AI summary · this digest is compiled by AI, not yet reviewed by Jason
Claude Haiku 5.5 challenging GPT-6 Luna at the same price, alongside LegalOn's 65% Codex cost cut through model routing, shows small models and cost optimization becoming key levers for AI app builders.
REA packages reverse engineering into a single MCP server. Coding agents can analyze apps from runtime behavior down to native binaries and explain how a feature works. Work that once required specialist toolchains and expertise becomes accessible through agents, which is useful for competitive research, debugging and security analysis.
diagram-design is a skill for Claude Code, Codex, GitHub Copilot, Factory Droid and Pi. It offers 42 diagram types and outputs self-contained HTML+SVG as an alternative to rough Mermaid charts. Version 2.5.10 adds ten more layouts, including Sankey, giving developers polished visuals for docs, decks and landing pages.
This open-source skill tackles a common annoyance: coding agents burying the key answer in long responses. It makes output direct and easy to scan. Installation takes one line pasted into the CLI, and it supports multiple languages including Chinese. It is a cheap way to improve everyday agent workflows.
Anthropic released the small model Claude Haiku 5.5, which Latent Space rates as stronger than OpenAI's GPT-6 Luna at the same price. For cost-sensitive agents and SaaS products with heavy call volumes, competition in the small-model tier directly affects unit economics. It is a good moment to revisit model choices.
LegalOn matched Astra, Sol and Luna models to tasks by difficulty and managed budgets strategically. It cut estimated daily Codex costs by 65% without slowing development. For teams facing rising AI coding bills, routing models by task is a cost-cutting playbook they can copy directly.
OpenAI's annualized revenue was previously reported at about $70 billion. A new report claims the real figure is roughly $20 billion lower. With AI giants eyeing public markets and investors focusing on sustainable revenue, the gap could reshape expectations about how fast frontier labs can commercialize.
The paper tests the open-source package galahad-kv. It saves the KV state of each ~16K-token block to encrypted local NVMe and reloads it byte-exact instead of recomputing. Running 50M tokens on one H100 with vLLM and Gemma 4, it was faster and cheaper than recompute. Long-context agents and RAG apps get cheaper long-term memory.
The team ran Karpathy's AutoResearch paradigm, where an LLM edits training scripts and keeps only changes that improve a metric, for 12 weeks on production recommendation embeddings. The paper documents failure modes under expensive multi-GPU iterations and proposes a multi-agent framework. Teams automating experimentation with agents get practical lessons.
Subscribe for daily AI updates + free playbook
📘 Subscribe Free