GPT-Rosalind billing starts October 5: verify API pricing and eligibility

Distinguish the billing date from model availability

OpenAI’s September 8, 2026 API changelog already recorded GPT-Rosalind (model ID gpt-rosalind-research) becoming available through a trusted-access program for approved internal life-sciences research, with billing beginning October 5, 2026. The current API pricing page and product help page also confirm this date.

At this article’s October 2 check, October 5 was still a future billing date, not evidence that charges had already occurred. OpenAI did not specify an exact effective hour or time zone; do not convert it to midnight Beijing time yourself. Organizations must verify historical trials, separate agreements, taxes and actual invoices. The start date alone does not imply that all historical usage was free.

Who can use it, and what the API may be used for

The product help page positions it as an enterprise life-sciences R&D product. Eligible organizations with Enterprise or Business agreements can apply through trusted access, still limited to approved users. ChatGPT Enterprise, Codex and the API are entry points whose approval scope must be checked separately. Having a paid plan, joining an organization or seeing prices does not establish model authorization.

The API currently supports only approved internal research tools, workflows and applications, not customer products or external commercial applications. Teams turning research prototypes into customer services should confirm the scope with OpenAI first; internal approval must not be carried over automatically to an external product. Rosalind Workbench can be used without GPT-Rosalind model eligibility, so seeing Workbench does not establish model approval either.

API dollar pricing and two credit units

The Standard table for API Specialized models lists gpt-rosalind-research at $5 for input, $0.50 for cached input and $25 for output per million tokens. It explicitly says cache-write pricing does not apply to this model; do not borrow cache-write multipliers from other models. These are standard model-token rates, not a fixed quote for each research task or the total price of every service.

The ChatGPT credits rate card separately lists GPT-Rosalind-Research Chat messages at approximately 20 credits. Work/Codex Standard rates per million tokens are 125 input, 12.50 cached-input and 625 output credits, with model access limited to certain Enterprise plans. These two credit tables and the API dollar table use different units. The rate card also notes that some Enterprise agreements bill tokens in dollars; the agreement determines applicable rates. Do not independently convert credits to dollars or treat a visible table as model eligibility for all Business seats.

Estimate API model costs from the actual token mix

Split input into uncached input U and cache-hit input C, and denote output as O, all in tokens. At Standard rates, the dollar estimate is (5 × U + 0.50 × C + 25 × O) ÷ 1,000,000. U and C are non-overlapping input portions; do not charge all input at full price and then add cached input again. The official caching documentation describes cached-usage fields in usage. Verify actual responses or invoice classifications rather than assuming every request hits the cache.

Suppose a set of calls uses 500,000 uncached input, 200,000 cached input and 100,000 output tokens. The model-token estimate is $2.50 + $0.10 + $2.50 = $5.10. If those 200,000 input tokens do not hit the cache, the same total input and output estimate becomes $3.50 + $2.50 = $6.00. Both are calculations, not our measurements, invoices or savings guarantees. Check tools, storage, applicable processing options and contractual extra charges separately; do not label the model-token subtotal as the entire research cost.

A checklist for before and after October 5

First, ask the organizational owner to confirm model approval, whether it covers the API or ChatGPT/Codex, and whether the research remains internal. Second, record the organization, workspace, billing channel and applicable agreement. Compare API usage with the dollar table and ChatGPT/Codex with your own rate card, retaining the verification date.

Third, use the team’s existing compliant call records to collect the model ID, uncached input, cached input and output volumes. Estimate with the formula above and allow headroom for output growth and cache misses. Fourth, compare actual usage and invoice classifications on the same channel around the official billing date. If the exact start time is unpublished, confirm it with the OpenAI account team rather than scheduling large batches from your own time-zone assumption.

Fifth, if actual charges or approval scope differ from expectations, stop adding batch research tasks, retain necessary usage, error and billing evidence, and ask organizational administrators or OpenAI Support to investigate. Do not change accounts, expand permissions or switch network egress to solve authorization or billing issues. This article has not entered accounts, changed permissions or initiated billable calls.

What to check first if the model remains invisible after approval

The official troubleshooting guidance says to verify that you are signed into the approved ChatGPT workspace or API organization and ask administrators to check groups, custom roles and relevant permissions. Codex/API also require approval for the corresponding entry point. Visibility in ChatGPT does not establish API access.

When distinguishing the model, API key, base URL and organization, work through the API connection-parameter checklist. Eligibility, permissions and billing ownership are managed by the service provider and organization. A proxy cannot increase model eligibility or change rates.

What is confirmed and what the organization must still verify

Confirmed items are the official billing date, the three Standard API token rates, the absence of cache-write pricing for this model, and the trusted-access and internal-research restrictions. This article has not verified the effective hour, time zone, tenant eligibility, contractual rates, invoice changes or actual charges.

This article helps teams check cost and access. It does not evaluate research accuracy or clinical uses, nor provide a general API deployment example. Before launching research workflows, the responsible team should verify its data, purpose and organizational authorization. Future price or eligibility changes should be substantively incorporated into this same page.