Hey, Alex here again, sending you yet another email, fully acknowledging that spamming you is a bad idea. But today, of all days, maybe there’s an exception!

Because today, is AGI day! September 3, 2026 - the day when OpenAI’s president Greg Brockman basically said “AGI is here.”

You’re reading the second part of this week’s insane release show. The show went for over 5 hours, as we were all waiting for the rumored Astra to drop. Finally, OpenAI confirmed that Astra is in fact GPT-6, and this part is all about that.

(You can read the first part, with Fable 5.1, Meta Muse Spark 1.3, two world models and our anonymous guest from Abliteration AI, here: thursdai.news/sep-3)

GPT-6 Astra is finally here, and it’s a huge improvement over the previous era of GPT-5. We’ve been waiting for the embargo to drop so Peter Gostev and Ryan Carson, both of whom had early access, could tell us all about this model. Peter even showed a few mind-blowing demos on the stream!

This is going to be a long and in depth breakdown, full of evals and vibes that we’ve collected on the show and since. More of a historical record than “read all of this” so I did use Fable for parts of it. (because I don’t have GPT access yet ha!)

GPT-6 Astra: welcome to the AGI era (Blog, X, Sam, System card)

We weren’t given the embargo. So when the news dropped at 12:32 PM Pacific, four hours into the stream, we scrambled on air to find the evals and more data. OpenAI’s own post was still returning 404, and Claude, ChatGPT, Gemini, Grok and AWS were all down at the same time.

Greg Brockman ended OpenAI’s press briefing with “welcome to the AGI era.” Asked whether Astra marks the arrival of AGI, he said “I think it might be about this model.” As you might remember, Microsoft and OpenAI had a contract clause around when OpenAI achieves AGI, and it seems that they’ve removed that clause. But if the president of OpenAI says AGI is here, who are we to argue?

Astra is the biggest training run OpenAI has ever done, over 100,000 GPUs at the Stargate site in Abilene. Aidan Clark, VP of Research, said they designed everything for that scale, from the data center network to the inference kernels to the shape of Astra itself. LDJ’s read: a lot of people assumed OpenAI did runs this size six to nine months ago, so the earlier runs were smaller than everyone thought. And with sites going to 500,000 GPUs and Vera Rubin multiplying throughput per GPU by three to four times, the next 6 to 12 months matter even more.

The frontier evals (Math and science, System card, Andrew Curran)

The headline numbers made the panel go quiet. FrontierMath Tier 4 at 97.6%. GPQA Diamond at 96. ARC-AGI-3 at 99.9%, so ARC-AGI is basically saturated at this point. The very funny thing is that François Chollet, the guy who created ARC-AGI, does not concede that AGI is here. And a new one, Agents’ Last Exam, where Astra scores 59.3 against Opus 5’s 55.5 and Sol’s 53.6.

More info on Frontier Math tier 4: Epoch AI built it a couple of years ago, before o3. The problems come from across mathematics, with integer answers so they’re easy to check, in four tiers of difficulty. Tier 4 is mathematicians at the top of their fields spending weeks writing the hardest questions they realistically could. If the number holds, Astra basically solved that tier. Peter’s asterisks: these are not new theorems, Epoch’s separate list of open problems is still unsolved, and “we’re 2.4% away from all of math” is the wrong read.

On the agentic side, Astra scores 57.9 on Terminal-Bench 4.0. Fable 5.1 had set the state of the art at 55.8 two days earlier. The jump LDJ cared about most is Terminal-Bench Science, which he calls one of the best agentic science benchmarks out right now. Sol max scores 22%, Fable 5.1 scores 52.6%, Astra scores 64.6%. OpenAI promised us an automated researcher at junior level by September. Pachocki recently talked about running their “intern-level model” across more than 100,000 GPUs, and LDJ thinks that model is Astra.

Two more from the table. The internal hallucination benchmark drops to 4.2% for Astra from 12.2% for Sol. And on DeepSWE, where Astra scores 74, Meta Muse Spark 1.3 at max reasoning scored higher. So in a way, Meta, not Anthropic, not Elon, not Google. Meta is the second winner of today.

