Hacktoberfest 2026: le issue che i maintainer hanno segnato per ottobre, aperte e adatte ai principianti. Sfoglia le issue Hacktoberfest

10 expensive OpenAI models ship without a price, because the $10,000 request cap was sized for Anthropic's rates

Aperta
  • Qualcosa non va?
#1,406 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
5/5
Tempo stimato
Più di una settimana
Idoneità per principianti
15/100
Tipo di issue
Funzionalità
Chiarezza
Da chiarire
Stato di attività
Attiva
Stack tecnologico
go
Ambito
backend

Direzione di ricerca

The decision sits between the $10,000 cost cap and the bundled pricing table; start with pricing.BundleableRate and TestMaxPlausibleRequestCostMicros_Derivation, which fails whenever a dear model is bundled. Read docs/pricing.md for the unpriced-traffic signal. The issue offers three options (raise the cap, make the margin rule per-model, or lower the 100× requirement) and none is chosen, so 'done' starts with a maintainer recording the decision before any code changes.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

Context for this: Cortex now bundles OpenAI's rates alongside Anthropic's, so OpenAI traffic gets a cost instead of showing blank. Ten of the dearest models are deliberately left out. This explains why, and what it would take to include them.

The short version

Cortex refuses to believe any single request cost more than $10,000 — above that it assumes a glitch rather than a bill, because a gateway can report a cost in an HTTP header and one nonsense value would land permanently in your spend history.

That $10,000 is not a round number someone picked. It is calculated, and the calculation assumes no model costs more than about $75 per million tokens — which was true when Anthropic was the only vendor in the table.

OpenAI sells ten models far dearer than that:

Model Price
o1-pro, o1-pro-2025-03-19 $600 / million tokens
gpt-5.5-pro (+ dated) $180
gpt-5.4-pro (+ dated) $180
gpt-5.2-pro (+ dated) $168
gpt-5-pro (+ dated) $120

Include those and the $10,000 limit stops having the safety margin it was built with. So the generator leaves them out.

Why the margin matters

The rule is that the cap should sit at least 100× above a realistically large request, so that a vendor raising prices does not suddenly cause Cortex to reject real bills.

Dearest bundled model A large real request costs Margin under the $10,000 cap
claude-opus-4 — $75/Mtok $79.80 125× ✅
o1-pro — $600/Mtok $638.40 15.7× ❌

That check is in TestMaxPlausibleRequestCostMicros_Derivation, and it fails if a dear model is bundled. It uses list prices and a real context window — no inflated assumptions — so it is a genuine statement about margin, not an artifact.

What this costs today: nothing

Those ten models have no price right now either, because the bundled table has only ever held Anthropic models. So nobody is worse off than before, while 142 OpenAI models gain prices they did not have (the table goes from 33 entries to 175).

They are not silently wrong, either. An unpriced model is reported as priceable but unpriced and named in unpricedBy, which is the designed signal — see Finding traffic that is not priced.

Workaround if you need one of them priced

Four lines in ~/.cortex/config.yaml, and your own entry always beats the bundled table:

pricing:
  endpoints:
    - hosts: ["api.openai.com"]
      models:
        "o1-pro":
          input_cost_per_million:  600.00
          output_cost_per_million: 600.00

To include them properly, pick one

  • Raise the cap so the margin rule is satisfied with a $600/Mtok model — roughly $64,000 instead of $10,000. Simple, and nothing needs skipping. The cost is a 6× looser bound on a figure that exists to catch garbage.
  • Make the margin rule per-model, comparing each model against the tokens it can actually accept rather than one global figure. o1-pro only takes 300,000 tokens, so its worst real call is $180 and its margin is 55× — better, but still under 100×, so the threshold would need revisiting too.
  • Lower the 100× requirement with a written argument for the new number. It is a judgement call, not a derivation, and 15× is still a wide margin by most readings.

No action needed unless someone actually runs one of these models through Cortex. Filed so the decision is recorded rather than rediscovered.

Note on the generator's rule

The skip is written on the rate, never the vendor — so the first Anthropic model past the bound is dropped on identical terms. It is not an OpenAI carve-out. See pricing.BundleableRate.

Separately: Codex's own model is still unpriced

Worth knowing if that is why you are reading this. gpt-6-luna, which Codex uses, is not affected by the skip — at $0.50/Mtok it is 1,200× clear of the bound. It is missing for an unrelated reason: the bundled table is pinned to a LiteLLM commit that predates that model's addition upstream. Pricing it needs a commit bump, which is its own change because it also moves Anthropic's rates.

Related: #1365 (Codex parity), #942.

Assisted-By: Claude (Anthropic AI) [email protected]

Lingua principale
Go
Stelle
16
Fork
41
Merge medio
11h 23m
PR unite (30g)
275

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di rossoctl/cortex

Tutte le issue di rossoctl/cortex

Issue simili

Altre issue su Go

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.