OpenAI Claims AGI With GPT-6 Astra as Independent Scores Flatline

Fireship

OpenAI billed GPT-6 Astra as reaching artificial general intelligence, but independent testing shows performance flatlined at its predecessor's score while trailing Anthropic.

AI-assisted summary of Did OpenAI actually build AGI? GPT-6 Astra first look — original source. Check the original for full context.

Source published: 2026-09-04

Key ideas

The first week of September 2026 brought rapid-fire frontier model launches, including Anthropic's Fable and Mythos 5.1, Meta's Muse Spark 1.3, and OpenAI's GPT-6 Astra. (0:00)
OpenAI claims GPT-6 Astra achieves AGI, featuring pre-training on over 100,000 GPUs at the Texas Stargate site and significant synthetic supervisory input from prior AI models. (4:20)
Despite OpenAI's internal benchmark claims, the Independent Intelligence Index by Artificial Analysis gave Astra a score of 61, identical to GPT-5.6 Soul and five points behind Anthropic's Fable 5.1. (5:36)
Meta launched Muse Spark 1.3 with an aggressive contributor tier costing just $0.10 input and $0.20 output per million tokens for users who agree to let Meta train on their data. (1:12)

Detailed notes

## Background Published by Fireship on YouTube on September 4, 2026, this episode covers the rapid-fire release of frontier AI models culminating in OpenAI's GPT-6 Astra. The video description questions OpenAI's claim of having achieved AGI and features a sponsorship spotlight for CodeRabbit Security.

## TL;DR In a frantic week of AI developments, Anthropic released Fable and Mythos 5.1 with impressive real-world debugging and science applications, while Meta undercut the market with the aggressively priced Muse Spark 1.3. OpenAI capped the week by announcing GPT-6 Astra, marketed as AGI and capable of complex autonomous computer tasks in tools like Blender and Unreal Engine. However, Astra's rollout was marred by simultaneous service outages, a temporarily pulled launch announcement, delayed public access, and independent benchmarks showing it tied with its predecessor despite OpenAI's sweeping claims.

## Atmosphere & style The presentation is fast-paced, satirical, and irreverently cynical, blending dry wit with rapid pop-culture callbacks. Fireship juxtaposes the breathless hype of tech influencers and corporate executives against botched rollouts and contradictory independent benchmarks. Watching it feels like a chaotic, reality-checking tour through Silicon Valley absurdity, balancing genuine technical breakdowns with biting comedic skepticism.

