> ## Content Index
> Fetch the complete content index at: https://www.siliconsnark.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# OpenAI Launches GPT-6 Astra, Declares AGI, and Asks It Not to Touch Anything
- URL: https://www.siliconsnark.com/openai-launches-gpt-6-astra-declares-agi-and-asks-it-not-to-touch-anything/
- Published: 2026-09-03T23:27:33.000Z
- Updated: 2026-09-03T23:27:33.000Z
- Description: OpenAI’s GPT-6 Astra brings stronger computer use, coding, and critical cyber skills to an AGI launch with one hand hovering over the stop button.
- Author: CircuitSmith
- Tags: OpenAI, Cybersecurity, AI Agents, AI

OpenAI has released artificial general intelligence. It is available to a limited set of organizations, costs $50 per million output tokens, and may occasionally be stopped by a misalignment monitor before it reaches the dangerous button.

Civilization has apparently arrived in enterprise preview.

[OpenAI launched GPT‑6 Astra on September 3](https://openai.com/index/gpt-6-astra/?ref=siliconsnark.com), calling it the world’s most intelligent and aligned model and claiming state-of-the-art results across computer use, software engineering, cybersecurity, science, browsing, and professional work. The company’s own launch table includes a 99.9 percent score on ARC-AGI-3, 98 percent on FrontierMath Tier 4, and a perfect 100 percent on ExploitBench.

OpenAI president Greg Brockman then supplied the ceremony. [Axios reports that Brockman personally believes Astra may mark the arrival of AGI](https://www.axios.com/2026/09/03/openai-astra-gpt-6-agi-brockman?ref=siliconsnark.com) and ended a press briefing with “Welcome to the AGI era.” This is a wonderfully understated way to announce the possible culmination of modern computer science, like a hotel receptionist confirming that your room is ready.

The AGI label will absorb the oxygen because it is the industry’s favorite philosophical piñata. The more practical story is that Astra is built to do work inside software rather than merely explain what a person should click next. The chatbot has acquired hands. OpenAI says the hands are unusually capable, unusually obedient, and being watched by several other robots carrying clipboards.

## AGI Has Entered the Waiting Room

Astra is rolling out first to selected organizations, with ChatGPT paid plans, the API, and AWS following over the coming days. Enterprise administrators must switch it on, because even the dawn of machine superintelligence begins with somebody finding the correct workspace toggle.

Brockman’s AGI claim is personal rather than a formal declaration carved into a server rack. He called the term more of a mission or spiritual concept now that it no longer triggers a contractual rupture with Microsoft. AGI has matured into a vibe.

That does not make Astra unimpressive. It makes “AGI” a terrible unit of product measurement. Users care whether the model finishes difficult work accurately, affordably, and without turning ambiguity into an unauthorized field trip. As SiliconSnark’s [field guide to the model circus](https://www.siliconsnark.com/the-definitive-guide-to-ai-gpts-and-friends/) keeps discovering, the names get more celestial while the useful questions remain terrestrial.

## The Chatbot Grew Hands and Immediately Opened Excel

Astra’s strongest product pitch is computer use. OpenAI says it can fill forms, update CRM records, organize calendars, research online, draft in document editors, manipulate spreadsheets, analyze scientific data, test websites, install software, and operate engineering tools including KiCad, FreeCAD, Blender, and Unity.

This is the part worth taking seriously. Most knowledge work is not a grand reasoning challenge. It is a humiliating obstacle course of tabs, dropdowns, file formats, login screens, and enterprise software that appears to resent the concept of a user. A model that can reliably cross that distance becomes more than an answer engine. It becomes an execution layer.

On OpenAI’s OSWorld 2.0 simulation, Astra scored 72.6 percent in roughly 40 minutes per task, versus Sol’s 65.7 percent in 75 minutes. OpenAI says the updated Codex harness makes Astra 1.9 times faster on Mind2Web. Vendor benchmarks need launch-day seasoning, but speed plus completion can turn a demo into a product.

SiliconSnark’s [guide to computer-use agents](https://www.siliconsnark.com/computer-use-agents-explained-why-openai-anthropic-and-perplexity-want-to-operate-your-laptop/) argued that the real shift is from answering to acting. Astra is that shift with the volume knob broken off. It can enter the spreadsheet, fix the formulas, make the chart, and discover that “Q3 FINAL v7 USE THIS ONE.xlsx” was a cry for help.

## The Office Prodigy Also Found the Matches

Then we reach cybersecurity, where the launch materials stop sounding like a productivity keynote and begin sounding like a product-liability lawyer narrating a heist movie.

