BYOK provider keys

Limited-access models

Some models require elevated provider-account access. What that means, why the two OpenAI pro variants are target-only, and how to upgrade your account tier.

Updated 2026-04-25 · By Jon Lasley

On this page

A small number of provider models require elevated account access beyond a standard API key. The Target Model picker marks these with a Limited access badge and a tooltip. As of August 2026 both limited-access models are also target-only, so you can author prompts for them but Prompt Assay never calls them on your behalf.

Which models

As of August 2026, the limited-access models in Prompt Assay are:

`gpt-5.5-pro` · OpenAI's highest-precision GPT-5.5 variant. Background mode only (no streaming).
`gpt-5.4-pro` · OpenAI's enhanced-precision GPT-5.4 variant.

Target Model only, not Workbench Model

Both -pro variants are target-only entries. You can set either as a prompt's Target Model and get the full metadata treatment: linter rules keyed to the model's characteristics, context-window warnings sized to its 1.05M window, cost estimates at its rates, and export. They are not offered as a Workbench Model, and they don't appear in the playground, in Compare models, or in the Behavioral Eval model picker.

Why they can't be a Workbench Model

OpenAI serves gpt-5.5-pro and gpt-5.4-pro on the Responses API only. Prompt Assay's OpenAI adapter speaks Chat Completions exclusively, so there is no code path that can reach them. They were previously offered as Workbench Models but could never actually complete a call · demoting them to target-only replaces a runtime failure with an honest, up-front absence from the picker.

Why these models are gated at all

OpenAI's 'pro' variants are gated by account tier, not by an opt-in form. Higher tiers unlock larger rate limits and access to the -pro family. The standard non-pro models (gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.4-nano) are available to every paid OpenAI account and are the ones you actually run against.

Access on your own infrastructure

Because Prompt Assay doesn't call these models, you won't see a 403 inside the product. The gate matters when you take the prompt to your own runtime: if your OpenAI account isn't entitled, a direct Responses API call to a -pro model returns 403 Forbidden. Your BYOK key is not the problem · the key is valid; the model just isn't on your account's allowed list.

How to upgrade

Open the OpenAI dashboard and check your usage tier (Tier 1 through Tier 5).
Tier eligibility is automatic based on lifetime spend on your OpenAI account. Tier 1 typically requires ~$5 of lifetime usage; higher tiers need more.
Once you're on a tier that includes the -pro family, the model becomes immediately available on the same API key · no key rotation needed.

Pricing reminder

Limited-access models are 6-12x more expensive than the standard variants. Both gpt-5.5-pro and gpt-5.4-pro are $30 input / $180 output per 1M tokens, against gpt-5.5 at $5 / $30 (6x) and gpt-5.4 at $2.50 / $15 (12x). Neither -pro variant offers a cached-input discount, so there is no cache rate to soften a long prompt, and both switch to $60 input / $270 output once a call exceeds 272K input tokens. The token-budget bar prices a target-only model at these rates so the estimate is honest before you take the prompt elsewhere.

Other 'limited-access' situations (none today)

The badge pattern is reusable. If a future model · from any provider · requires special access (a feature opt-in, an enterprise contract, a specific account tier), it will surface here under the same Limited access badge with the same tooltip pointing back to this page. As of August 2026, only the two OpenAI pro variants above qualify.