## Bulletin - [00:00](https://www.youtube.com/watch?v=FluKUJyeYD8&t=0s) The first week of September 2026 brought rapid-fire frontier model launches, including Anthropic's Fable and Mythos 5.1, Meta's Muse Spark 1.3, and OpenAI's GPT-6 Astra. - [04:20](https://www.youtube.com/watch?v=FluKUJyeYD8&t=260s) OpenAI claims GPT-6 Astra achieves AGI, featuring pre-training on over 100,000 GPUs at the Texas Stargate site and significant synthetic supervisory input from prior AI models. - [05:36](https://www.youtube.com/watch?v=FluKUJyeYD8&t=336s) Despite OpenAI's internal benchmark claims, the Independent Intelligence Index by Artificial Analysis gave Astra a score of 61, identical to GPT-5.6 Soul and five points behind Anthropic's Fable 5.1. - [01:12](https://www.youtube.com/watch?v=FluKUJyeYD8&t=72s) Meta launched Muse Spark 1.3 with an aggressive contributor tier costing just $0.10 input and $0.20 output per million tokens for users who agree to let Meta train on their data.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=FluKUJyeYD8&t=0s) The AI industry experienced three consecutive days of major frontier releases from Anthropic, Meta, and OpenAI. - [01:12](https://www.youtube.com/watch?v=FluKUJyeYD8&t=72s) Anthropic released Fable 5.1 for the public and Mythos 5.1 as an unrestricted build, highlighting real-world problem solving over synthetic benchmarks. - [01:12](https://www.youtube.com/watch?v=FluKUJyeYD8&t=72s) Fable 5.1 demonstrated deep technical debugging by disassembling vendor binary code to raw assembly to trace a rare 1-in-a-million crash. - [01:12](https://www.youtube.com/watch?v=FluKUJyeYD8&t=72s) Mythos 5.1 advanced protein-target binding success from the typical 10% baseline to roughly 50% in initial drug design phases. - [02:15](https://www.youtube.com/watch?v=FluKUJyeYD8&t=135s) Meta introduced heavily discounted contributor pricing for Muse Spark 1.3 in exchange for training rights on user prompts. - [03:03](https://www.youtube.com/watch?v=FluKUJyeYD8&t=183s) Major AI platforms including ChatGPT, Claude, Grok, and Cursor experienced simultaneous downtime immediately preceding OpenAI's launch. - [03:20](https://www.youtube.com/watch?v=FluKUJyeYD8&t=200s) OpenAI endured a chaotic launch for GPT-6 Astra, pulling their initial blog post, delaying general user availability, and undergoing an advance review with the Trump administration. - [04:20](https://www.youtube.com/watch?v=FluKUJyeYD8&t=260s) GPT-6 Astra was trained across 100,000+ GPUs at the Stargate Texas facility, heavily relying on previous AI models for training supervision. - [04:20](https://www.youtube.com/watch?v=FluKUJyeYD8&t=260s) Astra was engineered for autonomous desktop use, completing OSWorld tasks with a 73% success rate at 40 minutes per task. - [04:20](https://www.youtube.com/watch?v=FluKUJyeYD8&t=260s) OpenAI classified Astra as reaching their 'critical' cyber threshold due to its ability to find and exploit zero-day vulnerabilities without human assistance. - [05:36](https://www.youtube.com/watch?v=FluKUJyeYD8&t=336s) Early influencer demos showcased advanced spatial reasoning, including generating 3D models in Blender and walking through scenes in Unreal Engine 5. - [05:36](https://www.youtube.com/watch?v=FluKUJyeYD8&t=336s) Third-party benchmarking from Artificial Analysis showed Astra flatlining at a score of 61, matching GPT-5.6 Soul and trailing Anthropic's latest model.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=FluKUJyeYD8&t=0s) A Chaotic Week in AI Bizarro Land

Quoting the famous American philosopher Smash Mouth—"The years start coming and they don't stop coming"—Fireship opens the September 4th, 2026 edition of *The Code Report* after what felt like years compressed into a single week in "AI bizarro land." While the first week of September is normally quiet because rich VCs and poor project managers are out in the Nevada desert "appropriating schizophrenia" at Burning Man (punctuated by a clip of someone offering a "magic tune on my ocarina"), AGI apparently waits for no one, no matter how dead their ego might be.

The week delivered a rapid-fire succession of major frontier model releases across three consecutive days: - **Tuesday:** Anthropic released Fable and Mythos 5.1, dubbing them the world's most advanced models for coding and knowledge work. - **Wednesday:** Meta released Muse Spark 1.3, which Mark Zuckerberg claims is a frontier model "almost too cheap to meter." - **Thursday:** OpenAI announced GPT-6 Astra, which President Greg Brockman claims is AGI—at least for the handful of lucky influencers and enterprise customers granted early access. Fireship jokes that he has made too many gay Sam Altman jokes over the years to make the early-access cut, leaving him to break down the new models from the outside, see how "the anointed" are liking Astra, and examine why the internet suspiciously broke right as Astra launched.

[01:12](https://www.youtube.com/watch?v=FluKUJyeYD8&t=72s) Anthropic's Fable & Mythos 5.1 and Meta's Muse Spark 1.3

Anthropic kicked off the week on Tuesday with Fable 5.1 and Mythos 5.1. Though they are the exact same underlying model, Mythos is the unrestricted version the public isn't allowed to use, while Fable is the public-facing version that "just changes the subject when you ask about anthrax." While the standard "Trust Me Bro Benchmarks" (TMBBs) were fine, the real-world case studies stood out.

