExploitBench, GLM-5.3 produced end-to-end exploits in 12% of attempts versus 14% for Mythos Preview. A false cover story bypassed its safeguards 64% of the time, prefilled tokens 92%, and an abliterated copy 100% — at a cost of about $4,400.
Why does it matter?
Frontier-level exploit building is now in a model anyone can download. Anthropic notes several abliterated versions went public within days of GLM-5.3's release, and calls for government safety testing of its successors.
Who is it for?
Security teams, AI policy researchers, and open-model developers — assume attackers have these capabilities today.
|
|
|
|
MODEL
MAJOR
2026-09-28
Eleven v4 — ElevenLabs' speech models cover 90+ languages and clone from 10s
ElevenLabs' fourth-generation voices speak more languages, clone faster and take stage directions inline.
What is it?
Eleven v4 is ElevenLabs' most expressive text-to-speech model, released September 28, together with Eleven v4 Turbo for real-time voice agents. Both support 90+ languages (up from 70) with the biggest gains in Japanese, Brazilian Portuguese, Mandarin and Cantonese.
How does it work?
Inline tags like [laughs] or [said angrily in French accent] steer delivery, and several tags can be stacked. Instant voice clones need only 10 seconds of audio, and Turbo hits about 100 ms median inference latency.
Why does it matter?
Faster, more expressive speech at wider language coverage means fewer gaps in voice-agent products. ElevenLabs claims v4 ranked first on Artificial Analysis in September and won 65–81% of blind preference tests.
Who is it for?
Voice-agent builders, dubbing and audiobook teams. Creator plans and above get 3× credits until October 12 to try the new models.
|
|
|
|
SHOWCASE
MAJOR
2026-09-29
America.gov — the US government's AI chatbot runs on Gemini and Grok
The White House wants one AI chat box to be the front door to every federal service.
What is it?
America.gov is a new AI-powered site from the Trump administration that answers questions about US federal services — booking a campsite, getting a passport, replacing a Social Security card. It launched September 29 alongside an executive order making it "the single point of entry for Americans to access covered services online."
How does it work?
The chatbots run on Google Gemini and xAI Grok, with a knowledge base from about 29,000 government websites. Zero data retention agreements stop the AI providers from keeping or training on prompts; responses are cached by prompt hash for up to two hours.
Why does it matter?
This is one of the largest public deployments of an AI chatbot as a government front door. Google says Gemini will help over 100 million people reach public resources. Task completion (name changes, applications) is planned for 2027.
Who is it for?
US residents, gov-tech teams, and anyone watching public-sector AI deployment.
|
|
|
|
TOOL
MAJOR
2026-09-29
OpenAI Decisions API — GPT-6 Luna picks from your answers in 150 ms
OpenAI's answer to Jev: an endpoint that returns a choice, not text, about ten times faster than a normal Luna call.
What is it?
The Decisions API, announced at DevDay September 29, does not write free text. The developer sets a question and a fixed list of possible answers, and the API returns which answer fits best — OpenAI's response to TypeSafe AI's Jev decision model.
How does it work?
A specialized GPT-6 Luna reads text or image context and scores only the allowed options, returning the most likely one with a confidence score. Skipping text generation cuts latency to about 150 ms versus 1.6 seconds for a standard Luna call.
Why does it matter?
Classification and routing steps are often the slowest part of an LLM app. A 150 ms endpoint that can only return a valid option makes these steps fast enough for real-time use. (No price or accuracy figure published yet — benchmark before switching from Jev.)
Who is it for?
Developers building classifiers, routers and agent control loops. Currently in limited preview; broad rollout coming in days.
|
|
|
|
TOOL
MAJOR
2026-09-29
Claude Code 2.1.285 — a switch to turn off WebFetch and a --desktop handoff
Claude Code gets a kill switch for web fetching, a jump to the desktop app, and a long list of subagent and MCP fixes.
What is it?
Claude Code 2.1.285 adds CLAUDE_CODE_DISABLE_WEB_FETCH to turn off the WebFetch tool, a claude --desktop command to jump to the desktop app on the current session, and a new claude plugin configure command.
How does it work?
Admin controls get tighter: a new allowedProviders managed setting limits which API providers users can pick, and Team/Enterprise sessions now withhold WebFetch until the org policy has loaded.
Why does it matter?
Teams that need to keep agents off the open web now have one variable to set. The fix list is long and practical: fork subagents keep their plan and dontAsk modes, WebSocket MCP servers show in claude mcp list, and artifact publishes no longer overwrite newer files after a rewind.
Who is it for?
Claude Code users and team admins. Update with npm i -g @anthropic-ai/[email protected].
|
|
|
|
TOOL
NOTABLE
2026-09-29
Pi 0.99 — the minimal coding agent adds MCP and a codemode sandbox
Pi's team used to say no to MCP. Version 0.99.0 adds it, with a JavaScript sandbox that composes tool calls.
What is it?
Pi v0.99.0 lets Earendil's terminal coding agent connect to MCP servers over stdio or streamable HTTP with OAuth — a reversal of its earlier position against MCP, explained in a post titled "You Said No MCP."
How does it work?
Codemode is the key piece: instead of one tool call per turn, the model writes JavaScript that runs in a QuickJS sandbox and calls multiple MCP tools at once. Codemode activates automatically when MCP is configured.
Why does it matter?
Pi is known for its four-tool, small-prompt design, so MCP is a real change of direction. Running tool calls as sandboxed code keeps context small while letting the model combine tools — for example pairing Linear's MCP server with Jev to score issue sentiment.
Who is it for?
Developers who use or build on the Pi coding agent.
|
|
|
All releases at ai-tldr.dev
Simple explanations • No jargon • Updated daily
|
|
