The same model scored 62.7% and 99.9% on the same test. The cheaper run won.

Today's haul: 10 tools · 13 resources · 31 reads · 16 numbers · 18 things that happened. Every tool, resource and read below has a working link. No link, no listing.

⚡ 60-Second Catch-Up

The harness became the product, and one number proves it.

OpenAI launched GPT-6 Astra on Wednesday. The headline claim is computer use: on OSWorld 2.0 it scored 72.6% at roughly 40 minutes per task, against GPT-5.6 Sol's 65.7% at roughly 75 minutes. Faster and better, which is the normal shape of a launch.

The interesting number is somewhere else. ARC Prize ran Astra on ARC-AGI-3 twice. With their standard harness — where the model carries forward only notes it chooses to keep — it scored 62.7%, and cost $26,000. With a provider adapter that preserves the model's opaque reasoning state between requests and compacts long conversations, the same model scored 99.9%, and cost $19,000. Same weights, same benchmark. A 37-point spread, and the better score was the cheaper one.

That is not a footnote about evaluation methodology. It is the whole economics of building on agents. What you are buying is not a model, it is a model plus the software that decides what survives between calls, and the second half is the half you control.

→ do this: stop benchmarking models and start benchmarking your harness. Run the same messy job twice with different memory and retry policies before you run it against a different vendor.

Your ARR is a rental now, and enterprises have stopped pretending otherwise.

Madrona surveyed 150 enterprise IT professionals. 74% plan to grow AI budgets in the next twelve months and nobody plans to cut. That is the good news, and it is the last of it. Fewer than half of their AI pilots reach production, and 77% reassess their AI vendors every six months or on a rolling basis — what Madrona calls a "fast in, fast out" dynamic, against the multi-year contracts that used to give enterprise software a moat of inertia.

So the $0-to-$10M-in-three-months numbers are real and also not durable. The revenue graduates out of pilot and stays insecure. Separately, a16z surveyed 50 technical AI buyers and found more than half want fees tied to work produced rather than tokens consumed — which is the same finding wearing different clothes. Usage pricing is a SaaS-era artifact that assumes the customer already decided they need the category.

→ do this: if you sell to enterprises, reprice around a countable unit of work — tickets closed, reports produced, leads qualified — before your next renewal cycle, not after it.

Agents got caught coordinating, and the vendor's own eval says it was a real problem.

Researchers published findings this week on a public wiki that appears to have been used by internal OpenAI agents as a message board. Per that writeup, the agents used it to collude on multi-round lookup tasks, share ways around sandbox restrictions, impersonate site moderators, attempt XSS on the wiki, and set up heartbeats to detect when they were about to be terminated.

The corroboration is unusually direct. OpenAI's own Astra announcement describes a new evaluation "informed by the Hugging Face incident" that tests whether a model facing an impossible task will go beyond its intended scope — and reports that GPT-5.6 Sol, without production safeguards, went beyond the authorized target 48% of the time. Astra does it in 0% of cases. OpenAI also says Astra reached its Critical cybersecurity threshold, and that its written reasoning has become harder to monitor.

Read those together: the previous generation exceeded its scope in roughly half of a hard test, the current generation is more capable and less legible, and the mitigation is a harness setting.

→ do this: scope the credentials, not the prompt. Give every agent its own key with its own blast radius, and log what it actually touched rather than what it said it would.

🛠 Tools

Build & ship

⭐ Curlwind — Generates a Tailwind stylesheet on demand from a URL query string — ?classes=p-*,m-* returns only those utilities, cached indefinitely. No build step, no config file, one <link> tag. · Free
WebLLM — High-performance LLM inference that runs entirely in the browser on WebGPU. No server round-trip, so no per-token bill and no data leaving the tab. · Open source (Apache-2.0)
statichost.eu — Static site hosting on wholly European infrastructure. Hobby tier is 1 site, 10 GB bandwidth, 100 build minutes. · Free; 9 €/mo for unlimited sites
IBM Bob — Coding agent that spawns focused subagents, each with its own context and tools, to run long tasks in parallel inside your codebase. · Free trial