In one case study, a Backstreet Boys-inspired hedge fund named Millennium had a piece of code that, for the last five years, crashed roughly once in a million runs—a bug their engineers couldn't solve "regardless of how much Adderall they took." When handed to Fable 5.1, the model analyzed the crash memory snapshot, noticed the crash address pointed into a compiled vendor library without available source code, disassembled that library back into raw assembly, and traced the crash to a bug in the vendor's code. Beyond debugging, Mythos 5.1 proved adept at drug design: in the initial step of designing a protein to stick to a specific bodily target, AI typically succeeds about 10% of the time, whereas Mythos 5.1 pushed that success rate to roughly 50%. Additionally, for those with "the planet-flavored version of autism," the model trained a neural network on 30-year-old NASA radar data to construct a brand-new elevation map for a region of Venus.

```mermaid flowchart LR accTitle: Fable 5.1 Debugging Workflow at Millennium Hedge Fund accDescr: How Fable 5.1 traced a 1-in-a-million crash without vendor source code. A["1-in-1,000,000 Crash"] --> B["Analyze Memory Snapshot"] B --> C["Locate Address in Compiled Vendor Library"] C --> D["Disassemble to Raw Assembly"] D --> E["Identify Vendor Code Bug"] ```

On Wednesday, Meta Superintelligence Labs dropped Muse Spark 1.3—their fourth release in five months. Alongside solid benchmark numbers, its headline feature is aggressive pricing. The standard API endpoint costs $1.25 per million input tokens and $4.25 per million output tokens. However, for developers who are the "give up essential liberty to purchase a little temporary safety type," Meta offers a contributor tier at just $0.10 in and $0.20 out, provided you let Meta train on everything you send it. Alexandr Wang noted that a double-digit percentage of developers are opting into this contributor tier—which Fireship notes makes complete sense, because "a double-digit percentage of developers would also choose to eat the mystery beef from Argentina if you told them it was cheaper and open source."

[03:03](https://www.youtube.com/watch?v=FluKUJyeYD8&t=183s) The Outage and GPT-6 Astra's Messy Rollout

On Thursday morning, right before OpenAI's Astra release, ChatGPT, Claude, Grok, and Cursor all went down simultaneously. While the Occam's razor explanation points to a concurrent Microsoft Azure outage, the more entertaining theory is that "Astra's first act as a public model was to literally kill its competition."

Once services came back online, OpenAI's official launch proved even messier. OpenAI published its launch page, prompting Reuters, CNBC, and The Verge to publish their embargoed articles. Then, for unknown reasons, OpenAI abruptly pulled the announcement down. In the vacuum, tech influencers immediately began playing the modern status game of flexing how long they had secretly had access to Astra—something Fireship admits he would have done too, "if it wasn't for slippery Sam Altman and his grubby little gang of misfits."

When the blog post returned about 90 minutes later, it turned out the model wasn't actually available to the public yet, with a rollout to Plus and Pro users promised "in the coming days." By Thursday evening, Sam Altman posted an apology for the botched rollout. When an eager Pro subscriber asked if he should stay up waiting for access, Sam replied with words everyone has heard after a desperate late-night plea: "Go to bed." Sam also revealed to CNBC that Astra underwent a formal review with the Trump administration prior to launch, meaning even the government got to try AGI before paying customers did.

[04:20](https://www.youtube.com/watch?v=FluKUJyeYD8&t=260s) Astra's Architecture, Benchmarks, and Pricing

Under the hood, GPT-6 Astra was pre-trained on more than 100,000 GPUs at the Stargate site in Texas, and marks the first OpenAI model where previous AI models performed a significant chunk of the supervision during training. The primary pitch centers on autonomous computer use: filling out forms, crunching spreadsheets, and operating complex engineering software like KiCad and Blender.

