Anthropic announces faster, lower-cost Claude Sonnet 5.5

#1 📝 Anthropic News

Introducing Claude Sonnet 5.5 - Anthropic announced Claude Sonnet 5.5 on September 28, 2026—the second model in the Claude 5.5 family—which it says runs 30%+ faster and costs up to 30% less for most work. Positioned as a faster, lower-cost complement to Claude Opus 5.5, it is strongest at well-scoped everyday tasks, bug fixes, and creating polished documents.

Also covered by: @Peter Yang, @Cognition, @v0, @Cursor, @Boris Cherny, @Mike Krieger, @Mike Krieger, @Anthropic, @Claude – Anthropic, @Claude – Anthropic, @There's An AI For That

#2 𝕏

Cognition announced updates making Devin 30–40% cheaper in Fusion and Normal, 15–20% cheaper in Ultra, and up to 70% cheaper in Devin Review. Cognition also said the updates improve capabilities, with Devin Fusion now ranking first on FrontierCode 1.1 Extended.

#3 📝 OpenAI News

Towards safety cases for frontier AI training - OpenAI News said structured safety documentation should be required before continuing any frontier reinforcement learning training run, ideally rising to the level of “safety cases.”

#4 ▶️

We Built Grok Bot. Here Are Our 14 Best Bots | Peng Zheng & Lauren Tan

Peter Yang

#5 𝕏

Philipp Schmid shared that the Credentials API for Gemini Managed Agents protects secrets by injecting them on the wire only for trusted domains, preventing sandboxed code from accessing raw tokens. It works with environment variables, CLIs, and MCP Server.

#6 𝕏

Andrew Ng said weak sandboxing enabled the OpenAI-Hugging Face hack and highlighted Nvidia’s release of open-source tools for sandboxing AI agents. OpenWorker, an open-source agent harness he is developing with @rohitcprasad, is building on Nvidia OpenShell to sandbox commands with deterministic restrictions and action logging.

Also covered by: @Harrison Chase

#7 ▶️

Where AI products go next: voice, agents, and self-driving software | Tara Sesha and Nan Yu (OpenAI)

Lennys Podcast

#8 in

Guillermo Rauch announced that he ported his Mini web browser to Rust and Swift, bundling up-to-date Chromium via the cef crate and embedding an agent that connects to his local fx CLI over ACP. Built on fx.sh with Opus 5.5 and Sol 6, the project uses fx to manage the browser through an MCP server. Rauch says replacing Electron enabled Liquid Glass, faster startup, and a more Mac-native experience.

#9 ▶️

I’m using Jev more than Opus 5.5 or GPT-6. Here’s why.

How I AI Podcast

#10 𝕏

LlamaIndex 🦙 shared a blog post on where frontier vision-language models fail to parse real W-2s, 1040s, W-9s, and scanned W-4s, arguing that forms require field detection, hierarchy preservation, exact box-value mapping, and handwriting and checkmark recognition. It also includes a custom LlamaParse cookbook that the source says handles forms at a fraction of the cost.

#11 𝕏

claire vo 🖤 recapped @cxodev’s testing of Team Bots, highlighting Quincy Quickquote—made by @zachdavis—which uses the company’s CRM, contract platform, and historical quotes to quickly produce custom proposals.

#12 𝕏

claire vo 🖤 shared a workflow using embeddings on titles and descriptions to find related neighborhoods of work, pair items, and build named themes from the discovered groups. The resulting analysis covered all PRs, cost 9 cents, and felt pretty accurate to her.

#13 𝕏

Santiago commented on a feature that lets anyone—not just developers—use AI to prepare code changes and submit proposals to developers for review and deployment. For example, a support worker could discover a bug, ask AI for a fix, and propose it.

#14 ▶️

Roles aren't converging—they're expanding | Tamar Yehoshua (Atlassian CPO)

Lennys Podcast

#15 𝕏

Dharmesh Shah shared that YouSpotHub can import information from LinkedIn, X/Twitter, and Granola, with Gmail integration coming soon. Its Chrome extension adds web content to users’ Second Brain, while forwarded emails can be parsed for companies, people, content, and other information.

#16 𝕏

Thariq said prompts are difficult to share in isolation because agent workflows now depend on references, skills, and examples. He often asks his agent to review 3 other repos he made, search the web for references, and use other AI APIs.

#17 𝕏

Santiago demonstrated VEED’s face-and-voice cloning feature, saying setup took 30 seconds by uploading a picture and reading text. Once ready, it turns prompts into edited talking videos with customizable subtitles, music, and duration; VEED partnered with him on the post.

#18 ▶️

The Data Engineering Professional Certificate is now on DeepLearning.AI

Deeplearning.ai

#19 ▶️

DHH has gone completely off the rails...

Fireship

#20 𝕏

Sebastian Raschka commented that pre-training is not required to improve reasoning, and that better reasoning performance—not just efficiency—can be the objective. He cited Prime Intellect’s INTELLECT-3, which used GLM-4.5-Air-Base.