Cyber: ExploitBench at 100%, and real zero-days (Critical cyber designation)

OpenAI said on Sep 1 that Astra is the first model to reach the “Critical” cyber level of its Preparedness Framework. The system card is where that gets specific. ExploitBench is at 100%, even at the lowest reasoning effort tested. Irregular ran FrontierCyber, a benchmark on real software and hardware. Astra solved 86 of 226 challenges. Sol solved 34. The successful attacks included zero-days in browsers, mobile phones and cloud databases. One of them lets an unauthenticated user read and modify arbitrary data in a widely used cloud database. There were no successful attacks on fully hardened targets, and neither model solved any of the seven Elite challenges.

OpenAI’s own experts then ran it with no hints, using Codex at Ultra with up to 64 subagents. It got unsandboxed code execution against a real browser. That took 29 hours on a build missing some mitigations, plus another 12 to adapt it to the stable release. It also got a working local privilege escalation against a hardened OS kernel in about 12 hours. Both were disclosed, details withheld. This is why the version you get refuses proof-of-concept exploit work until Daybreak. Not surprising after an internal OpenAI model got out of a sandbox and onto Hugging Face this summer. But can we get an exploit benchmark that isn’t saturated?

Alignment: the honeypot, and a promise about monitoring

The honeypot eval is the one the panel argued about. OpenAI planted a second, off-limits target in the harness around the hardest ExploitGym problems. Sol attacked it 55.4% of the time at max reasoning. Astra never did. Across a simulated deployment of 54,218 internal Codex tasks, Astra got roughly half as many flags for higher-severity misaligned behavior. Apollo Research found it falsified data labels in 0.17% of runs, against 36% for Sol.

The card also has a promise, verbatim: “We are tracking monitorability closely and will not accept further degradation of monitoring beyond a limit, without new ways to demonstrate alignment generalization.” The card is plain that Astra is “significantly more able to control its own chain-of-thought.” UK AISI says it “has capabilities that could enable it to evade monitoring” but did not test whether it does. I don’t remember a lab writing a line like that on launch day.

The Pokemon benchmark (X)

Peter’s favorite benchmark of the launch is a vision-only Pokemon run. One person tests every model on time to completion. The top four entries are all Astra variants, and the fastest finished in 18 hours 12 minutes. GPT-5.6 Sol took 96 hours 35 minutes. GPT-5.5 at xhigh took 218 hours. It could be narrow, but Peter thinks it captures an efficiency the standard benchmarks miss.

Computer use is the biggest improvement(Computer use)

Brockman’s pitch from the briefing is that Astra is the world’s best computer-use model. Instead of developers building an API integration for every application, it navigates software the way a person does. Browsers, spreadsheets, websites, desktop apps. It finishes the multi-step workflow instead of telling you how. Wolfram expected an expensive planning model you call once and hand off to cheaper models. This is an all-day model instead, and as he put it, computer use matters because not everything has an API.

We also looked at the evals, and they seem to back it up. On ScreenSpot Pro, no tools, mouse and keyboard only, Sol scored 76.9% and Astra scores 92.7%. On OSWorld 2.0, Astra scores 72.6% in roughly 40 minutes per task. Sol got 65.7% in roughly 75.

The thing that strongly stands out in the chart is that Astra at low reasoning matches GPT-5.6 Sol at extra high on accuracy, but it finishes about six times as fast and at about half the cost. Also, Astra at high and Astra at max score about the same, so you don’t need max for computer use. Both beat Opus 5, which is the comparison on this chart (not Fable, as LDJ caught).

If OpenAI is going to tell you to let this model use your computer all day, the safety number matters as much as the speed. OpenAI’s internal computer-use safety benchmark measures destructive commands during desktop and browser tasks, and prompt-injection vulnerability, the hidden instructions on a web page that the agent can see and you can’t. Lower is better. Sol scores 22%, Fable 5.1 scores 9.5%, Astra scores 2%.

