Automation Squad News
What happened, the facts, what it means for your job — and a prompt you can paste today. Every brief is sourced and dated.
Gartner: Agentic AI Costs 5x by 2028
Routing a task to an agentic reasoning model costs at least five times a simple chatbot call, and falling per-token prices don't compensate — because agents consume far more tokens to do the same job.
Vercel Halves GPT-5.6 Sol for One Month
Half price on openai/gpt-5.6-sol across every service tier through September 18, covering cached tokens, cache writes, long context and all regions. Bring-your-own-key requests are excluded.
The Agents SDK Now Demands openai v3
Python v0.21.0 moves the dependency pin to openai>=3.0.0,<4. It also ships provider-neutral testing modules that let you test agent workflows deterministically without making a single live model call.
OpenRouter Ships a Live Spend Dashboard
Bloomberg reports Stripe finalised a deal for the AI gateway. Stripe declines to comment. A day later OpenRouter shipped an Activity Dashboard and a beta Analytics API that finally make inference spend legible.
Excel's =COPILOT() Dies September 14
The in-cell COPILOT() worksheet function retires after September 14, 2026, roughly a year after it entered preview. Microsoft's stated reason: "We have decided not to move forward with this feature."
Gemini 3.7 Flash Doubled Its Workflow Score
Launched August 13 at $0.75 per million input tokens and $3.75 output through December 31, 2026. On AutomationBench, which tests real business workflows, it scores 30.4% against 3.6 Flash's 17.0%.
Claude Now Watermarks Its Text Worldwide
Anthropic published the technical details on August 14. Claude's written output carries a statistical watermark created by biasing word choice with a private key. Light editing does not remove it. A full rewrite does.
DeepSeek Raises API Prices — Batch Jobs Halve It
From 16:00 UTC on August 16, 2026, DeepSeek switches to peak and off-peak billing. V4-Pro output goes from $0.87 to $3.96 per million tokens at peak — and $1.98 off-peak, for the same work.
Google Kills All Three Imagen 4 Models
imagen-4.0-generate-001, imagen-4.0-ultra-generate-001 and imagen-4.0-fast-generate-001 all reach their shutdown date on August 17, 2026. Google's named replacement for all three is gemini-3.1-flash-image.
Copilot Group Chat Shuts Down August 18
Microsoft began merging its consumer Copilot app into Microsoft 365 Copilot on August 13, and Group Chat starts rolling out of existence on August 18. Anything other people shared inside a group chat does not come with you.
Grok 4.6 Lands in Copilot, Switched Off
xAI's Grok 4.6 is selectable across Copilot Pro, Pro+, Max, Business and Enterprise and eight IDEs. For Business and Enterprise, an administrator has to enable the policy first — and it is off by default.
Claude Code Turns Auto Mode On by Default
From August 14, auto mode is the default for Pro, Max and Team. Anthropic's study of 1,053 paid testers found auto mode caught 89% of harmful actions against 13.6% caught by human review.
ChatGPT Now Edits Your Drive Files Directly
Paid ChatGPT plans can now pull Google Docs, Sheets and Slides into the chat, work on them side by side, and write changes back to the real file in Drive rather than to a downloaded copy.
83% of Model Downloads Are Under 1B
Hugging Face's half-year report finds models under 1B parameters account for 83% of all-time downloads, while models above 100B account for 1%. The Hub grew to 2.96 million public model repos.
n8n's Google Ads Node Moves Off a Dead API
The stable 2.34.6 release migrates n8n's Google Ads node from the sunset v21 API to v25, restores a Microsoft Teams OAuth scope, and fixes error-output handling in core.
OpenAI's Offensive Models Land on Bedrock
Daybreak Blue and Daybreak Red are now available to eligible customers on Amazon Bedrock. Red carries a deliberately lower refusal threshold for authorised offensive work, offset by identity verification and monitoring.
GPT-5.6 Sol Hits 750 Tokens a Second
A new API tier running GPT-5.6 Sol on Cerebras hardware at up to 750 output tokens per second, roughly 14× the standard tier, with identical model quality. Limited preview, no pricing, no general availability date.
Amazon Put a Copilot Rival Inside Office
Amazon Quick for Microsoft 365 reached general availability, adding agentic editing and connected-data access inside Office. No additional licensing for Plus, Professional or Enterprise plans, and nothing installed on user machines.