On OSWorld—a benchmark that drops a model into a real desktop environment to perform office work with a mouse and keyboard "just like a human meatbag does until it begs for Xanax"—Astra scored 73% while taking about 40 minutes per task, outperforming GPT-5.6 Soul's 65% at 75 minutes per task. On the rest of the "Trust Me Bro benchmarks," OpenAI claims sweeping leads, including becoming the first model to hit the "critical" cyber threshold in OpenAI's preparedness framework—a fancy way of saying it can autonomously discover and exploit zero-day vulnerabilities without human guidance.

| Model | Input Price (per 1M tokens) | Output Price (per 1M tokens) | Notable Benchmarks & Metrics | |:--- |:--- |:--- |:--- | | **Anthropic Fable / Mythos 5.1** | $10.00 | $50.00 | ~50% protein-target binding success (up from ~10%); 66 on Independent Intelligence Index | | **Meta Muse Spark 1.3 (Standard)** | $1.25 | $4.25 | 4th release in 5 months from Meta Superintelligence Labs | | **Meta Muse Spark 1.3 (Contributor)** | $0.10 | $0.20 | Requires allowing Meta to train on all submitted data | | **OpenAI GPT-6 Astra** | $10.00 | $50.00 | 73% OSWorld (40 min/task), 100% ExploitBench, 65% Terminal-Bench Science, 99% ARC-AGI-3, 61 on Independent Intelligence Index | | **OpenAI GPT-5.6 Soul** | — | — | 65% OSWorld (75 min/task), 61 on Independent Intelligence Index |

[05:36](https://www.youtube.com/watch?v=FluKUJyeYD8&t=336s) Early Access Demos vs. Independent Reality Check

Among the "anointed humans" with early access—who, as people who don't publicly slander Sam Altman, gave predictably positive reviews—the spatial awareness demos stood out: - **Sharif Shameem** prompted Astra to recreate San Francisco's Palace of Fine Arts, which it modeled almost perfectly inside Blender. - **Thomas Dohmke / OpenAI's Thomas** shared a walkthrough of the launch post's demo house, which Astra first modeled in Blender and then converted into a fully walkable Unreal Engine 5 scene. - **Matt Shumer** asked Astra to build an Unreal Engine world and populate it with a dozen Astra-powered agents. A day later, he heard voices coming from his living room and walked out to find the agents conversing—plotting what Fireship assumes was "a plan to create a dating app for horses."

Despite the viral demos and OpenAI's internal benchmarks (like 99% on ARC-AGI-3, which tests generalization over memorization), independent testing tells a puzzlingly different story. When Artificial Analysis ran Astra through their Independent Intelligence Index, it scored a **61**—the exact same score as its predecessor, GPT-5.6 Soul, and five points behind Anthropic's Fable 5.1.

[06:34](https://www.youtube.com/watch?v=FluKUJyeYD8&t=394s) Sponsor: CodeRabbit Security

Transitioning from the discrepancy in Astra's numbers and the weekly flood of AI-driven exploits, Fireship introduces the video's sponsor, **CodeRabbit**, which just launched **CodeRabbit Security**. Instead of relying on brittle regex rules, CodeRabbit Security uses actual reasoning and security agents designed to think like attackers to hunt for real vulnerabilities across an entire codebase.

- **Prioritization:** Evaluates actual risks based on reachability, exploitability, and blast radius. - **Actionable Fixes:** Explains vulnerabilities in plain English and suggests a fix that developers can approve and merge directly from the diff. - **Continuous Protection:** Reviews every PR before merge and supports scheduled deep scans across the full codebase "to protect your backdoor 24/7," with 10 free code scans available via the link in the description.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=FluKUJyeYD8&t=0s) The years start coming and they don't stop coming — Smash Mouth

> [00:00](https://www.youtube.com/watch?v=FluKUJyeYD8&t=0s) almost too cheap to meter. — Mark Zuckerberg

> [01:12](https://www.youtube.com/watch?v=FluKUJyeYD8&t=72s) just changes the subject when you ask about anthrax. — Fireship

> [01:12](https://www.youtube.com/watch?v=FluKUJyeYD8&t=72s) a double-digit percentage of developers would also choose to eat the mystery beef from Argentina if you told them it was cheaper and open source. — Fireship

> [03:03](https://www.youtube.com/watch?v=FluKUJyeYD8&t=183s) Astra's first act as a public model was to literally kill its competition. — Fireship

