Claim
Model progress is not tapering; every few months something previously impossible becomes easy, and the size of the leaps grows. The operating implication for any AI-native flywheel commitment: pin workflows to capability classes you can re-baseline each quarter, not to one model snapshot.
Mechanism
Workflows hardcoded to a specific model version inherit that model's failure modes and miss the gains of the next release. Capability-pinned workflows (e.g. "long-context summarization at quality X") survive model swaps and benefit from each upgrade.
Conditions
Holds when: the team has a quarterly cadence to re-evaluate model choice per workflow.
Fails when: a workflow depends on a model-specific quirk that capability framing doesn't capture (rare).
Evidence
"Every few months a new model arrives... something that was impossible becomes easy, while the size of the leaps grows each new release cycle."
— Ethan Mollick, One Useful Thing, 2026-04-23
Signals
- Workflow specs name capability requirements, not model IDs.
- Quarterly model bake-offs run as a standing scheduler task.
- Long-form fiction or other rough-edge tasks flagged with explicit caveats per release.
Counter-evidence
For regulated or audited workflows, model-pinning is required for reproducibility — capability framing can't override compliance.
Cross-references
- (none in current corpus)