AI news · Issue 2026-11-13-06-00
Version 1
In Brief
Details
OpenAI — ChatGPT dots get custom instructions
OpenAI
Dots in ChatGPT can now follow user-set custom instructions and rules.
- ChatGPT dots can now follow custom instructions, for example how to behave or which actions need the user's go-ahead before they run.
- Instructions can be given directly in the dot's thread, such as "Show me a draft before sending any emails."
- Custom instructions and rules can also be set in Customize > Behavior on mobile, or through the dot's menu on web and desktop.
- OpenAI's description of the update: faster replies, less jargon and better follow-through.
Sources:ChatGPT on X✓ official
Google — Gemini 4 Hydrogen released
Google
Google added Gemini 4 Hydrogen, a fast model for everyday coding and agentic work, across its products.
- Gemini 4 Hydrogen (gemini-4-hydrogen) is generally available in the Gemini API and Google AI Studio, including the free tier.
- Paid pricing is $1.50 per 1M input tokens and $7.50 per 1M output tokens, flat across the 1M-token context window.
- Adds a fourth xhigh thinking level alongside low, medium and the default high; it is now the default model for the Antigravity agent in Gemini Managed Agents.
- Google's results: 2.5x faster than Gemini 3.8 Flash; 51 on the Artificial Analysis Intelligence Index at default thinking (54 at xhigh) vs. 50 for Argon.
- Rolling out in the Gemini app, Gemini for business and globally in Google Search; Google also introduced Orchestrator in Antigravity, which switches between Carbon and Hydrogen in one session.
- Available in Cursor; Cursor's results: 54.3% on CursorBench at default effort vs. 39.6% for Gemini 3.8 Flash.
Sources:Google blog✓ officialGemini API changelog✓ officialGoogle Search blog✓ officialCursor on X✓ official
Anthropic — Claude Meetings in beta
Anthropic
Claude can join Google Meet calls as a participant on Team and Enterprise plans.
- Claude Meetings is rolling out in beta on Claude Team and Enterprise plans, in Claude and Claude Desktop.
- Claude can be added to a Google Meet call as a visible participant.
- During the call it takes notes, answers questions and builds artifacts such as docs, whiteboards and slides.
- Artifacts are built while the conversation is still going, according to Anthropic.
Sources:Claude release notes✓ officialClaude on X✓ official
Anthropic — self-hosted Claude Managed Agents
Anthropic
Claude Managed Agents can now run entirely on customer infrastructure, in private beta.
- Self-hosted Claude Managed Agents is in private beta.
- Customers run the whole Managed Agents service on their own infrastructure, in their cloud or on-premises, not only the sandboxes.
- Agents, sessions and session history are stored on the customer's side.
- The server sends model requests to a provider the customer chooses; Anthropic is the default.
- Access is requested through the Anthropic account team.
Sources:Claude Platform release notes✓ officialClaudeDevs on X✓ official
OpenAI — Codex CLI 0.168.0
OpenAI
New Codex release with model provider selection, side-by-side TUI panes and audio attachments.
- Model providers can be browsed under Sessions > Model providers, with a default chosen for new chats; open and resumed chats keep their saved provider.
- features.tui_workspace shows a side conversation next to the main chat in resizable panes.
- Recordings can be attached when using an audio-capable model.
- features.image_generation_animation generates or edits animated GIFs with frame count and timing controls.
- Local history and resume pickers now filter new threads by the signed-in ChatGPT account.
Sources:Codex release rust-v0.168.0✓ official
Thinking Machines — Inkling 2 Small open weights
Thinking Machines
Inkling 2 Small weights are on Hugging Face under Apache 2.0.
- Inkling 2 Small weights are available on Hugging Face under the Apache 2.0 license, in NVFP4 and MXFP8 checkpoints.
- Multimodal MoE model: text and image input, text output; 594B total and 17B active parameters, 8 of 384 experts per token.
- Context window of up to 1M tokens.
- The NVFP4 checkpoint needs at least 500GB of aggregated VRAM (e.g. 2x B300 or 4x H200); MXFP8 needs at least 800GB.
- Local inference recipes for SGLang, vLLM, TokenSpeed and Unsloth; fine-tuning via Tinker.
Sources:Hugging Face model card✓ officialThinking Machines on X✓ official
Microsoft — MAI-Transcribe-2 generally available
Microsoft
Microsoft AI's transcription model moved from preview to general availability in Microsoft Foundry.
- MAI-Transcribe-2 is now generally available in Microsoft Foundry for production use; it launched in public preview on Sep 3.
- Price: $0.10 per hour of audio (limited-time pricing).
- Covers 60 languages with speaker diarization, word-level timestamps and contextual biasing.
- Microsoft's results: 3.4% average WER on FLEURS (top 25 languages); 1 hour of audio transcribed in about 10 seconds of inference.
Sources:MAI-Transcribe-2 model page✓ officialMicrosoft AI on X✓ official