Anthropic kept its new model, Mythos, out of public release, only a douzaine of enterprises get limited access. The stated reason: security.
Yes, Mythos might be powerful enough to worry security teams, but part of the secrecy reads like a smart marketing move and not as a pure risk mitigation.
The bigger point is simple: models are getting stronger and adoption is exploding. Pew finds 62% of US adults use AI weekly, and McKinsey’s 2025 survey shows 88% of organizations use AI in at least one function.
Stronger models + surging usage = far more compute demand. Demand can spike exponentially; datacentre capacity grows far more slowly. You can’t plug power, add cooling, set up generator, connect network and racks servers overnight, datacentre projects take 6–24 months.
More, estimates put GPU demand in the millions of H100 equivalents, well beyond what the industry can produce today.
When demand outstrips supply, vendors ration access. The best models and largest number of tokens will go to the highest bidders; everyone else gets older models, throttled contexts, higher latency, or no access during peak times. Free tiers will, also, erode.
That creates a new business risk: workflows tuned to a specific model may only run reliably behind a paywall in the next months, sometimes an expensive one. The sponsored run we enjoy today will be over soon. Either you pay a premium that kills margins, or accept degraded service when vendors throttle subscriptions.
Don’t lock yourself to a single vendor. Choosing models by public benchmarks is risky; what matters is how a model performs on your data and in your workflows, not leader board scores.
If you are only using Claude Code, you should start working also with GitHub Copilot CLI, or/and OpenAI Codex or OpenClaude.
You need to run your workflows on several model, not as an experiment, but as a routing. There are to possible outcome.
- The output is similar, with or without a few adjustments, great you have a low lock-in risk
- The output is degraded; you have a decency you must mitigate.
Adopt a model agnostic mindset. Prompts and context that rely on vendor quirks (Claude’s long context, GPT’s tool syntax, Gemini’s multimodal shortcuts) create hidden technical debt, and fragile, costly systems.
Plan for scarcity: design for multiple models, measure outputs on your real data, and treat model access as a strategic dependency.
Top comments (0)