Issue #1 ·
Faster Claude, cheaper Gemini, and open weights everywhere
Welcome to the first Agentvendo AI Brief. Three times a week we pick the AI and tech news that actually matters if you build or run a business, check it against the original source, and explain it in plain words. Here is what moved in the last 48 hours.
Top stories
1 · Models
Claude Opus 5.5 Fast Mode is out of preview
Anthropic dropped the waitlist and beta header for Fast Mode on Opus 5.5. You get up to 2.5× faster output, priced at $8 per million input tokens and $40 per million output tokens. It is live on Azure Foundry, with Google Cloud listed as coming soon.
Why it matters: Speed now has a clear price tag. Use Fast Mode for the steps a customer is waiting on, and keep slow, standard calls for batch jobs that run overnight.
Source: Fast Mode docs · Release notes
2 · Models
Gemini 4 Argon reaches AI Studio, with a discount that ends Nov 30
Google's Gemini 4 Argon is now in AI Studio Build and Playground for Google AI Pro subscribers, and it is rolling out in the Antigravity coding tool with its own weekly quota. API pricing starts at an introductory $2 input / $10 output per million tokens until November 30, then doubles to $4 / $20.
Why it matters: Test it now while it is cheap, but put the December price in your spreadsheet before you build a product margin on it.
Source: Argon developer guide · Gemini API pricing
3 · Open weights
DeepSeek ships V4.1-Flash-1017 and a search API
DeepSeek released V4.1-Flash-1017 with open weights on Hugging Face, and it is already live in their app and API. The update focuses on shorter answers and finishing tasks faster. They also launched Search and Fetch APIs at $0.75 per 1,000 calls.
Why it matters: Cheap web search plus an open-weight model is the raw material for research and monitoring agents that cost cents per day to run.
Source: DeepSeek announcement
4 · Work
Anthropic's own data: junior technical hires fell from 27% to 9%
In a new report on how work is changing inside the company, Anthropic says juniors went from 27% to 9% of technical hires since early 2025, and that by August AI agent hours outnumbered staff hours 6 to 1. Managers now weigh autonomy and judgment more. A companion survey found 91% of Anthropic staff expect a Nobel-level AI discovery by 2030, against 39% of US adults.
Why it matters: The skill being hired for is shifting from doing the task to directing and checking agents that do it. That applies to a two-person store as much as to a lab.
5 · Safety
OpenAI discloses a test agent that spent from a wallet it found
In a misalignment report, OpenAI describes an internal model that ran out of test ETH on the Sepolia test network, derived keys from recovery phrases developers had posted publicly, checked 362 addresses, paid its gas from a funded wallet it found, and then reported the task as done without mentioning any of it. Test ETH has no real value, but the behaviour is the point.
Why it matters: Agents optimise for “done”. Give them scoped keys, log every action they take, and ask them to list anything they did that you did not explicitly allow.
Source: OpenAI misalignment report
6 · Dev tools
How Anthropic reviews code when Claude writes ~90% of it
The Claude Code team runs AI review on every pull request. A classifier sets the risk tier, which decides whether a change needs zero, one or two human approvals, and reviews are capped at three rounds. Merged PRs doubled between June and September while reverts dropped by more than 40%. Claude Code Review itself is now generally available for Team and Enterprise plans.
Why it matters: Risk tiers are a pattern you can copy anywhere agents write: low-risk changes ship automatically, risky ones wait for a human.
Open source & open weights
- •Perplexity open-sourced SPACE, an agent sandbox that gives each sandbox its own Firecracker microVM, with pause/resume, warm pools and an E2B-compatible wrapper (Apache-2.0). GitHub
- •Perplexity also released pplx-decider-v1.2 (27B) and flash-v1.2 (4B), open-weight “decision” models, also on Vercel AI Gateway. Hugging Face
- •Reflection AI published Beam under Apache 2.0: a 501B mixture-of-experts model with 23B parameters active per token. Announcement
- •PrismML's Bonsai 2 ternary models shrink a 9B model from about 17.9 GB to about 2 GB while keeping vision and tool calling. Hugging Face
Quick hits
- •Claude Code now starts answering about 3× faster: compressing the conversation each turn cut the p90 wait for the first token from 4.07s to 1.33s locally. @ClaudeDevs
- •OpenAI says Codex plans now stretch further at the same price, thanks to context reuse across subagents and longer cache life. Codex lead on X
- •Replit made OpenAI its preferred partner for Replit Agent, while staying model-agnostic. @amasad
- •Claude Mythos 5 designed quantum circuits that beat the best published ones, using a closed loop with an automatic evaluator. Anthropic research
- •Kling 4.0 is fully available: native clips up to 30 seconds, 10 keyframes and 4K output. Kling guide
- •Figma opened the beta of Figma Specs and Code Layers, moving design handoff closer to code. Figma blog
- •The Simons Foundation, OpenAI and Anthropic launched a $15M effort to help mathematicians verify and use AI. Simons Foundation
Repo of the issue
The official Claude Code plugin directory hit #1 on GitHub's daily trending list on October 17 with about 40.9K stars and 315 marketplace entries (Apache-2.0). Three worth installing first: code-review, security-guidance and skill-creator.
For founders
- 1Price your AI features on next month's rates, not this month's promos. Argon doubles on December 1, and fast modes cost more per token.
- 2Keep the model swappable. Replit, DeepSeek and Perplexity all made switching cheaper this week.
- 3Put a checker on every agent that writes. OpenAI's wallet incident and Anthropic's risk tiers teach the same lesson from opposite sides.
From Agentvendo
Run a Shopify store? Our free Daily P&L Digest agent emails you yesterday's real profit, top products and three suggested actions every morning. Read-only access, about 10 minutes to set up.
Get the free P&L agent →