A practical escalation rule
- Run a set of representative short tasks on Hydrogen. Score each against a written acceptance checklist.
- Group failures by cause, such as incomplete edits, missing evidence or context lost across steps.
- Run the failed tasks and a small sample of successful ones on Neon. Compare correction time as well as runtime.
Evaluate a real handoff
For support tickets, give both models the same notes and require an owner, next action and cited evidence. For coding, freeze the repo state and run the same tests. If both pass at the same rate, choose using measured cost and turnaround. Route more difficult jobs separately if the difference is concentrated there.
DeepMind Hydrogen model card · Benchmark worksheet
Primary sources
Google launch announcement · DeepMind Neon model card