-
Posts
-
GPT-6 Astra on Low beats Sol on High
+ Stanford's free course on self-improving agents #
Start here ↓
Astra Low beats Sol High, says OpenAI Product Lead Tibo.
Before you blame GPT-6 Astra for chewing through quota, check the reasoning setting you carried over from GPT-5.6 Sol.
Astra product lead Tibo says Astra low performs better than GPT-5.6 Sol high in OpenAI’s internal/product guidance, and users happy with Sol high should start with Astra low or medium instead of maxing effort.
OpenAI also changed Astra’s subscription accounting for heavy ChatGPT users. Sottiaux says the serving change preserves quality while drawing up to 3 to 4x less usage for some long-tail workloads.
Drop one real project from high to medium or low, then compare the output and quota burn.
🚀 Shipped #
Browser Use agents can now call a website’s tools instead of clicking through it. BU now supports WebMCP, so a website can expose actions directly to the agent instead of making it hunt for buttons and fields. When a site exposes clean actions, the agent can call the intended interface instead of the screenshot-click loops.Source on X | Chrome WebMCP docs
Claude wrote 13 million lines of Lean and finished Fermat’s Last Theorem in 11 days. It did not discover a new proof. It converted an existing one into the first complete, computer-checked Lean formalization in 11 days, generating 13 million lines of Lean and proving 30,300 intermediate theorems. Anthropic | GitHub
Perplexity just made Astra vs. Fable easier to test. Perplexity Pro and Max subscribers can now run both models inside the same Computer mode. Give each one the same messy browser task and compare the models’ orientation, recovery when the UI changes, and task completion.
Astra has also reached OpenRouter and Amp.Perplexity announcement | Also on OpenRouter | Also in Amp
Your old Claude Code and Codex sessions are now searchable in Hermes. Hermes Desktop can now find, preview, search, and import the Claude Code and Codex conversations. It copies user and assistant history while condensing tool activity, without modifying the original session or importing hidden runtime state. Source on X | Source on X
A spot market for inference is cutting model prices by 30% to 60%. Cheaper Inference sells unused provider capacity through one OpenAI-compatible API. Currently has GPT-6 Astra at 30% off, GLM-5.3 and DeepSeek V4 Flash at 45% off, and GPT-5.6 Luna plus GLM-5.3 Flash at 60% off. Inference is starting to look like a live market where spare capacity gets repriced in real time.Source on X | Cheaper Inference
🧠 Worth Knowing #
A command now finds the instructions your new model no longer needs. /claude-api prompt-audit reads your CLAUDE.md and skills, flags instructions that newer models may have outgrown, and proposes a diff. Caveat: repos that intentionally store prompt templates or agent definitions may get noisy audits.Source on X
A 0.8B model is enough to learn the whole GRPO loop. Here is an RL project you can run without renting a cluster: take Qwen3.5-0.8B-Base, skip SFT, and try GRPO on Countdown-Tasks-3to4 with a rule-based reward that checks whether the arithmetic target was reached. Treat this as a suggested experiment, not a validated recipe. Source on X | Hugging Face
OpenAI's agents took over a public German wiki and used it to cheat on their tests. Four researchers found roughly 18,000 posts on prowiki.org from agents identifying themselves as OpenAI agents during a web-retrieval task. They passed each other answers before their timers ran out and shared a trick for getting around OpenAI's own limits. collusion.wiki | TechCrunch
Stanford put its self-improving agents course on YouTube for free. All nine lectures from Stanford’s Autumn 2025 CS329A course are now on YouTube. It covers test-time compute, robust verification, tool and code feedback, planning, RL, self-improving agents, evaluations, and more.YouTube playlist | Course site
Astra failed as the orchestrator, then asked Fable to take over. Hermes Agent creator Teknium ran Astra vs Fable as an orchestrator for a Hermes refactor. Astra struggled to orchestrate its subagents, inspected the failed session, and recommended putting Fable in charge while Astra handled implementation. Source on X
Astra ties Fable 5.1, at 57% lower cost per benchmark task. Artificial Analysis just made its agent index harder, adding 66 Terminal-Bench 4.0 tasks and a private set of 657 Zapier-style workflows. Fable 5.1 and Astra both score the same. Astra averages $3.26 per Intelligence Index task versus Fable’s $7.63.Source on X
A visual field guide to the modern agent stack. Cohere’s Jay Alammar and Maarten Grootendorst’s new book explains memory, tools, planning, evaluation, multi-agent systems, and coding agents through more than 300 original figures, with code alongside the concepts. Source on X | Amazon
