OpenAI's GPT-6 Astra lands at 99.9%1 on ARC-AGI-3 and 97.6%2 on FrontierMath Tier 4 v2, and Greg Brockman says it may already clear the company's own AGI bar. The bill is $103 in, $504 out.
OpenAI's published benchmarks put Astra at 99.9%1 on ARC-AGI-3 and 97.6%2 on FrontierMath Tier 4 v2, ahead of GPT-5.6 Sol and Anthropic's Fable series. President Greg Brockman says the model may already satisfy OpenAI's own definition of artificial general intelligence, and none of it is externally verified.
Supervised autonomous rides began in London on September 3, the first autonomous ride-hailing service in the UK and a step ahead of Waymo in the city. Only 15 vehicles are licensed so far, each with a safety operator in the front seat.
The Trump administration imposed tariffs of up to 100 percent on imported drones, with the steepest rates on heavier airframes and thermal imaging payloads. It arrives the same week Ukraine sent 548 drones toward the Moscow region and Iran struck a US base in Kuwait.
A weekly operator deep dive, the full Dojo Pro prompt packs, and the complete prompt database. Founding members lock the launch price forever.
Join the founding list →OpenAI released GPT-6 Astra on September 3, positioning it as its flagship for computer use, coding, scientific research, and cybersecurity work. Rollout began with enterprise customers in the Daybreak cybersecurity program, with ChatGPT Plus, Pro, Business, Enterprise, API, AWS Bedrock, and Azure access following...
Continue readingIran launched missiles and drones overnight into Thursday at a US base in Kuwait, the first Iranian strike on Kuwaiti territory in the current conflict and a sharp widening of the Gulf front after last week's claimed attacks on Jordan and UAE bases.
Drones run through three of today's four world stories: 548 flew toward Moscow in 28 hours, Iran sent them at Kuwait, and Washington answered a Chinese-dominated market with tariffs up to 100 percent, so the cheapest airframe of the war is now also a trade weapon.
Ukraine's 548-drone operation against the Moscow region, among the largest reported of the war, and Iran's missile and drone barrage on a US base in Kuwait show how mass-produced airframes have become the defining weapon on two fronts at once. The same Thursday, the Trump administration imposed tariffs of up to 100 percent on imported drones, steepest for heavier airframes and thermal imaging models. The market being targeted is dominated by Chinese manufacturers, so the trade measure and the battlefield trend are two faces of one supply chain.
The WMO warned Thursday that a 'supersized' El Niño could keep intensifying into 2027 and may become the strongest ever recorded, raising odds of extreme heat, drought, and flooding across multiple regions. The agency flagged knock-on risks to hydropower output and grid stress in data center regions. For an industry shipping compute-heavy flagships like Astra the same day, climate-driven power constraints are now a direct operating risk rather than a background concern.
Three AI Rivals Crashed Together
Read full analysis
Tool of the day
7.516/ 10
The largest release in OpenClaw's history is worth installing, provided you delete the old one first and hand back permissions one skill at a time.
Version 2026.8.1, the release the project itself calls OpenClaw 2.0, rewrites installation, the web app, memory, skills, model setup, automation, browser and app control, plugins and the security model in one go, and upgrading in place is where people are getting burned: a GitHub discussion already reports exec and file access failing to restore after the update. Treat it as a fresh install, read the security notes before anything else, and rebuild your skills with a hard look at what each one can reach. Do that and you get the broadest open agent stack on the market, 389k GitHub stars deep, for the price of your own hardware and model tokens; skip it and you are running an agent with elevated permissions that reads untrusted web pages and messages, which the Reddit and Register coverage is right to worry about.
This is the largest update in OpenClaw's history, simplifying installation and rebuilding the browser app.Paraphrased · Hacker News
Easy-Peasy.AI has launched Marky Agent, an AI agent that runs on its own machine and completes tasks end to end rather than drafting text for you to finish. The pitch is delegated web and app work: give it a multi-step job like collecting data across sites or filling out a workflow, then review the result. Start with a task you already know the correct output for, so you can check its work before trusting it with anything unsupervised.
Try itAdobe's new Slack app exposes more than 70 tools through MCP, including Firefly, Photoshop, Premiere, Acrobat, InDesign, Illustrator, Stock, and Lightroom, so a teammate can request a resize, a background swap, or a stock pull without leaving the thread. The practical win is for marketing and support teams who currently ping a designer for small changes. Pilot it on one channel with low-stakes assets, and set clear rules on which edits are fine to make without a designer's sign-off.
Try itOpenAI's Codex CLI release 0.152.0 ships with the planning tool disabled by default, which speeds up small edits but removes the step where the agent lays out its approach before touching code. For refactors, migrations, or anything spanning several files, re-enable planning so you can catch a bad strategy before it becomes a bad diff. For one-line fixes and test scaffolding, the new default is the faster path.
Try itOpenAI's latest release notes add conversation branching on the web app and support for website-provided tools inside the built-in browser, so an agentic browsing session can call a site's own functions rather than scraping around them. Branching is the more immediately useful change: when a long research or coding thread reaches a decision point, fork it, explore the alternative, and keep the original context clean. One caution from the OpenAI community forum: a September 1 Codex desktop update was reported to wipe local projects, so back up before you accept the next desktop build.
Try itRun your own eval before you believe the card. Point gpt-6-astra at fifty real tasks from your product logs, not a public benchmark, and record cost per completed task alongside accuracy. OpenAI claims Astra cuts cost per completed task by roughly 57% versus Sol despite higher token prices, so verify that on your workload before you migrate anything.
Build a per-task budget ceiling into the router. At $10 in and $50 out per million tokens, and double that in fast mode, a single runaway agent loop is a real invoice. Cap output tokens, log spend per request, and keep GPT-5.6 Sol or Fable 5.1 as the fallback tier for anything that does not need frontier reasoning.
Re-audit permissions after the OpenClaw 2.0 upgrade. The update touches nearly every layer of the agent, so reinstall from scratch rather than patching over your old config, then re-grant filesystem and network scopes one at a time. While you are in there, wire Adobe for Slack via MCP if asset edits are still stuck in a file-request loop.
Astra's numbers came from OpenAI, the AGI framing came from OpenAI's president, and the pricing came with a faster tier that doubles the rate. That does not make the model weak, it makes the claim untested, and the only test that matters is yours. Meanwhile the outage that took ChatGPT, Claude and Grok offline together is the real lesson of the day: single-provider dependence is now a product risk, not an ops footnote. Build the eval harness, build the fallback, then decide what deserves $504 per million output tokens.
Method names the computation path behind the row, from a closed vocabulary: as published, yfinance close, 03B search corroboration, 04R dossier scoring, not recorded.
A figure is money, a percentage, a spelled magnitude, a multiple, a rate carrying its unit, a score over its denominator, a thousands separated number or an index close. Bare integers and years are not counted, so a phrase like "over 100 integrations" carries no mark. A chart's own scale line is the axis of the level above it and is counted once, with that level.
161 editions so far, one a day. Unsubscribe anytime.