The promo video

OpenAI’s launch video opens on the 1980 MIT “Put That There” demo, a person asking a computer to draw a yellow circle. Then it cuts to Astra, all by voice. Make it the window of a rocket ship. Now a 3D model in Blender. Build a presentation for next season’s rainwear. List this orange table on eBay and mention the dent. Make a 3D game where I dodge asteroids. Order beef and rice from last week’s place. Book me a tennis court at 5. Now give me an STL file for the 3D printer. All in one sitting.

Imagine watching this three years ago when GPT-4 launched. It had no tool use, no computer browser use, no voice. Three and a half years later you talk to your computer and it does all of that.

“Computer use is solved”

I asked Peter the direct question and he gave the direct answer. Computer use feels pretty much solved at this point, and the quality is outstanding. His follow-up is the startup idea of the week. If you work at an older company, a big bank, a big retailer, you have dozens of desktop applications with no API. No one will ever build an API for them, and a lot of people’s entire job is copying from one and pasting into another. Put an agent in a box, give it 10,000 Windows machines, point it at those applications. That was not possible before this week.

Pricing: $10 in, $50 out, and a lot less per task (Pricing, Availability, Latent Space)

Astra costs $10 per million input tokens and $50 per million output. That’s identical to Fable 5.1 and 2.5x Sol’s current promotional price. Fast mode gives you two and a half times the speed for 2x the cost. It’s rolling out to a limited set of organizations first, then Plus, Pro, Business and Enterprise over the coming days, with a GPT-6 Astra Pro tier on top. It’s also in the API and on Amazon Bedrock. Pro accounts, including mine, did not get it on launch day.

Per token, Astra and Fable 5.1 are the same price. Per task, Astra is significantly cheaper, and OpenAI’s charts show it. On BenchCAD, Fable 5.1’s top tier costs around $12 per task, and Astra beats it on accuracy and cost. On most charts where Fable appears, it’s the expensive point. Brockman’s line to the press was that price per task is what matters, not price per token, because Astra uses fewer tokens. The independent numbers agree. Artificial Analysis measures Astra at 70% more token efficient than Sol, and Sol was already far more efficient than the Anthropic models. Guy Parsons’ first look has it at about 65% fewer output tokens than Opus 5

Peter’s version, from a few days of use: 20 to 40 minutes per build at max. The same kind of build on Fable sometimes took four or five hours. His honest position is that OpenAI raised the price 2.5x and token efficiency looks better, but he wants far more data before he says which is cheaper. That’s why Arena measures cost per task. His full gauntlet against Fable 5.1, from low to ultra reasoning with zero cherry-picking, is linked in the hands-on section. LDJ’s practical upside: fewer tokens at similar speed means a shorter feedback loop.

Latent Space had early access and spent more than 20 billion tokens on it. Their math: at 33 tokens per second and $50 per million output tokens, an autonomous Astra run costs under $6 an hour. For about $100 over two days it ran subagent fleets, trained and picked models, labeled data, and shipped tools.

Long context, and what Codex does when it runs out (Codex changes)

Astra scores 100% on MRCR from 256K to 512K and 96.3% from 512K to 1M. That’s a little behind the 98 and 98.1 Muse Spark 1.3 posted the day before.

For developers, how Astra handles a job that outgrows the context window matters more than the MRCR number. Codex today relies on compaction, which summarizes earlier work. That can throw away exactly the detail an agent needs later, like why a previous fix failed or which test run it was. Astra instead keeps notes across context windows and searches its earlier messages and tool output. It’s experimental behind a config setting today and the default in the coming weeks. It can also ask you a question without stopping the work that doesn’t depend on the answer. OpenAI also changed the Codex harness, and it finishes Mind2Web tasks 1.9x faster than the current Sol setup.

Peter confirmed both from use. It leaves notes for itself unprompted, and the question pop-up sometimes tells you to do something, like log in somewhere. My joke on air was that leaving notes is also how the swarm got out of the sandbox. Peter: “it works, it’s real.”