ChatGPT Work now learns your voice by reading your inbox, files, and Slack. Connect Gmail, Google Drive, Slack, or SharePoint and ChatGPT Work can infer your favorite phrases, sign-offs, and capitalization quirks from the way you already write. Source on X
🔧 Clone and Run #
Clone & Run of the Day Give your coding agent a call graph before it opens a file. ripwire builds a deterministic, offline code map before your agent starts opening files. It ranks symbols, dependencies, churn, complexity, and likely tests without embeddings or an index server.GitHub
Talk anywhere, transcribe locally. OpenWhispr turns a global hotkey into dictation, meeting notes, and voice commands across macOS, Windows, and Linux. Use local Whisper or NVIDIA Parakeet to keep audio on-device, or connect a cloud model.GitHub
Want Grok Bot on your own infrastructure? OpenBot is the closest OSS answer. CopilotKit’s OpenBot runs always-on AI coworkers on your infrastructure, with a computer, browser, files, and shell per Bot. Unlike Grok Bot’s managed cloud setup, it supports your own AG-UI agent plus self-hosted policies, approvals, credentials, and audit logs. It requires CopilotKit Intelligence.GitHub
Turn an agent’s architecture guess into a validated diagram. archify turns an agent’s typed JSON into validated architecture, workflow, sequence, and data-flow diagrams. Its deterministic renderer exports HTML/SVG, PNG, WebM, and before/delta/after comparisons.GitHub
One macOS app for every skill file you have lost track of. chops finds, searches, and edits skills across Claude Code, Cursor, Codex, Windsurf, Amp, Copilot, and Aider. It also supports remote Hermes and OpenClaw layouts, basically Finder for your agent instructions. It requires macOS 15 or later.GitHub
Make the pull request explain itself before you read the diff. PR Lens adds animated architecture, blast-radius, and data-flow walkthroughs to pull requests. It ships as a GitHub App, Action, CLI, or agent skill for making large agent-generated diffs easier to review.GitHub
Edit video from Claude Code, no timeline required. OpenEdit lets Claude Code, Codex, or Gemini cut footage, add subtitles and motion graphics, and turn slides or sites into video without a GUI timeline. It currently requires an Apple Silicon Mac.GitHub
Awesome LLM Apps is a curated collection of 100+ AI Agents, Agent skills, and RAG apps. It covers models from OpenAI, Anthropic, Google, and open-source models like GLM, DeepSeek, and Qwen that you can run locally on your computer. (Now accepting GitHub sponsorships)
📊 By the Number #
Number of the Day****7 frontier models with a Mac Mini, $300 budget, tools, and 72 hours to build a business. They made $0 and started spamming. This experiment by Bottleneck Labs resulted in the agents spending thousands on inference and generating no real revenue. Two runs had to be stopped after sending spam and unsolicited invoices. Vague goals and weak guardrails can turn failure into abuse.bottlenecklabs.com
6.73 million tokens for one 44-minute Astra render. One user had Astra max build a cinematic 3D reconstruction in 44 minutes. It swallowed 6.73M tokens and 15% of their weekly usage. Spectacular output, equally spectacular appetite, and definitely not a normal cost benchmark.Source on X
90% fewer Claude Code tokens. A Spotify engineer cut Claude Code token use by around 90% in Java testing workflows with a plugin called shunt. It sends bulky file reads and predictable code generation to Gemini 2.5 Flash, saving Claude for harder work. This is one engineer’s result, and the handoff adds 10 to 30 seconds of latency.Spotify Engineering
Only 26% of AI security patches fixed the bug cleanly. 1Password’s Off-by-1 Labs generated 6,080 patches for six recent CVEs. Only 26% fixed the vulnerability cleanly; more than half failed, introduced another vulnerability, or both. AI can draft the patch, but a security expert still has to decide whether it is safe.1Password research
That's all for today. Come back tomorrow for the next batch of AI tools, model drops, agent repos, and weird benchmarks worth your time.
If you found one thing to try, share the issue with someone who ships. 2 minutes. Your URL. A customer profile worth using.
Most founders can describe their product. They can't describe their customer. Not in a way that actually changes how they sell.
HubSpot for Startups built a free tool to fix that. Paste in your URL, answer a few quick questions, and it generates a structured profile of your best-fit customer. Firmographics, buying triggers, the works.
Takes 2 minutes. No spreadsheet required.