AI.info
AI Tools Updates — Page 9
What changed in the AI tools this week — dated, from the makers’ own changelogs.
- Speakeasy SDK Generation — Workload sessions are labeled on the MCP Sessions page, and Platform MCP brings existing servers under management
MCP Sessions labels workload sessions, and Platform MCP can bring existing Claude Code servers under management.
- Speakeasy SDK Generation — Spend by product for pay-as-you-go organizations, and inference scans that stop redoing work
Pay-as-you-go organizations can view spend by product, and inference hooks skip only previously accepted content.
- AirOps — Reusable Campaigns, on-demand page tracking, and connected content reviews
Duplicate Campaigns, build Workflows with Claude Sonnet 5, send content review notifications to Slack, and capture page changes on demand.
- Google Vertex AI — CodeMender updates (v0.9.0)
Updated cm report --format html; added --open, scanning support for C#, Rust, Kotlin, Ruby, and PHP, and per-turn latency metrics.
- Tines Workflow Automation — 📸 Upload and capture images from camera in mobile chat workflows
Upload and capture images from a camera in mobile chat workflows.
- Ask Sage — v3.9.0
Reliable large file uploads (no gateway timeouts); cleaner API streaming errors
- Clay — Product Roundup Week of September 14, 2026
Save and reuse enrichment logic as Functions; change login email, sign-in method, and password; automatically purchase credits when balance runs low.
- LlamaIndex — v0.14.25
- Hermes Agent — Hermes Agent v0.21.4 (v2026.9.21)
Added a host-wide gateway singleton lock and had Desktop attach to the running host backend.
- Intercom Fin — See why an outbound call didn't connect
Outbound calls that don't connect now show a specific reason on the conversation; calls that can't be placed show why.
- Grok — Grok 4.7
Grok 4.7 is now available on the xAI API, with a 500k context window and reasoning effort supporting low, medium, high, and xhigh.
- AssemblyAI — Four NVIDIA Nemotron Models Now Available on the LLM Gateway
Four new NVIDIA models are now live in LLM Gateway: Nemotron 3 Nano 30B A3B, Nemotron 3 Super 120B A12B, Nemotron Lightning 3.5 30B A3B, and Nemotron Nano 9B v2.
- AssemblyAI — New LLM Gateway Models: NVIDIA Nemotron, DeepSeek v4.1 Flash, GLM 5.3, & GLM 5.3 Flash
Four new models are now live in LLM Gateway: NVIDIA's Nemotron models, DeepSeek's v4.1 Flash, and Zhipu AI's GLM 5.3 and GLM 5.3 Flash.
- Gram — Spend by product for pay-as-you-go organizations, and inference scans that stop redoing work
Pay-as-you-go organizations get product-level spend breakdowns, and inference scans skip previously accepted content.
- Gram — Workload sessions are labeled on the MCP Sessions page, and Platform MCP brings existing servers under management
MCP Sessions labels workload sessions, and Platform MCP can bring existing remote servers under management.
- Crush — v0.96.0
Added themes with live preview and the ability to create and edit custom palettes.
- OpenHands Cloud — openhands: v0.69.0
Bumped the OpenHands image tag to 1.62.0.
- Snowflake Data Cloud — Cortex AI Function Evaluation for measuring quality (Public Preview)
- Vapi — What's New: Week of September 21, 2026
Set serviceTier to fast to request OpenAI’s faster processing tier; call logs include provider request IDs for BYOK models.
- ComfyUI — v0.37.0
- Ada AI Agent — Dashboard tables remember your sort
Sortable tables in the dashboard now keep the column sort you choose.
- Riverside.fm — Dictionary
Correct a name or term once in your transcript and save it to your Dictionary, so it comes out right in every transcript from then on.
- Kilo Code — v7.7.6 (release)
Reintroduced automatic session cleanup and added creating Git projects and cloning repositories from Agent Manager.
- Tonic Structural — v1872
Bug fixes and other internal updates.
- Palantir AIP — AI FDE now includes tools to read and create or replace cover pages for Projects.
- Fern SDK Generator — 3.96.0
Object types with XML encoding metadata are generated as classes with toXml(), a static fromXml() parser, and a fluent Builder.
- Lovable — Use your connected tools from Lovable in Telegram
Lovable in Telegram can work in connected tools; it reads without asking and asks for approval before it creates, sends, or changes anything.
- Lovable — Jev for typed decisions in your app’s AI features
Use Jev Latest to return a category, a score, or a yes/no probability instead of generated text.
- Lovable — Chat with Lovable across your workspace or inside a project
Chat without changing code, ask about apps and workspace, work in connected tools, create files, and send work to a project.
- Lovable — Manage agent permissions from your account settings
Review and change every agent permission from one page, for each connector available in your workspace and for creating or editing projects from a chat.
- Lovable — Set a default design system for each group
Workspace admins and owners can set a default design system for a group; members get it pre-selected when they create a project.
- Hugging Face — tokenizers v1: encode, decode and scaling, measured
tokenizers v1 covers encode, decode, and scaling.
- ElevenLabs — ElevenAgents
Added enable_parallel_tool_calls, gpt-6-astra, Slack alerting, phone number search, GPT Image 2.5 models and cascade_timeout_seconds.
- ChatGPT — Privacy Center in ChatGPT
Privacy Center brings together information about chat privacy, memory, personalization, data use, connected apps, and account security.
- Napkin AI — v1.2.0 (2026-09-20)
Failed visual endpoint calls return JSON errors; X-API-Version is sent on every response; 429 errors use a code.
- Letta — Release v0.32.14
Removed the /remember command.
- Higgsfield AI — Grok 4.7 in Supercomputer
Grok 4.7 is now available in Supercomputer.
- Typefully — Choose Who Can Reply to Your X Posts
Added settings to choose who can reply to X posts published with Typefully.
- Open Interpreter — Open Interpreter 0.0.45
Adds DeepSeek Flash and DeepSeek V4 Pro to the model picker; DeepSeek Flash supports image input and a 1M-token context window.
- Keploy — v3.6.67
Replay now reports offline coverage and what each replay proved; generic exchange boundaries use the wire clock.