Technical Context
I dug into this story, and here's what emerges so far: the figure of 800B active parameters is floating around as an estimate from X discussions, not an official Anthropic release. Right there I hit the brakes, because for real AI implementation, rumors about a "huge model" are useless if you don't know what's actually accessible via API.
More consistent leaks suggest Mythos is an MoE model with a total size of around 10T parameters. The active part per forward pass, judging by restatements, could be somewhere in the 800B–1.2T range. That sounds more plausible than the bare phrase "800B base model."
Even more important: Mythos doesn't seem to be prepared as a mass-market public analogue to GPT-4 with open access for everyone. Multiple sources repeat the same idea: the model is being kept in early access, through restricted channels, due to cybersecurity risks. So the main question now isn't just about power; it's about who will even get to touch it.
Pricing is also murky. I've seen rumors of $25/$125 per million tokens, but that's still at the level of market whispers, not fact. So comparing Mythos to regular production models on TCO is premature.
What This Means for Business and Automation
If Mythos is truly that strong, winners will be teams needing heavy scenarios: security research, complex code, multi-step reasoning, agent pipelines with high answer quality. But if access stays restricted, most companies won't be able to just plug this model into their processes.
Losers will be those building a roadmap on hype. I wouldn't plan any AI automation around a model with no clear SLAs, no stable API, and no confirmed pricing.
In practice, the choice is usually more boring and more honest: either wait for restricted access and accept the risks, or build AI solutions for business on already available models with proper routing, validation, and a fallback layer. At Nahornyi AI Lab, we solve exactly these forks for clients: we don't guess on rumors; we design AI architecture so that business doesn't depend on a single flashy model. If you're facing a similar decision, I and my team can help you build AI integration without unnecessary magic and with clear economics.