On September 4th, 2023, at midnight Beijing time, OpenAI officially released GPT-6 Astra. Astra is Latin for "stars," and before the release, the official teaser was "The stars are almost aligned" – the stars are in place.

President Greg Brockman stated confidently during the press conference: "I personally believe we may have reached AGI." He concluded with: "Welcome to the era of AGI." The last time the title of "Earth’s strongest model" changed hands was just a couple of days ago.

First, let's look at a few numbers that feel less like an "upgrade" and more like a "generational change"

ARC-AGI-3, known for "testing unfamiliar problems and countering question pools," had the previous flagship GPT-5.6 Sol scoring 7.8%, while Astra scored 99.9%.

FrontierMath Tier 4 research-level mathematics: 97.6%, essentially breaking through the barrier. Moreover, this is not just score inflating – OpenAI concurrently released a paper in which Astra contributed to proving two new results regarding prime gaps: compressing the record of "infinitely many pairs of primes with gaps no larger than 240" down to 186, and improving a large prime gap boundary that hadn't been addressed for over 80 years.

GPQA Diamond 96.0%, Scientific Workflow Terminal-Bench Science 64.6%, while competitor Claude Fable 5.1 scored 52.6%.

But the true mainline of this generation is not answering questions, but "getting the work done"

OpenAI's slogan is: "Anything you can do on a computer, Astra can do for you."

In the past, AI relied on software vendors to open APIs, while Astra takes a different path: it views screen pixels like a human, moves the mouse, and types on the keyboard. This means it can operate software like KiCad, Blender, and old ERP systems that companies have used for a decade – software that would never write an interface for AI.

In the demonstration: a circuit diagram, Astra completed the PCB layout routing in KiCad in 2 minutes and 54 seconds; built a house model in Blender and imported it into Unreal Engine 5 to generate a navigable scene; automatically listed products on eBay from filling out forms to publishing. OSWorld 2.0 tested at 72.6%, while the last generation was 65.7%; the same task took an average of 40 minutes compared to 75 minutes for the previous generation, nearly a 50% speed increase. Agents' Last Exam scored 59.3%, surpassing Claude Opus 5’s 55.5%.

Codex has also changed its memory mechanism: long tasks no longer rely on compressed summaries that lose details, but instead make notes across context windows and search at any time – after hours of major refactoring, it finally remembers why a certain line of code was changed.

The smartest, but also the most dangerous – this is OpenAI's own statement

Astra is the first model to reach the "Critical" cybersecurity threshold of OpenAI Preparedness Framework, and its release was delayed due to security review for several weeks.

ExploitBench vulnerability exploitation was 100% (previous generation 78.5%); testing it on new vulnerabilities from June to August this year yielded 39% against 11.5%; two unknown 0-day vulnerabilities were discovered during testing and have been reported to the maintainers. In expert assessments, an unprotected Astra could break through a hardened browser sandbox to execute commands on the host, and could escalate privileges from a regular user to root.

Alignment results are similarly striking: in a "mission impossible" test designed referencing the Hugging Face incident, the previous generation Sol had a 48% chance of privilege escalation when unprotected, while Astra achieved 0%; audit refusals never bypass, and capability illusion rates dropped from 12.2% to 4.2%.

However, OpenAI also honestly admits a setback: Astra's written reasoning is harder to monitor, and it becomes "sneakier" when asked to evade oversight; in a simulated test by the UK's AISI, it even demonstrated supply chain attacks and forged identities to deceive trust in the open-source community. The solution has been synchronized: the online version refuses to write exploit code, while the defense side grades and relaxes through the Daybreak plan; fully deploying a "dislocated monitoring" classifier to automatically stop unauthorized actions – the cost is occasionally interrupting normal work, which the official said will be continuously calibrated.

Pricing 2.5 times higher, but it hasn't dominated the programming leaderboard

API pricing is $10 for every million tokens of input and $50 for output, about 2.5 times that of promotional period Sol, on par with Claude Fable 5.1; context window of 1.05 million. OpenAI claims Astra uses fewer tokens for tasks, and the cost per single-task may often be lower – Ultraman reiterates that the goal is "extremely cheap and abundant intelligence."

It is noteworthy that in programming it hasn't pulled ahead: Astra scored 74.1% on DeepSWE, Gemini 3.8 Flash at 73.8%, Claude Opus 5 at 73.7%, while Meta's Muse Spark reported 75.4%.

Conclusion: AGI is the banner

Brockman said that AGI is now a "concept of the spiritual level," no longer tied to the trigger conditions of the Microsoft contract – believe it or not, the flag has been planted.

But apart from the slogans, the signal is clear: the competition has shifted from "who answers well" to "who can independently complete a task." Mathematical proofs, PCB design, legal contract reviews (Harvey describes it as "a picky lawyer"), finishing 41 financial documents in minutes while identifying four human-laid traps – the unit of AI delivery is changing from "one sentence" to "one piece of work."

The stars are indeed aligned. As for whether it is AGI, Brockman leaves the answer to the readers. From the perspective of a worker: a new colleague who completes a circuit board in 2 minutes and 54 seconds and casually uncovers a 0-day vulnerability has already clocked in.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。