OpenAI put a price on GPT-Rosalind for the first time: $5, $0.50 and $25, which is Claude Opus 5's card to the cent and $5 under the GPT-5.5 it is built on. The meter does not start until October 5, and there is still no page saying how much context the money buys.
GPT-Rosalind has been running since April 16 without a rate card. On September 8 a line appeared in the API changelog saying it was generally available through the trusted-access program and, in a second sentence most coverage skipped, printed the price. On September 11 OpenAI stapled an update to the top of both Rosalind blog posts, and the pricing page gained a row at the bottom, under "Specialized models," below chat-latest and a Codex model. We read the row, the footnote under it, the help-centre article it links to, and the enterprise credit card that prices the same model a second way. The number is ordinary. What surrounds it is not.
Photo by Egor Myznik on Unsplash
The row and its footnote
Here is the whole of what OpenAI publishes about the price, from the pricing page: a row reading "Life Sciences, gpt-rosalind-research, $5.00, $0.50, $25.00" and a four-sentence note. "Billing for gpt-rosalind-research begins on October 5, 2026. Cache-write pricing does not apply to this model. Access is limited to approved internal research through the trusted-access program. All eligible organizations will continue to get access to the latest GPT-Rosalind models as they're released." The Fast mode tab of the same table lists only gpt-5.3-codex, so there is no Fast rate. There is no Batch tab and no Flex tab for the specialised section at all, and no long-context column. Rosalind is Standard tier, short context, or nothing, which puts it in the same position as GPT-5.6 Cyber, the other model OpenAI sells only to people it has vetted.
| Per 1M tokens | GPT-Rosalind | GPT-5.5 | Claude Opus 5 | GPT-5.6 Sol | GPT-5.6 Cyber |
|---|---|---|---|---|---|
| Input | $5.00 | $5.00 | $5.00 | $4.00 | $12.50 |
| Cached input | $0.50 | $0.50 | $0.50 | $0.40 | $1.25 |
| Cache write | none | none | $6.25 | $5.00 | $15.625 |
| Output | $25.00 | $30.00 | $25.00 | $20.00 | $75.00 |
| Batch or Flex | no row | $2.50 / $15.00 | $2.50 / $12.50 | $2.00 / $10.00 | no row |
| Long context | no row | 2x / 1.5x over 272K | flat to 1M | 2x / 1.5x over 272K | no row |
| Billing starts | October 5, 2026 | billing | billing | billing | billing |
Two columns to hold side by side. GPT-5.5 is the base: OpenAI's June 3 update says the current Rosalind "combines GPT-5.5's agentic coding and tool-use capabilities with stronger model intelligence in core drug-discovery domains." The card is GPT-5.5's on input and cache, and $5 lighter on output, five-sixths of the base rate. Neither model charges to write a cache, because the 1.25x cache-write line is a GPT-5.6-generation thing and 5.5 predates it, so the "does not apply" sentence is an inheritance from the base rather than a concession. Claude Opus 5 is the coincidence: $5.00, $0.50, $25.00 on the three lines both companies publish, and then Anthropic bills $6.25 to put a token in the cache and OpenAI bills nothing. On an agent loop that rewrites a 50,000-token context every five minutes, that one line is the whole difference between the two cards.
Against GPT-5.6 Sol, the model OpenAI tells everyone else to use, Rosalind is 1.25x on all three lines, because Sol's promotional $4 / $0.40 / $20 is 0.8x of the $5 / $0.50 / $25 Sol charged until August 21. That promo is "available at least through November 21, 2026" per the same page. If Sol goes back to $5 / $30, Rosalind becomes 1.0x on input and 0.833x on output against it too.
A specialist priced under its base
We could not find another case of OpenAI selling a fine-tuned or domain-trained model for less than the model it came from. GPT-5.6 Cyber, the Daybreak red-team model, is $12.50 / $1.25 / $75.00, which was 2.5x Sol when both launched and is 3.125x Sol now that Sol has been cut and Cyber has not. Anthropic's equivalent, Mythos 5.1, is Fable 5.1's $10 / $50 card with the safeguards lifted, the same price as its base. Rosalind bills five-sixths of its base on the line that dominates a reasoning model's bill. Whether that is a pricing decision or a token-shape decision we cannot tell from the outside, but OpenAI did publish token counts, which almost nobody does, in the June 3 post.
| Benchmark | Rosalind | GPT-5.5 | Tokens, per OpenAI | Cost per task vs GPT-5.5 |
|---|---|---|---|---|
| GeneBench | 21.6% | 20.4% | 31% fewer | 31.0% to 42.5% less |
| MedChemBench | 27.5% | 25.1% | 7.2% fewer | 7.2% to 22.7% less |
| LabWorkBench | 63.2% | 55.8% | 5.3% fewer | 5.3% to 21.1% less |
The last column is a range because OpenAI says "tokens" and not which kind. If every saved token is input, the saving is the token saving, since input costs the same on both models. If every saved token is output, take five-sixths of that again: 0.69 times 0.833 is 0.575, so a GeneBench task costs 42.5% less than on GPT-5.5. The truth is in between and OpenAI has the split. Note also what the table does not say. GPT-5.5 is the only model Rosalind is compared against on all three, and GPT-5.5 is four months old; the April post's LABBench2 comparison was against GPT-5.4. Nobody has put Rosalind next to Sol, Astra or Opus 5 on a life-sciences task in print, and since the model is gated, nobody outside the approved list can.
