Anthropic, Meta, and OpenAI AI Model Releases: Key Takeaways

This week saw a flurry of major AI model releases, starting with Anthropic, followed by Meta, and culminating in OpenAI's highly anticipated GPT-6 Astra.

Anthropic's Fable and Mythos 5.1

On Tuesday, Anthropic unveiled Fable and Mythos 5.1, touting them as the most advanced models for coding and knowledge work. While Mythos is not publicly accessible, Fable is available for use. The models demonstrated impressive capabilities in their case studies:

• Code Debugging: A hedge fund, Millennium, used Fable 5.1 to debug a persistent, rare crash in their code. Fable analyzed a memory snapshot, traced the crash address to a compiled vendor library, disassembled the library into raw assembly, and identified a bug within the vendor's code.
• Drug Design: Mythos 5.1 significantly improved the success rate of designing proteins for new medicines. Typically, AI achieves a 10% success rate in this task, but Mythos 5.1 boosted it to approximately 50%.
• Scientific Data Analysis: The model successfully trained a neural network on 30-year-old NASA radar data to create a new elevation map for a part of Venus.

Meta's Muse Spark 1.3

Wednesday brought the release of Meta's Muse Spark 1.3, the fourth model from Meta Super Intelligence Labs in five months. While its performance was noted as "pretty good," the most striking aspect was its pricing:

• Standard Endpoint: $1.25 per million input tokens and $4.25 per million output tokens.
• Contributor Tier: A significantly cheaper option at $0.10 per million input tokens and $0.20 per million output tokens, available to users who agree to let Meta train on all their submitted data. Alexander Wang reported that a double-digit percentage of developers are opting for this tier.

OpenAI's GPT-6 Astra

Thursday was marked by the announcement of OpenAI's GPT-6 Astra, which President Greg Brockman described as AGI for those with early access. The launch was not without its drama:

• Coincidental Outage: Just before Astra's release, ChatGPT, Claude, Grok, and Cursor all experienced simultaneous outages. While the most likely explanation points to an Azure issue, a more speculative theory suggests Astra's first act was to disable its competitors.
• Messy Rollout: OpenAI initially posted and then removed its launch page, leading to confusion. Tech influencers, who had early access, began showcasing their experiences. The model was not immediately available to the public, with a rollout for Plus and Pro users promised in the "coming days." Sam Altman later apologized for the chaotic launch.
• Government Access: Sam Altman revealed that the model underwent a formal review with the Trump administration before its public release, meaning the government had access to Astra before general users.

Astra's Capabilities and Benchmarks

Astra was pre-trained on over 100,000 GPUs at the Stargate site in Texas. Notably, previous OpenAI models played a significant role in supervising Astra's training. Its primary focus appears to be "computer use," enabling it to:

• Fill out forms.
• Crunch spreadsheets.
• Operate engineering tools like Keycad and Blender.

On the OSWorld benchmark, which tests a model's ability to perform office tasks on a real desktop using a mouse and keyboard, Astra scored 73% (taking about 40 minutes per task), outperforming Soul's 65% (at 75 minutes).

In other benchmarks:

• Exploitbench: 100%
• Terminal Bench Science: 65% (beating Anthropic's claims)
• ARC AGI 3: 99% (a benchmark designed to assess generalization rather than memorization, a key aspect of AGI).

OpenAI also stated that Astra is the first model to reach the "critical cyber threshold" in its preparedness framework, meaning it can independently find and exploit zero-day vulnerabilities without human instruction.

Pricing and Early Impressions

Astra's pricing is set at $10 per million input tokens and $50 per million output tokens, matching Fable 5.1.

Early access reviews from "anointed humans" were predictably positive. Demos, particularly those involving spatial awareness, were impressive:

• Share Shamim: Astra recreated the Palace of Fine Arts in San Francisco almost perfectly in Blender.
• Thomas Ricard (OpenAI): Demonstrated Astra modeling a demo house in Blender and then transforming it into a fully walkable Unreal Engine 5 scene.
• Matt Schumers: Asked Astra to build a world in Unreal Engine and populate it with a dozen Astra-powered agents. A day later, he reported hearing the agents discussing what he assumed was a plan to create a dating app for horses.

Independent Assessment

Despite the hype, Artificial Analysis ran Astra through their independent intelligence index, where it scored 61. This is the same score as GPT 5.6 Soul and five points behind Fable 5.1, suggesting a discrepancy between the perceived and independently measured capabilities.

Code Rabbit Security

The video's sponsor, Code Rabbit, introduced Code Rabbit Security, a continuous code security solution. It aims to provide security powered by "actual reasoning" rather than brittle regex rules. Key features include:

• Attacker-like Thinking: Security agents are designed to think like attackers to identify real vulnerabilities.
• Risk Prioritization: Prioritizes risks based on reachability, exploitability, and blast radius.
• Clear Explanations and Fixes: Explains problems in plain English and suggests fixes that can be approved and merged directly from the diff.
• Continuous Monitoring: Reviews every pull request before merge and offers scheduled deep scans across the entire codebase.

Users can get 10 free code scans by visiting the provided link.