AI News · 5 October 2026 · Reporting cutoff: 18:01 Asia/Karachi (13:01 UTC).
OpenAI’s published API billing for GPT-Rosalind takes effect on 5 October 2026, with standard rates of $5 per million input tokens, $0.50 for cached input and $25 for output. The change matters for research teams moving recurring workflows into a paid operating budget. OpenAI specifies the effective date without an exact activation time or timezone. These are published terms, rather than charges we have observed on an account. OpenAI API pricing.
The model remains restricted to approved internal research. A public price does not mean a developer can immediately add it to a customer-facing SaaS product. Below, we separate the billing change from the earlier release, calculate an illustrative workload and explain what teams should establish before committing a budget.
What changes today, and what was announced earlier?
GPT-Rosalind is OpenAI’s specialist life sciences research model. Its original announcement was published on 16 April 2026; OpenAI describes applications in biological research, scientific evidence synthesis and experimental planning. Its 11 September update says the model was leaving research preview for eligible organizations globally and identifies today as the pricing effective date. Original announcement and September update.
There is a separate dated record: the 8 September API changelog describes general availability through the trusted-access program and lists the October billing date. General availability here applies within that approved program. The announcement update and changelog have different publication dates; neither establishes a fresh model launch on 5 October. OpenAI API changelog.
For a lab already evaluating the model, the immediate decision is how to fund continuing use. For a team considering it for the first time, approval and permitted use come before price comparisons. Today’s change supplies a budget baseline; it does not establish better scientific results than the version a team previously evaluated.
GPT-Rosalind API pricing: the published standard rates
| Usage component | Standard rate |
|---|---|
| Input | $5.00 |
| Cached input | $0.50 |
| Output | $25.00 |
OpenAI says cache-write pricing does not apply to this model. This table covers model tokens at the listed standard rates, not a complete procurement quote or ChatGPT subscription price. Confirm the applicable agreement and any separate tool, infrastructure or regional charges when estimating a deployment. Official pricing and billing note.
A worked cost example for a research workflow
Illustrative arithmetic, not a test result: suppose one completed workflow uses 40,000 uncached input tokens and 8,000 billed output tokens across its model calls. Applying the rates in the September pricing announcement, input costs $0.20 and output costs $0.20, for $0.40 in model-token charges. One hundred identical workflows would cost $40.
| Scenario | Input cost | Output cost | Total |
|---|---|---|---|
| 40,000 uncached input; 8,000 output | $0.20 | $0.20 | $0.40 |
| 20,000 uncached + 20,000 cached input; 8,000 output | $0.11 | $0.20 | $0.31 |
The second scenario assumes half the input is actually billed as cached. Across 100 workflows, that assumption changes model-token cost from $40 to $31. It is a planning scenario, not a guaranteed cache hit rate. Reusing the same document in an application is not, by itself, evidence of a billed cache hit.
Budget for the complete task rather than its final answer. An agent may make several calls while collecting evidence, revising a hypothesis or recovering from a failed step. Record those calls together. A concise final report can still follow a costly sequence of intermediate requests.
Who can use it, and can it power a commercial app?
OpenAI’s current access guidance describes globally eligible organizations with Enterprise or Business agreements, approved users and trusted-access requirements. It identifies ChatGPT Enterprise, Codex and approved API workflows as access surfaces. It explicitly excludes customer-facing products and external commercial applications at present. GPT-Rosalind availability and access guidance.
That distinction changes a developer’s planning. An internal research assistant and a service sold to outside customers are different deployment proposals. Obtain approval for the intended workflow before investing in a production integration. A Business agreement alone should not be treated as automatic provisioning on every surface.
If an approved researcher cannot see the model, OpenAI recommends checking the approved workspace or API organization and confirming the user’s assigned permissions with an administrator. Access on one surface should not be assumed to provision another. Official access troubleshooting.
How to evaluate value once a team has access
Choose a repeatable internal task with an outcome researchers can judge, such as producing a cited evidence summary from a fixed collection of permitted documents. Keep the question and source collection stable between runs. Ask a domain expert to assess source fidelity, missing evidence and the amount of correction needed before using the output.
Track cost per acceptable result alongside elapsed time. A cheaper run that needs extensive revision may be less useful than a more expensive run that saves expert time. Equally, a polished report is weak evidence of value if its citations do not support its conclusions. Compare the same task with the team’s current method before expanding use.
Use a small initial batch, review its failures, then estimate the monthly workload from observed usage. Establish an owner for the budget and a stopping point for repeated unsuccessful runs. These are suggested evaluation steps, not a claim that AIX Radar has tested GPT-Rosalind or demonstrated scientific benefits.
For context on the broader model releases, read our GPT-6.1 Sol, Claude Sonnet 5.5 and Gemini 4 weekly roundup. Follow the AI News archive for further verified developments.
Reporting method: AIX Radar checked OpenAI’s original announcement, dated changelog, API price table and access guidance. The cost examples are our calculations from published rates. We did not call the model, inspect a customer invoice or independently evaluate its scientific performance.
