The Model Menu Is Getting Out of Hand

Vercel's AI Gateway added Kimi K2.7 Code, Nemotron 3 Ultra, Grok Imagine Video 1.5, and DeepSeek via Azure inside the same news cycle. Claude Fable 5 was suspended, then reinstated. The message is clear: the model layer is moving too fast for anyone to track, including the platforms distributing them.

This creates a real problem for product teams. Choosing a model used to be a quarterly decision. Now it's a weekly one, and the switching costs—prompt tuning, evals, cost recalibration—are non-trivial. Anthropic still dominates spend according to Vercel's own data, but DeepSeek is entering the fight for token volume. That gap between spend and volume is worth watching. It suggests teams are running DeepSeek for high-frequency, lower-stakes tasks while keeping Anthropic for work that matters more.

The practical move for teams right now: stop optimizing for the best model and start building model-agnostic evals. If your product's quality depends on a specific model staying available—Claude Fable 5 was suspended and back within days—you're one vendor decision away from a bad week.

The AI SDK update that now supports Claude Code, Codex, and Pi as agent harnesses is relevant here too. Abstraction layers exist precisely because the underlying models are unstable. Use them.