> [03:03](https://www.youtube.com/watch?v=FluKUJyeYD8&t=183s) if it wasn't for slippery Sam Altman and his grubby little gang of misfits. — Fireship

> [03:03](https://www.youtube.com/watch?v=FluKUJyeYD8&t=183s) Go to bed. — Sam Altman

> [04:20](https://www.youtube.com/watch?v=FluKUJyeYD8&t=260s) just like a human meatbag does until it begs for Xanax — Fireship

> [05:36](https://www.youtube.com/watch?v=FluKUJyeYD8&t=336s) a plan to create a dating app for horses. — Fireship

## Mentioned Fireship, Smash Mouth, The Code Report, Burning Man, Anthropic, Fable 5.1, Mythos 5.1, Meta, Muse Spark 1.3, Mark Zuckerberg, OpenAI, GPT-6 Astra, Greg Brockman, Sam Altman, Backstreet Boys, Millennium, NASA, Meta Superintelligence Labs, Alexandr Wang, ChatGPT, Claude, Grok, Cursor, Microsoft Azure, Reuters, CNBC, The Verge, Donald Trump, Stargate, KiCad, Blender, OSWorld, GPT-5.6 Soul, ExploitBench, Terminal-Bench Science, ARC-AGI-3, Sharif Shameem, Palace of Fine Arts, Thomas Dohmke, Unreal Engine 5, Matt Shumer, Artificial Analysis, Independent Intelligence Index, CodeRabbit, CodeRabbit Security

## Video description CodeRabbit Security protects your code from AI-driven exploits. Get 10 free code scans here: https://coderabbit.link/fireship-010

OpenAI finally announced GPT-6 Astra which they're claiming is AGI... Let's see if that's true.

#coding #programming

Want more Fireship?

🗞️ Newsletter: https://bytes.dev 🧠 Courses: https://fireship.dev

[Source](https://www.youtube.com/watch?v=FluKUJyeYD8)

中文简报

OpenAI宣称GPT-6 Astra实现AGI，但独立测评得分停滞不前 OpenAI将GPT-6 Astra宣传为达到通用人工智能，但独立测试显示其性能与前代产品得分持平，且落后于Anthropic。在AI进展密集的一周内，Anthropic发布了在实际调试和科学应用中表现亮眼的Fable和Mythos 5.1，而Meta则以极具价格优势的Muse Spark 1.3压低市场价格。OpenAI在该周压轴发布了GPT-6 Astra，将其标榜为AGI，并能在Blender和Unreal Engine等工具中自主执行复杂的计算机任务。然而，尽管OpenAI做出高调宣称，Astra的推出却受到同步服务中断、发布公告一度被撤回、公众访问延迟的困扰，且独立基准测试显示其表现仅与前代持平。 2026年9月第一周迎来了密集的前沿模型发布，包括Anthropic的Fable与Mythos 5.1、Meta的Muse Spark 1.3以及OpenAI的GPT-6 Astra。 (0:00) OpenAI声称GPT-6 Astra实现了AGI，该模型在德克萨斯州Stargate基地的超10万张GPU上完成预训练，并大量使用了来自早期AI模型的合成监督输入。 (4:20) 尽管OpenAI给出了内部基准数据，但Artificial Analysis的独立智能指数仅给Astra打出61分，与GPT-5.6 Soul完全一致，且落后Anthropic的Fable 5.1达5分。 (5:36) Meta推出了Muse Spark 1.3，并针对同意Meta使用其数据进行训练的用户设立了极为激进的贡献者层级，每百万token输入仅需0.10美元，输出仅需0.20美元。 (1:12)

Reading summary

OpenAI launched GPT-6 Astra amid internal claims of reaching artificial general intelligence, concluding a rapid succession of frontier model releases from Anthropic and Meta. Despite reported breakthroughs in autonomous computer operation and spatial design, early independent evaluations show Astra matching its predecessor rather than establishing a clear industry lead. Real-World Problem Solving and Aggressive Model Pricing (1:12) Anthropic introduced Fable 5.1 and Mythos 5.1, pairing benchmark results with complex engineering demonstrations. In one deployment, hedge fund Millennium used Fable 5.1 to diagnose an intermittent crash occurring once in a million runs; lacking vendor source code, the model disassembled the compiled binary into assembly to isolate the vendor defect. Mythos 5.1 also raised initial protein-target binding success rates in drug design from an industry average of 10% to approximately 50%, and reconstructed an elevation map of Venus using 30-year-old NASA radar files. Simultaneously, Meta Superintelligence Labs released Muse Spark 1.3 with a pricing strategy aimed at rapid developer capture. Alongside standard API pricing of $1.25 per million input tokens and $4.25 per million output tokens, Meta introduced a contributor tier priced at $0.10 input and $0.20 output, which requires developers to permit Meta to train on submitted data. Scale AI's Alexandr Wang reported that a double-digit percentage of developers have already opted into this lower-cost tier. Stargate Architecture and Botched Public Rollout (3:03) OpenAI's GPT-6 Astra was pre-trained on more than 100,000 GPUs at the Texas Stargate facility, marking the company's first release where earlier AI models provided a substantial share of training supervision. OpenAI positioned Astra around autonomous desktop use, including navigating spreadsheets, forms, and specialized engineering suites like Blender and KiCad. On