One worked turn, since a rate card is only as real as the request you put through it. Take 20,000 input tokens of literature and assay data and 5,000 output tokens of reasoning and answer, uncached. GPT-5.5: $0.10 plus $0.15, $0.25. Rosalind: $0.10 plus $0.125, $0.225. Opus 5: $0.225, the same, until the next turn caches the 20,000 and Opus charges $0.125 to write them where Rosalind charges $0.10 to read them cold. Sol: $0.18. Astra: $0.45. Cyber: $0.625. A team running a thousand of those turns a month is looking at $225 on Rosalind from October 5, and $0 until then.
Six output prices, cheapest to dearest
#1
$20
GPT-5.6 Sol (promo)
#2
$25
GPT-Rosalind
#3
$25
Claude Opus 5
#4
$30
GPT-5.5
#5
$50
GPT-6 Astra
#6
$75
GPT-5.6 Cyber
Per 1M output tokens, Standard tier, short context. OpenAI and Anthropic pricing pages read September 15, 2026. Sol is on a promotional rate; its pre-August 21 price was $30.00, level with GPT-5.5.
The bar that matters is not Rosalind's. It is the gap between Rosalind and Cyber, two models with the same access model, the same single-tier pricing, and a threefold difference in what OpenAI thinks vetted customers will pay. Biology got a discount to the base. Security got a surcharge. We are not going to guess at the reason; we will note that the cyber model picked up a $1 billion subsidised-access programme three weeks after launch and the life-sciences one launched with McKinsey, BCG and Bain named as "advisory partners" in the April post, which suggests two very different ideas about who the buyer is.
172 days before the first invoice
The dates, in order. April 16: research preview, in ChatGPT, Codex and the API, US enterprises only, no price. June 3: the GPT-5.5-based update, access expanded to eligible organisations globally, Novo Nordisk named, still no price. September 8: the API changelog says gpt-rosalind-research "is now generally available through the trusted-access program for approved internal life sciences research." September 11: an italic paragraph is added to the top of both blog posts saying the same thing and that "published pricing will take effect on October 5, 2026." October 5: billing begins. That is 172 days from the first shipped version to the first billable token, 124 from the current version, and 27 from the changelog entry that called it GA.
What we cannot do is price the free period, because nothing in any of the documents says how many organisations are approved or how many tokens they ran. The April post names Amgen, Moderna, the Allen Institute and Thermo Fisher Scientific and Los Alamos; the June post adds Novo Nordisk. Six names and no volume. It would not be a large number by Anthropic-report standards, since this is a gated model for internal research, but it is not nothing either, and the trusted-access structure means OpenAI knows the figure to the token and has no reason to publish it.
There is a second price sheet, and it disagrees with the first in a way worth a paragraph. The ChatGPT rate card for Enterprise credits lists GPT-Rosalind-Research at 125 credits per million input, 12.50 cached and 625 output, which at the 25-credits-to-the-dollar rate implied by Sol's 100 / 10 / 500 is exactly $5 / $0.50 / $25, consistent with the API. Then the same page prices Chat messages by the message: Sol 10 credits, Rosalind 20. So a Rosalind message costs twice a Sol message in Chat while a Rosalind token costs 1.25x a Sol token everywhere else. Either a Rosalind chat message uses 1.6x the tokens, which for a model built to reason longer is plausible, or the per-message number was set without reference to the per-token one. OpenAI does not say.
What $5 buys is not written down
Every other model on OpenAI's models page has a card that opens a spec sheet: context window, max output, knowledge cutoff, rate limits by tier, supported endpoints, snapshots. Rosalind's card is there, with the text "Life sciences reasoning for approved organizations," and it links to the pricing page. The URL that every other model gets, /api/docs/models/{id}, returns a 404 for gpt-rosalind-research. So we do not know the context window of a model whose base has a 1,050,000-token window with a surcharge at 272,000, and the pricing row has no long-context column to hint at either. We do not know the max output for a model sold on "longer reasoning for complex tasks." We do not know whether it takes reasoning.effort, whether it runs on Chat Completions or only Responses, or what a Tier 1 organisation's TPM is. For a model whose whole pitch in the help centre is omics data you point it at rather than paste in, the context window is the one number that would tell a bioinformatician whether the price is relevant, and it is the one number missing.
