You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Following #166, a question about the other side of model identity: models that are absent from model_cost/model_cost.json entirely, and specifically how promotional or preview endpoints should be handled.
What happens today
A prediction naming a model with no price entry is not scored as an expensive route — it is dropped before evaluation. Running the harness locally on 5,273 full-split rows generated by stealth/ox-alpha:
ERROR - Error converting model name 'stealth/ox-alpha' to universal name:
Model name stealth/ox-alpha not found in universal_names or mapping
INFO - Total: 5273 | Evaluated: 0 | Skipped: 5273 | Failed: 0
ERROR - Fatal error: No entries have valid cost values.
Cannot compute RouterArena score without cost data.
Every row is skipped, then compute_router_metrics refuses to score. In a submission that routes to a mix of priced and unpriced models, those queries are effectively graded as failed inference. That is a defensible default, and we are not reporting it as a bug — the behaviour is safe. The question is what the intended path is for adding a model.
The concrete case, with the awkward part stated up front
stealth/ox-alpha is currently served on OpenRouter at $0 / 1M tokens in and out (1M context). We measured it on the full split under the standard zero-shot \boxed{} contract:
stealth/ox-alpha
gemini-3-flash-preview
Accuracy (full split)
78.18%
78.44%
Accuracy, external corpus (2,600 items)
78.85%*
78.85%*
Mean output tokens
256
97
Empty responses
3.9%
0%
*measured on an identical 52-item subsample; the 2,600-item run is still in progress at the time of writing.
So it is a genuinely capable model. But a $0 price entry would let any submission that routes to it saturate the cost term, and that advantage would be an artifact of a temporary promotion rather than of routing quality. Given #166 was precisely about a submission benefiting from a price that no longer reflected what was really being served, we would rather raise this than quietly exploit it — and we are not asking for a $0 entry.
What is the intended policy for free / promotional / preview endpoints? Options we can see, with no preference beyond wanting the rule to be stated:
exclude them until they carry a standing commercial price;
admit them at a reference price for their capability tier rather than at $0, so cost reflects the compute rather than the promotion;
admit them at $0 but flag such entries on the leaderboard as promotionally priced;
admit them only once a price is published, and re-grade if it changes.
Should the checker warn on unpriced model names at submission time? Right now the failure surfaces only at evaluation, as a skip and then a fatal error, which is easy to misread as a broken artifact. Fix Grok 4.1 Fast → Grok 4.3 mispricing on the leaderboard (#166) #172 added a preflight warning for retired/redirected slugs — the unpriced case seems like a natural neighbour to it.
Happy to open a PR for either the checker warning or a documentation note, whichever is more useful. And happy to be told the answer is simply "not eligible" — that is a perfectly good answer and we will design around it.
Hi RouterArena maintainers,
Following #166, a question about the other side of model identity: models that are absent from
model_cost/model_cost.jsonentirely, and specifically how promotional or preview endpoints should be handled.What happens today
A prediction naming a model with no price entry is not scored as an expensive route — it is dropped before evaluation. Running the harness locally on 5,273 full-split rows generated by
stealth/ox-alpha:Every row is skipped, then
compute_router_metricsrefuses to score. In a submission that routes to a mix of priced and unpriced models, those queries are effectively graded as failed inference. That is a defensible default, and we are not reporting it as a bug — the behaviour is safe. The question is what the intended path is for adding a model.The concrete case, with the awkward part stated up front
stealth/ox-alphais currently served on OpenRouter at $0 / 1M tokens in and out (1M context). We measured it on the full split under the standard zero-shot\boxed{}contract:stealth/ox-alphagemini-3-flash-preview*measured on an identical 52-item subsample; the 2,600-item run is still in progress at the time of writing.
So it is a genuinely capable model. But a $0 price entry would let any submission that routes to it saturate the cost term, and that advantage would be an artifact of a temporary promotion rather than of routing quality. Given #166 was precisely about a submission benefiting from a price that no longer reflected what was really being served, we would rather raise this than quietly exploit it — and we are not asking for a $0 entry.
What we are asking
Is there a documented process for requesting a model be added to
model_cost.json? A3M Router submission - what is the official listing process? #138 asks a related question about listing routers; we could not find the equivalent for models.What is the intended policy for free / promotional / preview endpoints? Options we can see, with no preference beyond wanting the rule to be stated:
Should the checker warn on unpriced model names at submission time? Right now the failure surfaces only at evaluation, as a skip and then a fatal error, which is easy to misread as a broken artifact. Fix Grok 4.1 Fast → Grok 4.3 mispricing on the leaderboard (#166) #172 added a preflight warning for retired/redirected slugs — the unpriced case seems like a natural neighbour to it.
Happy to open a PR for either the checker warning or a documentation note, whichever is more useful. And happy to be told the answer is simply "not eligible" — that is a perfectly good answer and we will design around it.
Thanks for the work on the benchmark.