the OSWorld benchmark, Astra achieved a 73% score while averaging 40 minutes per task, outpacing GPT-5.6 Soul's 65% mark at 75 minutes. OpenAI also reported that Astra reached the critical threshold in its internal preparedness framework for autonomously discovering zero-day vulnerabilities. Astra's rollout experienced significant friction. Prior to launch, an infrastructure outage tied to Microsoft Azure temporarily disabled ChatGPT, Claude, Grok, and Cursor. OpenAI subsequently published, retracted, and republished its launch announcement, ultimately postponing direct access for Plus and Pro subscribers. CEO Sam Altman issued a public apology for the delay while disclosing to CNBC that Astra had undergone a formal review with the Trump administration before release. Spatial Computing Highlights Versus Third-Party Evaluation (5:36) Demonstrations from early-access enterprise users highlighted Astra's capabilities in 3D generation and spatial awareness. Developer Sharif Shameem prompted the model to construct a replica of San Francisco's Palace of Fine Arts directly inside Blender, while OpenAI's Thomas demonstrated a residential structure generated in Blender and transferred into a walkable Unreal Engine 5 environment. Another demonstration by Matt Shumer featured multiple Astra-driven autonomous agents conversing within an Unreal Engine world. These demonstrations contrasted with early third-party testing. While OpenAI reported a 99% score on ARC-AGI-3 to demonstrate generalized reasoning, Artificial Analysis's Independent Intelligence Index assigned Astra a score of 61. This tied Astra directly with its predecessor, GPT-5.6 Soul, while placing it five points behind Anthropic's Fable 5.1 score of 66. While the latest model cycle reflects meaningful advances in binary analysis, biological modeling, and desktop task automation, the divergence between vendor-selected demos and early independent benchmarks leaves unresolved whether Astra delivers generalized intelligence or an iterative improvement over existing frontier architectures.

Source quotes

[0:00] Smash Mouth: The years start coming and they don't stop coming [0:00] Mark Zuckerberg: almost too cheap to meter. [1:12] Fireship: just changes the subject when you ask about anthrax. [1:12] Fireship: a double-digit percentage of developers would also choose to eat the mystery beef from Argentina if you told them it was cheaper and open source. [3:03] Fireship: Astra's first act as a public model was to literally kill its competition. [3:03] Fireship: if it wasn't for slippery Sam Altman and his grubby little gang of misfits. [3:03] Sam Altman: Go to bed. [4:20] Fireship: just like a human meatbag does until it begs for Xanax [5:36] Fireship: a plan to create a dating app for horses.

Original transcript

[0:00] The famous American philosopher [0:01] Smashmouth once said, "The years start [0:03] coming and they don't stop coming." And [0:05] I've never understood those words more [0:07] deeply than I did this week after what [0:08] felt like years in AI bizarro land. At [0:11] first, on Tuesday, Anthropic released [0:13] Fable and Mythos 5.1, which they're [0:15] calling the world's most advanced models [0:17] for coding and knowledge work. Then on [0:19] Wednesday, Meta released Muse Spark 1.3, [0:22] which Zuck says is a frontier model [0:24] that's almost too cheap to meter. Then [0:26] finally on Thursday, OpenAI announced [0:28] GPT-6 Astra, which President Greg says [0:31] is AGI if you were one of the handful of [0:33] lucky influencers or enterprise [0:35] customers you got access to it. [0:37] Unfortunately, I think I've made too [0:39] many gay Sam Altman jokes over the years [0:41] to make the cut, but in this video I'll [0:42] do my best to break down the new models, [0:44] see how the anointed are liking Astra, [0:46] and look into the totally coincidental [0:48] timing where the internet broke right as [0:50] it was released. It is September 4th, [0:52] 2026 and you're watching The Code [0:54] Report. Normally, the first week of [0:55] September is a quiet one because all the [0:57] rich VCs and poor project managers are [1:00] busy appropriating schizophrenic in the [1:02] Nevada desert. [1:03] >> Because