The access terms are written down, and they narrow the price further. API use is "for approved internal research tools, workflows, and applications" and "not available for customer-facing products or external commercial applications at this time," which means the $25 output line is for a cost centre, never a product margin. Approval needs an Enterprise or Business agreement, or the OpenAI-managed workspace the June post introduced for organisations without one. And a detail from the setup guide that we read three times: to use the model in ChatGPT, "open the model picker, go to Legacy models, and select GPT-Rosalind." The model that went generally available on September 8 is filed under Legacy in the product it is generally available in.
For comparison, Anthropic's answer to the same market is not a model. Claude for Life Sciences, from October 2025, is the standard Claude line plus connectors to Benchling, PubMed, 10x Genomics and the rest, at the standard card, which today means Opus 5 at the same $5 / $0.50 / $25 or Sonnet 5 at $2 / $10. Anyone with a credit card can call it, for any purpose, with a published context window. OpenAI's answer is a separate model at Opus's price that only approved organisations can call, only internally, with no spec sheet. Whether the biology is worth the gate is a question the trusted-access list can answer and we cannot.
Three smaller gaps sit behind the big one. The footnote promises access to "the latest GPT-Rosalind models as they're released" and says nothing about whether the $5 / $25 card comes with them. Batch and Flex, which halve GPT-5.5, have no Rosalind row, and Cyber has not gained one in five weeks of being in the same position. And the input-output split behind "31% fewer tokens" is the difference between a GeneBench task costing 31% less and 42.5% less, and only OpenAI has it. The Hacker News thread on the April launch had 102 points and 30 comments; we could not find a thread on the September pricing at all, which is what you would expect for a model that most of the people who read Hacker News cannot call. That does not make the number less real. From October 5 it is the price of biology at OpenAI, and it is the same as the price of everything at Anthropic.
The pages we read
- OpenAI: API pricing - gpt-rosalind-research at $5.00 / $0.50 / $25.00 under Specialized models; the October 5 billing note and the cache-write sentence; Sol at $4.00 / $0.40 / $5.00 write / $20.00 with the November 21 promo floor; Astra at $10.00 / $1.00 / $12.50 / $50.00; Cyber at $12.50 / $1.25 / $15.625 / $75.00; GPT-5.5 at $5.00 / $0.50 / $30.00 with no cache-write line. Read September 15, 2026
- OpenAI: API changelog - September 8, 2026: "GPT-Rosalind (gpt-rosalind-research) is now generally available through the trusted-access program for approved internal life sciences research. Standard pricing is $5 per 1M input tokens, $0.50 per 1M cached input tokens, and $25 per 1M output tokens. Billing begins on October 5, 2026."
- OpenAI: Introducing GPT-Rosalind - April 16, 2026 by the Hacker News timestamp, with the September 11 update paragraph at the top. Amgen, Moderna, Allen Institute, Thermo Fisher; BixBench and LABBench2 vs GPT-5.4; Dyno Therapeutics 95th and 84th percentile; McKinsey, BCG and Bain as advisory partners
- OpenAI: Introducing new capabilities to GPT-Rosalind - June 3, 2026. "Combines GPT-5.5's agentic coding and tool-use capabilities"; MedChemBench 27.5% vs 25.1% with 7.2% fewer tokens; GeneBench 21.6% vs 20.4% with 31% fewer; LabWorkBench 63.2% vs 55.8% with 5.3% fewer; Novo Nordisk; the managed workspace for organisations without an Enterprise account
- OpenAI Help Center: GPT-Rosalind for life sciences research - "Published pricing will take effect on October 5, 2026"; Enterprise or Business agreements; "not available for customer-facing products or external commercial applications at this time"; the Legacy models instruction
- OpenAI Help Center: ChatGPT rate card - GPT-Rosalind-Research at 20 credits per Chat message against Sol's 10; 125 / 12.50 / 625 credits per million tokens against GPT-5.5's 125 / 12.50 / 750 and Sol's 100 / 10 / 500
- OpenAI: Models - the Rosalind card links to the pricing page; /api/docs/models/gpt-rosalind-research returned 404 on September 15
- Anthropic: Pricing - Claude Opus 5 at $5.00 input, $6.25 cache write, $0.50 cache read, $25.00 output, batch half; Mythos 5.1 at Fable 5.1's $10 / $50
- Anthropic: Claude for Life Sciences - October 20, 2025. Connectors to Benchling, BioRender, PubMed, 10x Genomics and others on the standard models, no separate price
- Hacker News: GPT-Rosalind for life sciences research - posted April 16, 2026; 102 points, 30 comments