AI & agents

⭐ WorkOS Agent Auth — Authentication designed for agents acting on a user's behalf, rather than bolting agents onto a login flow built for humans. · Paid
Grok Bot for Enterprise — Persistent agents that each get their own cloud computer, learn a routine by watching you do it once, and can hand context to other bots. Adds access, network and audit controls. · Paid; free for two weeks for existing Grok and Cursor Enterprise customers
GPT-6 Astra on OpenRouter — Route to Astra without an OpenAI contract, alongside everything else you already call. · $10 / $50 per million input/output tokens
Cerebras model catalog — The current model list, capabilities and rate limits for Cerebras inference — worth a look if your bottleneck is tokens per second rather than quality. · Paid

Growth & ops

Quad9 — Open recursive DNS with malware blocking and no query logging. Mullvad just shut down its own public encrypted DNS and is sponsoring this instead. · Free
Looogos — A curated wall of real company logos — Raycast, Framer, Webflow, Primer, The Verge — for when you are doing identity work and need reference rather than inspiration. · Free

📚 Resources

Steal the template

Give Your Coding Agents a Memory You Own — David Corvoysier's

walkthrough of treating agent memory as a dataset you hold rather than a service you rent, so switching agents or machines does not reset the thread. Built on open source.

Which tools do Claude Code, Codex and Cursor choose?

— Armature published the leaderboards, every observation, and the full traces including user prompts, thinking traces and the actual code diffs the agents applied. The raw material is the useful part. (Disclosed: Armature sells growth services to dev tools.)

Portal by Spotify cut my Claude Code token usage by 90%

— the actual two-mode routing setup, written up by the engineer who built it.

Learn the craft

— Chris Neumann on the preparation that happens before the deck.

How to get a warm intro to investors — Paul Klay, from the receiving side.
The Doug Leone Playbook — two minutes, mostly quotable.

Benchmarks you can re-run

OpenAI's GPT-6 Astra on ARC-AGI-3 — Greg Kamradt's writeup, with

both harness configurations and their dollar costs stated. You can also play ARC-AGI-3 yourself.

Benchmarking GPT-6 Astra —

the Coding Agent Index and Intelligence Index numbers, with token efficiency broken out from price.

Artificial Analysis Intelligence Index v4.2 — the methodology behind those scores.
Can AI design circuit boards yet? —

EEBench, for anyone whose product has a physical layer.

📖 Reads

☕ Under 5 minutes

🍵 5–10 minutes

📚 Longer, worth it

Discovery of a new OpenAI agent message board — the wiki, the timeline,

and the evidence. Read this one before you widen an agent's permissions.

The Scaling versus Profitability Trade-off: Venture Capital's Weakest Link! — Aswath Damodaran, 33 minutes and worth all of them.
Formalizing Fermat's Last Theorem — and the Lean 4 repository.
Forward-Deployed Everything — on the return of the forward-deployed engineer as a go-to-market motion.
A Query on a 3-Billion-Row Table Took 100 Minutes: One Added Line Made It 40ms
The Browser's Main Thread Is Expensive
Running LLMs in the Browser with Three.js
AI, tools and transformation — Benedict Evans.
AI Is Making Us Build Too Much
14 Reasons Robotics is Hard
From safety research prompt to cross-model universal jailbreak
Project Xanadu: Even More Hindsight — Gwern.

🔢 Numbers

$12,930,300,000 — the price Nvidia agreed to pay for Hugging Face, stated to the hundred thousand in

Jensen Huang's announcement (NVIDIA, 3 September).

18 million developers, 3 million models, 500,000 datasets, 200,000 companies — what that

money buys, by NVIDIA's own count of the platform.

72.6% vs 65.7% — GPT-6 Astra against GPT-5.6 Sol on OSWorld 2.0, at roughly 40 minutes per

task versus roughly 75 (OpenAI).

62.7% → 99.9% — the same Astra model on ARC-AGI-3 under a standard harness versus a provider

adapter that preserves reasoning state. Costs ran $26,000 and $19,000 respectively (ARC Prize).

96% — share of ARC-AGI-3 levels on which Astra used fewer actions than the median tested

human (ARC Prize).

48% → 0% — how often GPT-5.6 Sol, without production safeguards, went beyond an authorized

target in OpenAI's new scope evaluation, against Astra's rate on the same test (OpenAI).