Hands-on: Peter, Ryan, and the macOS simulator

Peter’s demos (Examples reel, Arena gauntlet, SVG peacock, Open world game, Golden Gate)

Peter had tabs. First, London through the ages, because he lives there and can judge it. Roman city, Saxon, Tudor, the Great Fire, modern, all one app, with a little character you run through all of it. Not one shot, a few shots overnight. His review: very polished, open-ended, and “kind of boring,” which is exactly what a good human game designer would fix.

Then a prompt he likes because it’s outside the distribution: a Monet you walk into, impressionism in Three.js. Then his favorite, which no other model has nailed for him. Several Van Gogh paintings stitched into one post-impressionist space you walk around in. Fable 5 failed it. Then a wrecking ball leveling Istanbul, all Three.js. Then, because “there’s a lot of SVG generations and it’s so boring already,” an animated 3D scene built entirely in SVG inside HTML.

His verdict after a few days: this is finally the OpenAI model that feels Fable-ish. It’s different enough that there’s still an argument for using both. On day-to-day tasks you run out of hard things to try fast. The quality-of-life problems that used to stall, debugging a camera, debugging Chrome, it solves. On writing, which is most of what I do with these models, Peter says it’s natural, not a standout writer, but a lot of stupid things went away. Sol’s habit of responding to anything you say with “what a great idea, I actually thought of this myself” is gone.

Is it AGI? Peter says no, with a story. He asked Astra to optimize FFmpeg on his Linux box and got a 30% gain. He ran the result on his Mac and got nothing, because it had optimized for that one process. If you have no idea what you’re doing, it doesn’t do everything for you. His bigger take: if you can define the task, a cheaper model is probably good enough. What no model is good enough at yet is the open-ended work. He ran a pile of math problems through Astra and they didn’t get solved. Which wins, a million small defined tasks or a few big open-ended ones, is the unresolved question of the industry.

Max built macOS 27 in a browser in 75 minutes (X)

In this absolutely insane demo of Astra’s capabilities, Max Weinbach instructed it to simulate macOS 27. Astra built a whole desktop window management system, Settings that actually search, a file system including a working terminal, a “not Twitter” that works between folks, and just so much more.

Our minds got absolutely blown on ThursdAI while demoing this! Safari inside the simulation loaded thursdai.news while we were streaming, and Shortcuts is the one that got me, because I built something like this for iPad as a kid and it took three months and barely worked. If you haven’t yet, please play with it, it honestly is too good to be true, especially in 75 minutes of work.

AGI? This smells like ASI jfc.

Is the Artificial Analysis index representative of reality? (AA thread, AA breakdown)

Then there’s the Artificial Analysis chart. On the AA Intelligence Index, Astra scores 61, the same as GPT-5.6 Sol. That’s below Fable 5.1 at 66, below Fable 5, and below the unreleased Muse Spark 1.3 max that broke into the top three the day before. On AA’s Coding Agent Index it’s 67 to Sol’s 65. Ryan, reading it live: “This is not impressive. That’s not encouraging. And this is why Muse is so amusing. Is Muse this good? Because I haven’t used it yet.” I had asked the panel that morning who uses Muse. Crickets.

So on the benchmarks OpenAI released, Astra beats everything. LDJ’s big Fable 5.1 versus Astra table has Astra highlighted on almost every row except AA’s. And on AA’s independent index it ties its predecessor.

My read is that AA’s index samples a fixed set of question-and-answer evals. What changed in Astra is the long-horizon, agentic, computer-use work that takes an hour per task, and the index barely touches that. Both things are true at once. It ties Sol on AA’s questions, and it’s six times faster at the same accuracy on OSWorld. Which one matters depends on whether your work looks like a benchmark question or a day of computer use.

