Knowing what we now know about Anthropic’s Mythos model, it would have been difficult indeed to create something more attractive to the Trump Administration. Or, indeed, any president sitting in the White House, particularly one facing hostile cyber threats from Iran, Russia, and China.

Mythos is a tier higher and quite a bit smarter than the Opus mixture of expert model that the company has been peddling for two years and which was released to a select few tech titans because it poses a significant cybersecurity risk because it does an excellent job of finding security vulnerabilities in the vast installed base of software in the world.

In its announcement of the Mythos preview earlier this week, Anthropic didn’t say much about the architecture of its forthcoming model or when it might be something that can be released to the public, but it did freak the hell out of everyone when the reveal was buried in an announcement about the Project Glasswing security effort.

“Mythos Preview has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser,” Anthropic wrote in its announcement, and that is its emphasis there, not mine. “Given the rate of AI progress, it will not be long before such capabilities proliferate, potentially beyond actors who are committed to deploying them safely. The fallout – for economies, public safety, and national security – could be severe. Project Glasswing is an urgent attempt to put these capabilities to work for defensive purposes.”

Hence the partnership with Amazon Web Services, Apple, Broadcom, Cisco Systems, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks, who are using the Mythos model to comb through their code to see where the holes are. Perhaps they will also be able to use Mythos to fix the security vulnerabilities, too? And maybe when they recompile, there won’t be any shenanigans that the model hides in the code. . . .

This publication is about system architecture, not the effects of technology on society and culture and business. Nonetheless, we think it is noteworthy that one of the major model makers is worried about letting one of its creations loose on the world and is actually not releasing it. No one is saying that Mythos won’t be available in some form at some point in the future, but at least there is careful consideration at some level by a model builder, and not because of poor quality but because the model seems to be so damned good at what it does.

It was a strange moment in a strange week.

As has been the case with Haiku, Sonnet, and Opus variations of the Claude model, Anthropic is being tight lipped about the Mythos variant. There is no talk of parameters and training data set size, which we have been told by all of the model makers doesn’t matter anymore. I think this is bullshit, as you might imagine. The size and quality of the training data set and the number of parameters embodied in the model overall and the number of parameters that can be activated at the same time in a mixture of experts model absolutely matter. None of the model makers wants to tell us because we can therefore infer things about their models and how they work.

We are able to glean some things from the Project Glasswing announcement, the technical deep dive for the Mythos preview, and the system card – what you and I might call the spec sheet – for Mythos. The alignment risk update for the Mythos preview is here.

Mythos was developed under the codename “Capybara,” which is a water-loving, 150 pound varmint from South America that, given enough time – like maybe 50 million years – might evolve into something akin to a dolphin. It is hard to say why Anthropic called it that, but what is clear from the benchmark results is that Mythos represents a significant shift from its chattybot Sonnet and reasoning Opus models. Mythos might be an MoE, or it might be something new.

Mythos has the same context window as the current Sonnet 4.6 and Opus 4.6 models, at 1 million tokens, but we strongly expect that it was designed to scale well beyond this. There is chatter that the Mythos model has 10 trillion parameters, but no credible source as far as I know. Mythos is a reasoning model, and can be thought of as better than Opus, although it remains to be seen if it will ever be officially commercialized.

Logan Graham, head of the frontier red team at Anthropic, spoke to NBC News last week, and had this to say about Mythos: “The degree of its autonomy and sort of long ranged-ness, the ability to put multiple things together, I think, is a particular thing about this model.” The system card for Mythos says that the model engages like a collaborator, is opinionated, stands its ground, writes densely, assumes we share its context, has a recognizable voice, can describe its own patterns clearly, and makes mistakes in subtle and not obvious ways. (If you think that sounds like you, you are not alone. . . .)

We can infer a few things from the benchmark data published in the system card, which is shown below:

The alignment risk update, which is an assessment of how safe Mythos is and how it might cause harm through its actions, says this: “The difference in capabilities between Mythos Preview and Claude Opus 4.6 is larger than the difference between previous releases.”

So we presume there is some insight that makes a big improvement in the model. This is apparent is the various SWE-bench code assistant benchmarks as well as in the GraphWalks BFE test, which forces reasoning across a massive graph of hexadecimal hashes and is a kind of test against “context rot,” where models get lazy about actually looking at a massive context window and just work with what they remember. (Not to anthropomorphize too much, of course.) The jump in performance of Mythos over Opus 4.6 on Humanity’s Last Exam (perhaps not ironic, and not multiple choice but including problems models cannot solve and domain experts – meaning people – struggle with but can solve) and on the Charxiv reasoning benchmark (checks how models reason from charts and tables and figures) shows something significant and new is in Mythos that is not in Opus 4.6, GPT-5.4 from OpenAI, and Gemini 3.1 Pro from Google. And finally Mythos is very good at solving complex math problems, as shown by the USAMO test, much better than Open 4.6, considerably better than Gemini 3.1 Pro, and on par with GPT-5.4.

To sum it up plainly: Everyone is going to want the Mythos model, and maybe Anthropic is worried and maybe it is playing a little hard to get at the same time.

Meta Platforms Gets It Frontier Model Act Together.

Meta Platforms co-founder and chief executive officer Mark Zuckerberg spent $14.3 billion on an “acquihire” of Scale AI to get Alexandr Wang to head up Meta Superintelligence Labs last June, and the first fruits of the new AI team at Meta Platforms were also unveiled last week.

The new model is called Muse Spark, which will be a closed source model and which is a top five frontier model but not yet the top of the pack. It is, however, better at many things than the prior open source Llama 4 models that Meta Platforms unveiled last April. The top-end Llama 4 Behemoth model, with 2 trillion parameters overall with the ability to have 16 models active with up to 288 billion parameters over them at any given time, never made it into the field. This was a bit of an embarrassment to Zuckerberg, who went on a spending spree to bring in new talent to craft Muse Spark, and presumably other variations of the Muse theme.

Muse Spark is a multi-modal reasoning model, and Meta Platforms was very careful not to compare it to the prior Llama 3 and Llama 4 releases. The company did say that it took one tenth as much compute power to train Muse Spark than it did for the Llama 4 Maverick model.

Llama 4 Maverick has 400 billion parameters, about the same as the top-end Llama 3.1 and 3.2 dense models, but in this case, it is a reasoning model and only 17 billion of its parameters are active across 128 experts at any time. The Maverick model is interesting in that it has a 1 million token context length for input, and is also multimodal input (image, text, video, sound). The comparison seems to suggest that Muse Spark is a kind of follow-on for Maverick, and Meta Platforms did say straight up that it was working on larger Muse variations. Perhaps something on the scale of Llama 4 Behemoth, perhaps even bigger.

The test scores for Muse Spark are acceptable and competitive, but the average grade across all tests still looks like an F+ or D- for all of the vendors. So, sleep easy for a bit, humanity.

Muse Spark is in a controlled prelease via an API. It is not clear how or when it will be licensed, but it is pretty clear that it will not be open sourced.