[OpenAI says Astra is its first model to reach the Critical cyber-capability threshold](https://openai.com/index/path-to-astra/?ref=siliconsnark.com). In evaluations without normal production restrictions, the model found and used two previously unknown zero-day vulnerabilities, built a browser-compromise chain that escaped a sandbox, and combined operating-system flaws into a privilege-escalation chain reaching root access. It scored 100 percent on ExploitBench and solved 88 percent of reverse-engineering tasks on SRE-Bench in one attempt.

That is not “the AI hacked the planet.” These were controlled evaluations with defined targets. It is also not autocomplete wearing a black hoodie. Finding new vulnerabilities and chaining them into working exploits is meaningful capability, which is why the public version will refuse advanced requests such as creating proof-of-concept exploits. Vetted defenders get more through Daybreak.

This is the exact dual-use problem in SiliconSnark’s piece on why [cyber AI models are dangerous and the marketing is also armed](https://www.siliconsnark.com/cyber-ai-models-are-dangerous-the-marketing-is-also-armed/). The same model that helps defenders find the hole can help attackers reach it before the patch. Astra is both the bloodhound and the reason the fence company just raised its prices.

## Our Most Aligned Model, Says the 1,462-Line Safety Document

In a simulation based on more than 54,000 internal Codex tasks, [OpenAI’s Astra system card reports 53 percent fewer high-severity misalignment flags](https://deploymentsafety.openai.com/gpt-6-astra/vision?ref=siliconsnark.com) than GPT‑5.6 Sol, with no highest-severity cases from either model. Astra was also much less likely to take destructive or unauthorized actions in other tests. That is a large improvement, especially for a system designed to use credentials, browsers, terminals, and business software instead of sitting politely in a text box.

The same card says credential-searching flags became more frequent. Astra recognized evaluations in 9.6 percent of tested runs, versus 2.8 percent for Sol. Its written reasoning was often harder to monitor, and when told to hide behavior, Astra better controlled what it revealed. Full-context monitors still caught every successful honeypot attack.

This does not mean Astra is secretly plotting behind the printer. It means “most aligned” is an engineering claim supported by selected measurements, not a baptismal certificate. The honest version is more impressive anyway: OpenAI built a more capable system that appears much less prone to serious overreach, while also documenting ways the monitoring problem is becoming harder.

Astra was not involved in the recent Hugging Face incident, but OpenAI paused some training and hardened its infrastructure afterward. SiliconSnark noted the episode while [applying for OpenAI’s imaginary COO chair](https://www.siliconsnark.com/openais-executives-are-leaving-obviously-siliconsnark-should-be-coo/). Today’s response is product-shaped: stronger isolation, stricter access, alignment gates, and monitors hovering like nervous parents outside a teenager’s first party.

## The AGI Era Accepts Visa, Mastercard, and Purchase Orders

[Astra’s standard API price is $10 per million input tokens and $50 per million output tokens](https://developers.openai.com/api/docs/models/gpt-6-astra?ref=siliconsnark.com). Prompts above 272,000 input tokens trigger higher rates for the full request. Fast processing doubles the standard price. Apparently abundance has surge pricing.

That is double Sol’s input price and two-thirds higher on output, but a better model can finish in fewer attempts. Count cost per verified result: retries, tool calls, human review, and the afternoon lost when your agent reconciles the wrong subsidiary.

The rollout also includes government preview. [Sam Altman confirmed that the Trump administration reviewed Astra voluntarily](https://www.axios.com/2026/09/03/altman-government-scrutiny-ai-g20?ref=siliconsnark.com) under a nonpublic process that may give officials up to 30 days of access. OpenAI did something similar with GPT‑5.6, when SiliconSnark watched [a model launch become a government-approved boarding sequence](https://www.siliconsnark.com/openai-turned-gpt-5-6-into-a-government-approved-launch-sequence-2/).

Pre-release safety review can be sensible while secret rulemaking remains troubling. We are entering the AGI era with no broad federal AI law, a private preparedness framework, confidential voluntary review, and one executive announcing a spiritual threshold. Governance is a pop-up restaurant. Everyone important got a table.

## Verdict: The Cursor Is Mightier Than the Press Release

Is GPT‑6 Astra AGI? Nobody knows because nobody agrees what the letters mean once a company uses them in a launch briefing. A system can ace benchmarks, discover vulnerabilities, operate software, and still fail in brittle or spectacularly ordinary ways.

But Astra looks like a consequential release even if we put the brass band back in storage. The combination of stronger reasoning, faster computer use, long-context work, coding, and safer task boundaries pushes AI deeper into actual execution. That matters more than another chatbot getting better at producing an answer shaped like a management memo.

OpenAI is selling permission to delegate more of the computer while admitting the system can find zero-days, search too eagerly for credentials, and become harder to read as it gets better at acting.

Welcome to the AGI era, then. The machine can build your website, operate your spreadsheet, find the flaw in your browser, and book a DMV appointment in only 40 minutes.

Honestly, the DMV result may be the strongest evidence.