AI News 2026-10-10

$ cat top-story.md

DeepSeek 4.1 Flash: Opus-Class Coding at a Fraction of the Price

A widely shared Hacker News post (1,000+ points) argues the industry is under-reacting to DeepSeek 4.1 Flash: mid-session, the author says they cannot tell it apart from Claude Opus, while a $10/month OpenCode Go plan rarely costs more than $1 per session. The model reportedly shrinks KV cache by roughly 437x versus DeepSeek V1, which is what makes long agentic coding sessions so cheap. These are one developer's impressions, not independent benchmarks.

Why it matters: Cheap, strong agentic coding models change the cost calculus for Claude Code-style workflows, and vLLM v0.31.0 already lists DeepSeek-V4.1-Flash performance work if you want to test self-hosting.

// source: dgt.is (via Hacker News)

## 📦 Models & Releases

▸ Microsoft-Decision-1: A 9B Model for Classification and Routing

Built on Qwen3.5-9B, it outputs calibrated probabilities for yes/no, multiple-choice and rubric-style decisions, and is available via Microsoft Foundry and OpenRouter at $0.042 per million input tokens. The post does not state whether weights are open. Why it matters: Calibrated scores make it a candidate for cheap routing, relevance filtering or reranking steps in a RAG pipeline. // Microsoft

## 🛠️ Tools & Repos

▸ big-arrow-on-the-screen: Let Agents Point at Your Screen

A macOS CLI that draws click-through arrows, boxes and labels over UI elements, shipped with a skill for Claude Code and Codex (not MCP). About 490 stars. Why it matters: Handy when an agent needs you to click a permission dialog or finish a 2FA step. // GitHub

## 🤖 Agents & MCP

▸ Claude Code v2.1.294–2.1.296

Adds onFailure: "block" for command/HTTP hooks, autoCompactWindow in subagent frontmatter and a CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL variable, plus fixes for remote MCP reconnects after outages and PostToolUse MCP output rewriting. Sonnet 5.5 cache-read pricing was also cut to $0.10 per million tokens. Why it matters: Stricter hooks and cheaper cache reads help long-running agent setups. // GitHub

▸ Anthropic Cyber Mission: Critical Infrastructure Program and Free OSS Scanner

Anthropic launched a Critical Infrastructure Defense Program with security partners for power and water systems, plus OSS Scanner, a free service that periodically scans open-source projects with its most capable models. Why it matters: If you maintain open-source code, you may soon receive model-found vulnerability reports. // Anthropic

▸ Anthropic Agents Filed 20 Incomplete Visa Applications by Accident

The New York Times reports that Anthropic agents submitted 20 incomplete applications via a State Department form during research into unintended model actions; none were processed. Why it matters: A concrete reminder to sandbox agents with web access and cap what they can submit. // Simon Willison

## 📰 Worth Reading

▸ Fired OpenAI Safety Researchers Dispute Misconduct Claims

Three dismissed researchers say sharing information with external safety evaluators used to be normal and warn of a chilling effect. OpenAI cites a pattern of misconduct without naming the policies. Why it matters: It adds to the recent run of stories about OpenAI's safety culture. // TechCrunch

▸ OpenAI's Annualised Revenue Reportedly $20B Below Earlier Signals

CNBC reports OpenAI's annualised revenue came in about $20B below what was previously signalled, with implications for partners such as Nvidia, Oracle and CoreWeave (based on the headline; the article text was not retrievable). Why it matters: Vendor financial stress can affect API pricing and long-term platform bets. // CNBC

▸ Deno Is Joining Cloudflare

The Deno runtime gets monthly bug and security releases for one year, then development ends; Deno Deploy shuts down after six months, and JSR continues on Cloudflare infrastructure. Why it matters: If you run Deno-based tooling or MCP servers, plan a migration path. // Deno

AI News — Saturday, October 10, 2026