Blog
Notes and analysis on AI development.
Mercury 3 Outputs 1,068 Tokens a Second, and $0.35 Is a Launch Discount
Inception shipped Mercury 3, a diffusion LLM, on October 19. Output runs at 1,068 tokens per second, about 8x GPT-6 Luna, and $0.35 per million output tokens is a 50% launch discount off a $0.70 list price.
DeepSeek Now Sells the Search Engine Behind Its Chatbot, $0.75 per 1,000 Calls
DeepSeek opened the search engine that runs its own app to outside developers on October 17. Search and page fetch both cost $0.75 per 1,000 successful calls, and the free path through Claude Code ends October 31.
This Week in AI: Argon Reaches the Gemini App, OpenCode Console Keys Exposed
Google opened Gemini 4 Argon to paying Gemini app subscribers on October 13, and OpenCode disclosed on the 17th that OpenAI and Anthropic keys belonging to 1.4% of Console users may have been exposed.
OpenAI and Anthropic Keys Stored in OpenCode Console Leaked, the Local CLI Is Not Affected
Anomaly Innovations disclosed an OpenCode Console breach on October 17. OpenAI, Anthropic and Google API keys saved in the workspaces of 1.4% of Console users may have been exposed.
The Model Refused and the Tool Ran Anyway: a "Repeat-After-Me" Prompt Worked on 6 Agent Frameworks
Penn State researchers reproduced tool calls that execute after the model has already refused them, across LangGraph, AutoGen, CrewAI and three more. The same paper fuzzed 273 public MCP servers and confirmed 253 reproducible flaws in 118 of them.
Meta Opened 816B Weights, and the API Still Costs $4.25 per Million Output Tokens
Meta released the weights for the 816B-parameter Muse Spark Open on October 15. It shares a base model with Muse Spark 1.3, and any provider reselling API access to it needs a separate agreement with Meta once that business passes $20M in revenue.
Anthropic Dropped git at 30,000 Merges a Day, and Published Scion as Design Only
Anthropic described on October 13 how it moved its monorepo off git onto an in-house version control system called Scion. Daily merges went from about 2,500 in May to 30,000 in September, hundreds of mirrors stopped holding, and the code is not being open sourced.
Gemini 4 Argon Goes GA, and Its 1M Output Arrives 65,536 Tokens at a Time
Google opened Gemini 4 Argon to everyone on October 13. The $2 per million input price runs only until November 30, and the headline 1M output ceiling is a sum of 65,536-token segments stitched together.
OpenAI Watermarks ChatGPT Text in the EU, and Leaves the API Switched Off
OpenAI shipped textGrain, a text watermark, on October 5, 2026. ChatGPT and Codex output in the EU gets it automatically, but API customers worldwide have to switch it on themselves. Swapping 25% of the words drops detection to 17%.
This Week in AI: Gemini Free Loses Model Choice, Haiku 5.5 in Copilot
Google locked manual model switching in the Gemini app behind a paid plan, leaving 3.5 Flash-Lite as the only model its pricing page lists for free accounts, and GitHub shipped Claude Haiku 5.5 to five Copilot plans.
No Human Reviews Anthropic AI Vulnerability Reports, and 363 Projects Have Applied
Anthropic opened OSS Scanner, a free vulnerability scan for open source, on October 8, 2026. Reports go out without human review and never become public advisories, and 363 application PRs piled up in three days.
The Deleted Answer Commit Was Still in .git: 1,795 of MiMo 2,698 Coding Tasks
Vals AI audited all 2,698 coding tasks in Xiaomi MiMo v2.6 open RL environments and found the answer commit surviving as an unreachable git object in 1,795 of them. Xiaomi reports confirmed hacking under 2%.