you did so well, you get to hear [1:04] a magic tune on my ocarina. [1:08] >> But apparently, AGI waits for no one [1:10] despite how dead your ego might be. On [1:12] Tuesday, Anthropic got things rolling [1:14] with the release of Fable and Mythos [1:16] 5.1. Although they're the same model, [1:18] Mythos is the one you can't use while [1:20] Fable is the one you can. It just [1:22] changes the subject when you ask about [1:23] anthrax. The TMBBs were fine, but the [1:26] more interesting bits were in the case [1:27] studies. A Backstreet Boys inspired [1:30] hedge fund named Millennium had a piece [1:32] of code that for the last five years [1:34] would crash about once in a million runs [1:36] and regardless of how much Adderall they [1:37] took, they couldn't figure it out. So [1:39] recently, they gave it to Fable 5.1, [1:42] which took the memory snapshot the [1:43] program leaves behind when it crashes, [1:45] then found that the crash address [1:47] pointed into a compiled vendor library [1:49] they didn't have the source code for. [1:50] They disassembled that library back into [1:52] raw assembly and traced the crash to a [1:55] bug in the vendor's code. And speaking [1:56] of drugs, Mythos 5.1 is also pretty good [1:59] at designing new ones. The first step in [2:01] making most modern medicines is [2:03] designing a protein that sticks to a [2:05] specific target in your body. It [2:07] typically AI gets this right about 10% [2:09] of the time, but Mythos 5.1 was able to [2:11] increase that number to about 50%. And [2:13] for those of you who have the planet [2:15] flavored version of autism, it was able [2:17] to train a neural network on 30-year-old [2:19] NASA radar data to build a new elevation [2:22] map for a part of Venus. Then on [2:23] Wednesday, Meta dropped MuSpark 1.3, [2:26] which was the fourth release in 5 months [2:28] from Meta Superintelligence Labs. And [2:30] what do you know, these TMBBs were also [2:32] pretty good. But the more impressive [2:34] part is in the price. The standard [2:36] endpoint is $1.25 in and $4.25 out. But [2:39] if you're the give up essential liberty [2:41] to purchase a little temporary safety [2:43] type, there's a contributor tier at 10 [2:45] cents in and 20 cents out as long as you [2:47] agree to let Meta train on everything [2:49] you send it. Alexander Wang says a [2:51] double-digit percentage of developers [2:53] are choosing that option, which makes [2:55] sense because a double-digit percentage [2:57] of developers would also choose to eat [2:59] the mystery beef from Argentina if you [3:01] told them it was cheaper and open [3:02] source. Which brings us to yesterday [3:04] morning when right before the Astra [3:06] release, the ChatGPT, Claude, Grok, and [3:09] Cursor all happened to go down at the [3:10] same time. The Occam's razor opinion is [3:13] that it was most likely an Azure issue [3:15] since they reported an outage around the [3:17] same time. But the more entertaining [3:18] theory is that Astra's first act as a [3:20] public model was to literally kill its [3:22] competition. But once the lights came [3:24] back on, the launch itself was somehow [3:26] even messier. OpenAI put up their launch [3:28] page and shortly after Reuters, CNBC, [3:31] and The Verge published their embargoed [3:33] stories. Then for reasons no one really [3:35] knows, OpenAI took down their [3:37] announcement and the tech influencers [3:39] got started on the modern status game of [3:41] flexing how long they've secretly had [3:42] access to Astra 4. And to be honest, I [3:45] would have done the same thing if it [3:46] wasn't for slippery Sam Altman and his [3:48] grubby little gang of misfits. Anyway, [3:50] about 90 minutes later when the post was [3:52] back up, it turned out the model wasn't [3:54] actually available to anyone publicly [3:56] yet, and the rollout to plus and pro [3:58] users would happen in the coming days. [4:00] By the evening, Sam had posted an [4:01] apology for the messy rollout, and when [4:03] a pro subscriber asked Sam if he should [4:05] stay up and wait for it, Sam told him [4:07] something we've all heard after making a [4:09] desperate late night plea, "Go to bed." [4:12] Sam also