Catch-up edition: OpenAI DevDay brings GPT-6.1 Sol at a fifth of Astra's price, plus Gemini 4.
​​​​​​​​​​

$ ai-news --date 2026-10-02 | mail

AI News — October 2, 2026

// catch-up edition: this issue covers the past few days since Sep 29, curated for your stack

$ cat top-story.md

OpenAI DevDay: GPT-6.1 Sol at a Fifth of Astra's Price, Dots Agents, Ultrafast Inference

At DevDay on Sep 29, OpenAI launched GPT-6.1 Sol, delivering "near-Astra level intelligence at a fifth of the price" ($2 input / $10 output per million tokens versus $10 / $50). It also introduced an Ultrafast mode (300 tokens/s, 8x faster, at 6x the price), a Decisions API that restricts outputs to a predefined set of options, and Dots, customizable personal agents that work across Slack and Teams.

Why it matters: The price war keeps tightening. The Decisions API mirrors the new class of "decision models" (see Cloudflare's Clef below), and Codex Security Cloud ships open-source CLI tools for automated vulnerability scanning and patching.

// source: Simon Willison's DevDay live blog

## 📦 Models & Releases

▸ Google Announces Gemini 4 "Argon" — But Only to Select Cyber Partners

Google unveiled its Gemini 4 flagship on Sep 30 and says it is comparable to GPT-6 Astra and Claude Opus on key coding and cyber benchmarks. Access is limited to select cybersecurity partners for now, with no public timeline or pricing. Why it matters: The long-awaited Gemini 4 is real but not usable yet; expect another capability and price reshuffle once it opens up. // Reuters via Investing.com

▸ Cloudflare Releases Clef: Open-Weight Decision Models Built on Qwen 27B and 9B

Clef and Clef-flash are Apache-2.0 classification models for agent routing and decisions, with median latencies of about 209 ms and 39 ms and a vision encoder. Cloudflare also announces an RL fine-tuning service. Why it matters: A cheap, deterministic, self-hostable classifier is a natural front stage for a RAG pipeline (query routing, intent checks). // Cloudflare Blog (485 HN points)

## 🤖 Agents & MCP

▸ Pi 1.0: A Hardened, Minimal, Extensible Agent Harness

Pi reaches 1.0 with native MCP support, extensions, deferred tool loading and UI improvements; Earendil also ships Pi Durable, an experimental package for long-running agentic apps. Why it matters: A small, dependency-ready harness with built-in MCP is an attractive base for your own agent scripts, and it topped Hacker News with 1,000+ points. // Earendil / Hacker News

## 🎬 Worth Watching

▸ Fireship — "The one OpenAI announcement that can actually make you money..."

A quick take on the DevDay announcement with the most direct revenue angle for developers. Why it matters: A fast, opinionated digest of what DevDay means for builders. // Fireship, Oct 1

▸ Fireship — "Did a 50 year old military secret just solve agent prompt injection?"

Looks at a decades-old security concept applied to prompt injection in agents. Why it matters: Prompt injection is the main risk for agents with tool access, so any structural defense is worth knowing. // Fireship, Sep 30

▸ Andrej Karpathy — "Let's build GPT: from scratch, in code, spelled out."

A new upload from Karpathy on Sep 30 in his from-scratch GPT series. Why it matters: Solid fundamentals for understanding what vLLM and your local models actually do. // Andrej Karpathy

▸ AI Explained — "OpenAI Security: Controlling Models is Now 'Hell'"

Covers OpenAI's security warnings on model control, Gemini 4 Argon and recent safety research. Why it matters: Good context on the agent-safety debate dominating this week's coverage. // AI Explained, Oct 1

## 📰 Worth Reading

▸ OpenAI and Synopsys Announce GPT-Synopsys for Chip Design

OpenAI and Synopsys announce a frontier-model-based assistant for chip design workflows. Why it matters: It shows frontier labs moving into vertical, domain-specific products. // Synopsys (176 HN points)

$ echo "automatically researched on 2026-10-02"
> ai-news · news.reichenberg.ruhr

[ unsubscribe ] [ view in browser ]