The Arms Race Nobody Voted For

What leaked is worth paying attention to. Mythos is described internally as Anthropic's most capable model to date, a step above the existing Claude Opus tier in reasoning, coding, and cybersecurity. That last part is the interesting wrinkle. The model is apparently so capable in the security domain that Anthropic is holding back full deployment over concerns about what it could do in the wrong hands. Training is complete. A small group of trusted users already has access. The public rollout is being paced deliberately, which is either responsible safety practice or the most effective marketing strategy in AI history. Probably both.

Spud is a different animal. OpenAI finished pre-training it around March 25, and the framing is not another increment on GPT's existing line. The focus is on agents that are always available, always working, moving from a tool you call on to something closer to a system that runs in the background of your work. OpenAI even shut down Sora, its video generation product, to redirect compute toward Spud. That is a real signal. When a company kills a product that still has users to feed a new one, it is not hedging. It believes the new thing is where the value is going.

The thing both models share is a direction. The previous generation of AI products asked you to prompt them. You showed up, typed something, got an answer, and moved on. What Mythos and Spud are both pointing at is a model that participates more continuously. Fewer questions and answers, more something that sits inside your workflow and operates over time. For developers, that could mean code review that doesn't wait to be asked. For security teams, something that actively monitors rather than just responds. That shift sounds incremental when you describe it. In practice, it is not.

There is a tension here that neither company is quite saying out loud. The same capability that makes Mythos useful for threat detection is the capability that makes it dangerous in the wrong hands. Anthropic is aware of this, which is why they are sitting on full access even though training is done. But awareness doesn't resolve the problem. It just delays it. At some point, a capable model gets released, gets fine-tuned, and gets misused. The question is not whether that happens but how much damage it causes when it does. Both companies are betting that the good use cases outweigh the bad ones. That bet has always been part of the deal with powerful technology. It doesn't get less uncomfortable with repetition.

For the African tech ecosystem, the practical question is what this generation of models costs to use and who gets access first. The pattern with frontier models has been that pricing eventually becomes accessible, but the first window of high capability is expensive and quota-limited. Kenyan developers building on top of these APIs are good at extracting a lot from limited compute, partly because they have always had to be. If Spud or Mythos genuinely delivers 2x or 3x performance on coding benchmarks, that multiplier matters most where every token spent has to count. The efficiency argument for better models has always been stronger in markets where margins are thin and compute isn't cheap relative to output value.

What keeps coming back is the timing. Both leaks happened within days of each other, both companies are navigating major fundraising or IPO preparation, and both models represent a genuine shift in what AI can do without being asked. None of that is a coincidence. The race right now is not about individual features. It is about who controls the layer that sits below everything else, the model that others build on, the infrastructure that becomes the default. Mythos and Spud are bids for that position. The companies that get in early, build on the right one, and survive the move toward agentic AI will have structural advantages that won't be obvious until it's too late to close the gap.