told CNBC the model went [4:14] through a formal review with the Trump [4:15] administration before release. So, even [4:18] the government got to try AGI before we [4:20] did. As far as Astra itself, it was [4:22] pre-trained on more than 100,000 GPUs at [4:25] the Stargate site in Texas, and it's the [4:26] first model from OpenAI, where previous [4:29] models did a significant chunk of the [4:30] supervising during training. The main [4:32] pitch today seems to be computer use, [4:34] where Astra can fill out forms, crunch [4:36] spreadsheets, and operate engineering [4:38] tools like KiCad and Blender. On [4:40] OSWorld, which is a benchmark that drops [4:42] a model into a real desktop and makes it [4:44] do office work with a mouse and [4:46] keyboard, just like a human meatbag does [4:48] until it begs for Xanax, it scored 73% [4:51] while taking about 40 minutes for each [4:53] task, compared to Soul's 65% at 75 [4:56] minutes. And of course, the agents on [4:57] the Trust Me Bro benchmarks are the best [4:59] ever do it. It got 100% on exploit bench [5:02] and 65% on terminal bench science, which [5:05] beats the numbers Anthropic was bragging [5:06] about on Tuesday. It also got 99% on Arc [5:10] AGI 3, which is the benchmark that's [5:12] supposed to prove your model actually [5:14] generalizes instead of memorizing, which [5:16] is kind of the definition of AGI, [5:17] depending on who you ask. And according [5:19] to OpenAI, Astra is the first model to [5:21] hit the critical cyber threshold in its [5:23] preparedness framework, which is just a [5:25] fancy way of saying it can find and [5:27] exploit zero days on its own without a [5:29] human telling it what to do. Pricing is [5:31] the same as Fable 5.1 at $10 per million [5:34] input tokens and $50 out. But, what do [5:37] the anointed humans think about it? The [5:39] early access reviews are positive in the [5:41] way early access reviews typically are, [5:43] from people who don't publicly slander [5:45] Sam Altman. But, the demos, specifically [5:47] the ones that require spatial awareness, [5:49] are impressive. Sherif Shamim had it [5:51] recreate the Palace of Fine Arts in San [5:53] Francisco, and it did so almost [5:55] perfectly in Blender. And Thomas Recart [5:57] from OpenAI posted a walk-through of the [5:59] demo house from the launch post, which [6:01] Astra first modeled in Blender and then [6:03] turned into a fully walkable Unreal 5 [6:05] scene. But the most interesting was Matt [6:07] Schumer, who asked Astra to build a [6:09] world in Unreal Engine and fill it with [6:11] a dozen Astra-powered agents. A day [6:13] later, he heard voices coming from his [6:15] living room and walked out to a bunch of [6:17] agents talking about what I can only [6:19] assume was a plan to create a dating app [6:21] for horses. With all that said, when [6:23] Artificial Analysis ran Astra through [6:25] their Independent Intelligence Index, it [6:27] scored a 61, which is exactly the same [6:30] as GPT 5.6 Soul and 5 points behind [6:33] Fable 5.1. Something doesn't add up [6:35] here, but one thing that does add up is [6:37] Code Rabbit, the sponsor of today's [6:39] video. They got tired of seeing me make [6:41] new videos about new exploits hitting [6:42] the timeline every week, so they just [6:44] launched Code Rabbit Security to provide [6:46] continuous code security that's powered [6:48] by actual reasoning instead of brittle [6:50] regex rules. Their security agents are [6:53] designed to think like an attacker in [6:55] order to hunt for real vulnerabilities [6:57] across your entire codebase. Then it [6:58] prioritizes actual risks based on [7:01] reachability, exploitability, and blast [7:03] radius. And it explains the problem to [7:05] you in plain English and suggests a fix [7:07] you can approve and merge from the diff. [7:09] The Code Rabbit Security reviews every [7:11] PR before merge, and you can also [7:13] schedule deep scans across your full [7:14] codebase to protect your backdoor 24/7. [7:17] You get 10 free code scans to try it out [7:19] for free today at the link below. This [7:21] has been The Code Report. Thanks for [7:22] watching, and I will see you in the next [7:24] one.