The cost of running AI fell again yesterday. The cost of supervising it did not.
- Cheaper agents. Anthropic cut Fable 5.1 cache reads 75%, taking roughly 45% off heavily agentic workloads.
- Critical, contained. OpenAI says Astra is the first model to cross its Critical cyber threshold, so access will be rationed.
- Agent inventory. CrowdStrike launched Falcon Guardian to find and stop the agents already running on your endpoints.
- Local by default. Perplexity now routes the confidential parts of a task to a model on the user's own Mac.
- Fewer blue links. Google is expanding AI Overviews into full AI Mode answers without the extra click.
- The renewal audit. A working CFO published how he decides to cancel, and he opens the conversation four months early.
- Output, not time. Linear's data shows agent teams open six times the pull requests and saved no time at all.
- Factory floor. GE Appliances used AI to catch errors and forecast demand while adding 600 jobs in Georgia.
- Prompting the new model. Anthropic published fourteen habits for running long agent jobs on Fable 5.1.
- Packaged judgment. Every open-sourced the plugin one writer spent six months encoding her method into.
The pattern to carry into your week: token prices are falling and oversight costs are rising, and only one of those shows up on an invoice.
Developments
Product & Technology
September 1: Anthropic cut the price of the part of agent work you pay for most
Anthropic released Claude Fable 5.1 and Mythos 5.1 on September 1, cutting cache reads 75% to $0.25 per million tokens while leaving headline input and output pricing unchanged at $10 and $50. The company estimates typical workloads now cost about 25% less, and heavily agentic ones roughly 45% less. Fable 5.1 scored 52.6% on Terminal-Bench-Science against 24.7% for Fable 5, and safety interventions fell about 60% on cybersecurity work and 85% on biology questions. Mythos 5.1 is the same model with fewer guardrails, restricted to vetted researchers.
So what
This is a repricing of long-running agent work rather than a price cut. Agents pay over and over to reread the same code, documents and conversation history, so a 75% discount on that reread is where the savings actually land. If you shelved an agentic workflow on unit economics, rerun the model against the new cache price before your next board deck. Then check whether your own AI product charges for the expensive part of the job or the cheap one.
Hype check
Artificial Analysis found that a maximum-effort task costs about 20% more on 5.1, because the model writes roughly 1.7 times as much text. The 38-hour unattended experiment and the 82% Browserbase result both came from Anthropic's launch partners, not from independent testing.
Confidence: Established fact. Pricing and benchmark figures are published by Anthropic and corroborated in independent coverage.
Go deeper Anthropic · The Neuron
Technology
September 1: OpenAI says its next model is the first that can find and exploit zero-days on its own
OpenAI published Path to Astra on September 1, declaring Astra the first model to cross the Critical cybersecurity threshold in its Preparedness Framework. In expert-led assessments the model found previously unknown vulnerabilities in a hardened browser and operating system and chained them into working exploits, and it scored 100% on ExploitBench. OpenAI restarted a training run it had frozen after the Hugging Face breach. Release is promised soon, with access to the cyber capabilities restricted.
So what
A frontier lab has now told the market that its own model is too capable at offense to ship openly, and that precedent will arrive in your security questionnaires within a quarter. Expect enterprise buyers to start asking which model tier your product runs on and what your vendor discloses about capability thresholds. Ask your security lead this week whether your incident response plan still assumes an attacker working at human speed. That assumption is the one being retired.
Confidence: Established fact. OpenAI published the classification and multiple outlets reported it independently.
Go deeper OpenAI · The Rundown AI
Technology
September 1: CrowdStrike launched a product to find the AI agents already running inside your company
At its Fal.con conference on September 1, CrowdStrike introduced Falcon Guardian, which discovers known and previously unknown AI agents across Windows, macOS and Linux endpoints, records what they do, enforces which agents may run where, and blocks unauthorized activity at runtime. The company is calling the category AI detection and response. It also launched cybersecurity models built on Nvidia's open Nemotron models and a harness called SafeMind, claiming 52% more legitimate threats detected than general-purpose models.
So what
The premise of this product is that you do not know which agents are running inside your company, and for most growth companies that premise is correct. Before you evaluate any tool, spend an hour building the list yourself: every agent holding a credential, who deployed it, and what it can reach. If nobody can produce that list, you have a governance gap that no procurement cycle will close for you. Treat the 52% detection claim as vendor marketing until a third party runs the test.
Confidence: Established fact. The launch and its capabilities are documented in CrowdStrike's press materials and independent trade coverage.
Go deeper CrowdStrike · The Deep View
Product & Technology
September 1: Perplexity started running the confidential part of a task on the user's own machine
Perplexity launched Hybrid Compute on September 1, a version of its Computer agent that scans files on the device for personal and financial identifiers and routes those steps to a local model on Apple silicon while the rest of the task continues on frontier models in the cloud. Enterprises can set organization-wide sensitivity policies and observe what left the machine. It is available to Pro, Max and Enterprise subscribers, and Perplexity is open-sourcing the classifier that decides what stays local.
So what
Data residency has been the objection that stalls enterprise AI deals, and the answer is shifting from a contract clause to an architecture decision. If your product touches customer records, legal documents or health data, a hybrid path is now a competitive feature rather than a research project. Ask your product lead what it would take to keep the identifying fields off your vendor's servers entirely. The buyers raising that question will not wait twelve months for your answer.