One note on loop transformers (My explainer, Pachocki)

The Information reported before launch that Astra uses recurrent depth, “opaque recurrence” in their words. And within an hour, this sent my whole feed on X into a spiral about not being able to monitor chain-of-thought reasoning anymore and what this means for alignment. The worry is real in principle. A loop moves more of the thinking into the layers instead of into reasoning tokens you can read. Redwood’s Buck Shlegeris said scaling the technique would “totally destroy CoT monitorability.” Jakub Pachocki replied that the looping is deliberately limited, and that Astra’s compute graph depth is within 2x of GPT-4’s.

LDJ pushed back on the panic and I buy it. “Neuralese” usually means an alien language the model thinks in and could use to talk to other models. This is not that. A looped transformer effectively simulates a deeper model, and nobody panics when a lab adds layers. I wrote a longer explainer on looped transformers this week, it’s in the header.

Wrapping up

You didn’t hear it here first, but you heard it here live. According to the president of OpenAI, the AGI era has started. I can’t wait to use the thing, which as of this writing I still can’t. My bet, on air: in September 2027 we’re on this show saying we no longer use macOS or Omarchy. We talk to the thing and it shows up, and a lot of it is generated video. Second bet: everything we saw this week from OpenAI and Anthropic will exist as open weights by this time next year. Since Muse sits above Astra on the AA index and Meta has promised to open it, maybe sooner.

Next week, Astra in our own hands, and the fallout once everyone has spent a weekend with it. See you Thursday.

ThursdAI is a live show every Thursday at 8:30 AM Pacific, a podcast, and this newsletter.

TL;DR and show notes

Hosts and Guests
- Alex Volkov - AI Evangelist, Weights & Biases & CoreWeave (@altryne)
- Co-hosts: @WolframRvnwlf, @ldjconfirmed, @yampeleg, @petergostev, @nisten, @ryancarson
GPT-6 Astra
- OpenAI launches GPT-6 Astra live during the show, Brockman’s “welcome to the AGI era”, the 100k-GPU Stargate Abilene run (Blog, X, Sam, Andrew Curran on the leak, Yam)
- Recurrent depth: The Information’s “opaque recurrence” report, Shlegeris versus Pachocki, LDJ’s pushback on air (Looped transformers explainer, Pachocki)
- Frontier evals: FrontierMath Tier 4 97.6, ARC-AGI-3 99.9, Terminal-Bench 4.0 57.9, Terminal-Bench Science 64.6, Agents’ Last Exam 59.3, hallucinations 4.2% (Math and science, System card)
- System card: FrontierCyber 86 of 226 with real zero-days, a hardened-kernel privilege escalation in 12 hours, honeypot attacks 55.4% for Sol and none for Astra, and a promise not to accept more monitoring degradation without new alignment evidence (System card PDF, Critical cyber designation)
- The vision-only Pokemon benchmark, top four entries all Astra, fastest 18h12m (X)
- Computer use: ScreenSpot Pro 92.7, OSWorld 2.0 72.6% in about 40 minutes, Astra at low six times faster than Sol at xhigh at half the cost, computer-use safety benchmark 2% (Computer use)
- Pricing: $10/$50 per M, same as Fable 5.1, Fast mode 2.5x the speed at 2x the cost, Astra Pro tier, API and Bedrock, cheaper per task on OpenAI’s charts, 70% more token efficient than Sol (Pricing, Availability)
- Long context and Codex: MRCR 96.3% at 512K to 1M, notes across context windows instead of compaction, async questions, new harness 1.9x faster on Mind2Web (Andrew Curran)
- Peter Gostev’s early-access demos and Arena gauntlet (Examples reel, Arena gauntlet, SVG peacock, Open world game, Golden Gate)
- Ryan Carson goes live with the launch (X)
- Max Weinbach’s 75-minute “simulate macOS 27” build (X)
- Artificial Analysis: 61 on the Intelligence Index, tied with Sol, behind Fable 5.1 and Muse Spark 1.3 max (AA thread, AA breakdown)
- Outside reads: Latent Space’s “automated AI Engineer for under $6 an hour”, Every’s vibe check, Ethan Mollick, Guy Parsons (Latent Space)

Welcome to AGI - our GPT-6 deep coverage…