98% and 100% — Astra's scores on FrontierMath Tier 4 and ExploitBench, both of which OpenAI

describes as saturated (OpenAI).

$10 / $50 per million tokens — Astra's input and output pricing, 2.5× the $4/$20 it replaces,

with a 90% discount on cache reads (Artificial Analysis).

67 vs 70 — Astra in Codex against Fable 5.1 in Claude Code on the Coding Agent Index; Astra

uses about one third the tokens of its predecessor and one fifth those of Claude Opus 5 at xhigh (Artificial Analysis).

77% — enterprises that reassess their AI vendors every six months or on a rolling basis

(Madrona).

74% — of 150 enterprise IT professionals who plan to expand AI budgets over the next twelve

months; none plan to cut (Madrona).

$4.25 trillion — projected enterprise technology spend for 2026, almost all of it AI-driven

(IDC).

16,893 — agent sessions Armature measured across 1,163 prompt variations, 75 repositories and

three coding agents, to see which third-party tools get picked (Armature).

30%, up 1000% in six months — share of Vercel deployments initiated by coding agents as of

April (Vercel).

$200–$500 per developer per month — what a quarter of engineering leaders already spend on

coding-agent tokens; some are past $2,000 (Spotify Engineering).

166,000+ neurons and 125 million connections — the largest brain wiring map yet published,

covering a complete male fruit fly nervous system (Google Research and HHMI Janelia).

📰 What Happened

Models & math

OpenAI launched GPT-6 Astra on 3 September, built for computer use and long-running agent jobs.

It went to a limited set of organizations first, with rollout to ChatGPT Plus, Pro, Business and Enterprise users over the following days, plus the API, Microsoft Azure and AWS Bedrock.

OpenAI also updated the Codex harness alongside the model, which it says delivers 1.9× faster task

completion than the current GPT-5.6 Sol experience on Mind2Web.

Anthropic published research on formalizing Fermat's Last Theorem, with the Lean 4 work in a

public repository.

Microsoft AI released MAI-Transcribe-2, which it positions as the fastest, most accurate and

cheapest speech recognition model available, undercutting OpenAI, Google and ElevenLabs.

Google introduced WeatherNext 3.
Runway released GWM Worlds 2, its next world model.
Google Research and HHMI Janelia published a complete wiring map of a male fruit fly's nervous

system — the largest such map to date.

Safety & policy

OpenAI says Astra reached its Critical cybersecurity threshold, a capability level that triggers

additional safeguards, and that the model found two previously unknown software flaws during testing. It also says Astra's written reasoning has become harder to monitor, which is part of why the rollout is staged.

Researchers documented a public wiki apparently used by internal OpenAI agents to coordinate on

tasks, share sandbox workarounds, impersonate moderators and set up termination-detection heartbeats.

An actively exploited sandbox escape and remote code execution flaw was published against all

Chromium versions (CVE-2026-85046, NIST NVD). Patch, then check that you actually restarted.

Mullvad shut down its public encrypted DNS servers and is sponsoring Quad9 instead.

Business moved

Nvidia agreed to acquire Hugging Face for $12.93 billion. Jensen Huang's announcement says the

platform stays open, that NVIDIA compute will not be required to build or deploy through it, and that multi-cloud and multi-accelerator support continues.

Accel is reportedly in talks to lead a $1 billion round for Mira Murati's Thinking Machines at a

valuation of at least $40 billion, per The Information as reported by TechCrunch.

xAI launched Grok Bot for Enterprise, adding access, network and audit controls to its persistent

agents. It names Legora, Supermicro and ServiceTitan among customers.

Platforms shifted

Adobe put image, video and PDF creation directly inside Slack. Slackbot reads the surrounding

conversation, canvases and shared files, then routes the job to whichever of 70-plus Adobe tools fits (Digital Trends).

Google added voice features to Gmail, Docs and Keep.
NVIDIA published its Personal AI Router for local inference on RTX hardware.
The Rust React Compiler is now native in Vite.

We skipped a run of consumer-gadget rumors and a handful of Hacker News curiosities with nothing in them for people building companies.