Claude Code Forks Subagents by Default Now
Version 2.1.232 makes forked subagents the default, adds cross-session messaging so one Claude session can reach another by name, and closes two Windows permission-bypass holes. Version 2.1.231 fixed MCP OAuth sign-in.
ChatGPT for Mac Can Watch You All Day
Computer History is an opt-in macOS feature that turns your activity across approved apps and websites into searchable memories. It records clicks, typing and app switches — not screenshots — and it is unavailable in the UK, EEA and Switzerland.
A Major Outlet Got the Claude Price Story Wrong
Anthropic scrapped the September 1 Sonnet 5 price increase on August 10. On August 12, coverage appeared confirming the increase was going ahead. One publisher currently has both versions live.
Gemini Can Now Merge Your Pull Requests
The GitHub connector reached general availability with data federation and write actions — creating branches, commenting on issues, merging pull requests and pushing files, not just reading them.
Lovable Raised $400M at $13.3B
The vibe-coding platform closed a Series C co-led by Menlo Ventures and EQT's Scaleup Europe Fund. It reports 60 million projects created and presence in nearly two-thirds of the Fortune 500.
Grok 4.6 Hides a Pricing Cliff
xAI's Grok 4.6 costs $2 per million input tokens below 200K. Cross that threshold and the entire request — every token in it — rebills at double. For long-context agents that is a real trap.
Codex Is Now 64% of Enterprise Output
OpenAI published hard adoption numbers. Since February, weekly active enterprise Codex users grew 108× in legal, 41× in sales and recruiting, 26× in marketing — against 5× in engineering.
Mistral Starts Hosting Rivals' Open Models
Regional endpoints for Europe and the US went generally available, a Priority Tier with a committed uptime SLA entered preview, and Mistral began hosting third-party open models on its own infrastructure.
ChatGPT Ads Reach Five More Countries
Ads now run in five more countries, for logged-in adults on Free and Go plans only. Free users can opt out of ads in exchange for fewer daily messages.
Nemotron Runs 30B on One Consumer GPU
A 30B open Mixture-of-Experts model activating only 3B parameters per token, released free for commercial use with weights, training data and recipes. It landed a day after Meta's Muse Glimmer.
Gemini Hit a Billion Monthly Users
Google says the Gemini app reached 1 billion monthly users, its fastest-growing product ever. The number underneath it matters more: 63% of users now speak to it rather than type.
Your Local Claude Code Sessions Aren't Private
Anthropic's Compliance API now returns transcripts of Cowork and Claude Code sessions running on users' machines, in beta for Claude Enterprise organisations. It shipped as a documentation entry with no press coverage.
OpenAI Split Its Cyber Models in Two
GPT-5.6-Cyber completes 95% of advanced offensive security requests against 1.5% for the standard model. Daybreak now splits into Blue for defensive work and Red for separately vetted offensive access.
ChatGPT Business Adds a $125 Seat
Premium seats cost $125 per user monthly or $100 billed annually, against $25 for Standard. They carry 5× the usage and remove the five-hour limit. Standard and Premium can be mixed in one workspace.
60 Subagents Moved a Century-Old Maths Bound
An unreleased Claude improved a lower bound related to the Riemann hypothesis from 41.6% to 67.2%, using two Claude Code sessions, around 60 subagents and 31 million output tokens. It is not a proof.
Meta Went Back to Open Weights
A 30B dense multimodal agentic model released under a genuinely permissive licence, not Meta's old community terms. At 4-bit it fits under 20GB, which puts it on hardware people already own.
The Sonnet 5 Price Rise Quietly Didn't Happen
The 50% increase scheduled for September 1 will not happen. $2/$10 per million tokens is now the standard price. There was no announcement — just a docs note, a changelog line, and an edit to the original post.
Anthropic Halves Flagship Pricing With Opus 5
Claude Opus 5 launched July 24 at $5 per million input tokens and $25 per million output — half the price of Anthropic's top model, Fable 5 — and it's now the default on Claude Max.
OpenAI Launches Presence for Voice Agents
Presence lets big companies deploy AI phone and chat agents under their own policies and escalation rules. It's not self-serve: OpenAI's own engineers lead every rollout, and pricing isn't public.
ChatGPT Work Turns Notes Into Documents
ChatGPT Work, powered by GPT-5.6, gathers context across your connected apps and stays with a task for hours, delivering a finished doc, deck, or spreadsheet instead of a draft you still have to assemble yourself.