Confidence: Established fact. Announced by Perplexity and covered independently by VentureBeat and Engadget.
Go deeper VentureBeat · The Deep View
Go-to-Market
Google is expanding AI Overviews into full AI Mode answers without the extra click
Google has begun showing fully expanded AI Overviews for some queries, complete with the "Ask anything" follow-up box that previously appeared only after clicking "Show more." Google confirmed the behavior, saying overviews may dynamically expand for topics where its systems judge it most useful. The change was first documented by search practitioners rather than announced. The practical effect is that the ten organic links move further down the page.
So what
If a meaningful share of your pipeline still begins with an organic Google click, the top of your funnel just got shorter and nobody sent you a notice. Pull the last sixty days of organic sessions and conversions split by query intent, and find out whether your informational queries have already gone quiet. Then decide whether the next quarter of spend belongs in ranking pages or in the product and pricing pages an answer engine can quote directly. Comparison and roundup pages are the weakest place to put that money now.
Confidence: Established fact. Google confirmed the expanded behavior, though it published no formal announcement.
Go deeper Search Engine Roundtable · TLDR Marketing
Perspectives
Operations & Finance
A CFO published the exact way he decides to cancel your software
Writing at OnlyCFO and surfaced by TLDR Founders, a working software CFO itemized the ten reasons he cuts vendors. He names a workflow tool that quoted him roughly double what he pays for Claude, and a $50,000-a-year point solution whose function a larger vendor now bundles at no charge. His own practice is to open renewal conversations four or more months early, and he reckons thirty to sixty minutes spent demoing an alternative buys real negotiating leverage. The backdrop is the Bessemer cloud index up 55% in six months.
So what
This is your renewal call written from the other side of the table, and two of his ten reasons are things you control this quarter. Price your AI features against what the buyer pays a frontier lab directly, because that is now the comparison being made out loud in procurement. Then audit which of your capabilities a platform vendor could plausibly bundle for free within a year, and move the roadmap toward the ones they cannot. Assume every renewal conversation starts four months before you planned to have it.
Confidence: Informed perspective. One CFO's stated playbook rather than survey data, though the pricing pressure it describes is corroborated across sources.
Go deeper OnlyCFO · TLDR Founders
People & Organization
Teams with coding agents open six times the pull requests, and none of it came back as time
Linear's analysis of its own product data shows that teams which connected a coding agent went from 21 weekly pull requests to 65 over two years, while teams without one moved from 8 to 10. The output gain did not convert into time saved. Time spent on existing tasks held roughly steady and AI work arrived as a new layer on top, so total time on product development went up rather than down. Linear also notes that managers already see only about 40% of what a report does in a day.
So what
The bottleneck moved from writing code to reviewing it, and the person absorbing that shift is your engineering manager or tech lead. Before you fund another round of agent seats, decide who reviews the additional output and what you will stop asking that person to do. If your player-coach roles are still expected to ship and manage at the same time, the six times figure is a burnout forecast rather than a productivity one. Report merged and shipped work to your board, not pull requests opened.
Confidence: Established fact. The figures come from Linear's published analysis of its own platform, which reflects its customer base rather than the whole market.
Go deeper Linear · TLDR Founders
Resources
Product & Technology
Anthropic published the habits that make long-running agents cheaper and less annoying
Anthropic's Fable 5.1 prompting guide, highlighted by The Neuron, says most existing Claude prompts still work but long-running agents need new habits. Re-test effort levels from scratch, since medium can approximate Fable 5 at lower cost. Keep conversation history append-only so cached context is not invalidated, batch independent tool calls to avoid one-call-per-turn loops, and state explicitly that the model should finish the whole assignment rather than stop to ask permission. Keep scope tight, and ask for surgical file edits instead of whole-file rewrites.
So what
These are configuration decisions with a direct line to your cloud bill, and most teams will never read the document. Hand it to whoever owns your agent workloads with a two-week brief: re-benchmark effort settings, verify the cache is actually being hit, and measure token spend before and after. The savings from yesterday's new cache price only materialize if your call pattern lets the cache work. That is an engineering task rather than a procurement one.
Confidence: Established fact. The guidance is published by the model vendor in its own developer documentation.
Go deeper Anthropic docs · The Neuron
AI Strategy & TransformationNew Angle
Every open-sourced the plugin one writer spent six months encoding her method into
Every published Compound Writing, a guide and installable plugin in which staff writer Katie Parrott packaged her methods for interviewing, outlining, drafting and reviewing into reusable skills. The design point is accumulation. A recurring editorial objection becomes a guardrail, and a sentence-level preference becomes an instruction the agent reads on the next piece. The plugin is on GitHub alongside a guide that shows the back-and-forth behind one of her essays. This extends the same publisher's earlier work packaging its editor-in-chief's judgment into an internal skill, with the difference that this one is public and installable.
So what
The reusable artifact here is the pattern rather than the writing plugin itself. Pick the one recurring judgment call in your business that currently lives in a single person's head, and have that person spend two weeks writing it down as instructions an agent can follow. You will learn quickly whether the expertise is a method or a reflex, and only one of those scales past its owner. Start with a process that runs weekly, so the corrections have time to compound.
Confidence: Informed perspective. A documented internal practice from one publisher, with the artifact public and inspectable.
Go deeper Every guide · Every
