Stay
Your current option clears the workload; switching cost buys no material improvement.
Cheap → default → frontierSame job. Same bar.
The upgrade has to do work.
A useful comparison does not end with “Model B is better.” It shows the first task where that difference changes success, human effort, or risk enough to matter.
Try the protocol previewInteractive protocol preview
Choose a coding-job difficulty.
This demonstrates test order only. It makes no claim about any current model.
Cost of a missed error
Selected job
Isolated function
Begin with the cheap serious contender. Define the acceptance bar, repeat the run, and stop if it clears reliably.
Escalation trigger: a repeatable miss, material cleanup, or a failure that is hard to detect.- Test 01Cheap
Always establish the cost floor first.
- Test 02Default
Run only when the first tier misses or is borderline.
- Test 03Frontier
Make the premium answer for the hardest unresolved work.
Why no model names? The supplied research defines a method, not current performance evidence. A real comparison will bind every result to exact model IDs, dates, settings, tools, repetitions, and artifacts.
The useful outcomes
A comparison should change an action.
Switch
The alternative changes success, review time, or failure risk enough to justify migration.
Route
Use a lower-cost default and reserve the premium tier for a named breakpoint.
Every head-to-head record
Same objective. Same materials. Same bar.
Change one decision variable at a time.
| Record | Cheap tier | Default tier | Frontier tier |
|---|---|---|---|
| Exact identity | Model ID + date | Model ID + date | Model ID + date |
| Acceptance bar | One job-specific definition of successful work | ||
| Repeatability | Repeated runs | Repeated runs | Repeated runs |
| Human effort | Correct + verify | Correct + verify | Correct + verify |
| Failure shape | Detect + recover | Detect + recover | Detect + recover |
| Decision | Enough, worth upgrading, or overkill—for this job only | ||
Current external-research comparison
Luna versus Terra: is 10× worth it?
The price ratio is current and checkable. The workflow payoff still needs matched repository evidence.
Decision rule
Evaluate Luna first. Escalate a diagnosed failure.
Compare the exact IDs, rate conditions, illustrative workflow baskets, public evidence boundaries, and the protocol needed to decide.
Open the full comparison →Before accepting the conclusion
Read what the comparison can—and cannot—prove.
A score is not permission to make a bigger claim than the job, bar, setup, and repetitions support.
See the evidence hierarchy