gpt-rosalind-discovery is the name of a model OpenAI has not announced, and it now sits on the company’s own API price list. The row reads $5.00 per million input tokens, $0.50 per million cached input tokens and $25.00 per million output tokens, which is exactly what OpenAI charges for gpt-rosalind-research, the life sciences model it took out of research preview in September. There is no blog post, no changelog line, no model page and no SDK entry. There is only the row.
We went past the screenshots that circulated on 10 October and read the primary record instead: the live pricing page and its Markdown twin, the last Wayback Machine capture before the change, the pull request that put the new id into a widely used open-source cost map, OpenAI’s SDK source, the developer changelog, the ChatGPT credit table, the Help Center article and the Rosalind product page. Teams that budget for AI models line by line will find the timing more useful than the screenshot. The row went live on the evening of 9 October, about ten hours before anyone posted about it.
We have covered this model family twice. The April launch is in GPT-Rosalind: 7 Critical Facts About OpenAI’s New Life Sciences Research Model. This article sets out what gpt-rosalind-discovery is on the record, what it is not yet, and how a research team should treat a price for a model nobody has described.
Table of contents
- What the Pricing Page Shows for gpt-rosalind-discovery
- When gpt-rosalind-discovery Appeared: An 85-Minute Window
- What OpenAI Has Not Published About gpt-rosalind-discovery
- The GPT-Rosalind Story So Far
- Reading the Price of gpt-rosalind-discovery
- What gpt-rosalind-discovery Could Be: Four Readings
- What Life Sciences Teams Should Do Now
- gpt-rosalind-discovery FAQ
- References and Further Reading
What the Pricing Page Shows for gpt-rosalind-discovery
The evidence is one table and one footnote on developers.openai.com. Both are worth quoting exactly, because most of the coverage has paraphrased them, and the paraphrases have quietly added things the page does not say.
The two rows, word for word
The table sits under the heading “Life sciences models”, with the subheading “Our latest Rosalind models” and the note “Prices per 1M tokens”. It has three price columns and no cache-write column.
| Model | Input per 1M | Cached input per 1M | Output per 1M |
|---|---|---|---|
| gpt-rosalind-research | $5.00 | $0.50 | $25.00 |
| gpt-rosalind-discovery | $5.00 | $0.50 | $25.00 |
The two rows match on all three columns. That is the whole of what OpenAI has published about gpt-rosalind-discovery: a name, an order on the page (second, beneath the research model) and three prices.
The footnote that names only one model
Beneath the table is the paragraph that used to sit under gpt-rosalind-research alone, unchanged. “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.” It goes on: “All eligible organizations will continue to get access to the latest GPT-Rosalind models as they’re released.” The final sentence is the standard 10 percent uplift for regional processing endpoints on models released on or after 5 March 2026.
Read literally, the billing date and the cache-write exemption apply to gpt-rosalind-research only. Nothing on the page says when gpt-rosalind-discovery would start billing, whether cache writes are free for it too, or whether the trusted-access wording covers it. The footnote’s grammar is singular (“this model”) because nobody rewrote it when the second row arrived.
What else changed in the same edit
The new row did not arrive on its own. Until 9 October the research model lived in a “Specialized models” table, with a “Category” column that filed it as “Life Sciences” after chat-latest (ChatGPT) and gpt-5.3-codex (Codex). After the edit the Specialized table holds only those two models. Rosalind has its own top-level section further up the page, directly after one headed “Cyber models” with the subheading “Our latest Daybreak models”.
The two new headings share a template: a domain name, then “Our latest” followed by a brand. OpenAI rebuilt the page so that its two trusted-access model families present the same way. The developer models catalogue has not caught up. It still lists a single “GPT-Rosalind” entry, “Life sciences reasoning for approved organizations. Model ID: gpt-rosalind-research”, linked to a “#specialized-models” anchor that no longer contains it.
| Page element | 9 Oct, 18:12 UTC capture | Live page, 10 Oct |
|---|---|---|
| Where Rosalind sits | “Specialized models” table, category “Life Sciences” | Own section, “Life sciences models” |
| Subheading | “Prices per 1M tokens.” | “Our latest Rosalind models.” |
| Rosalind rows | 1 (research) | 2 (research, discovery) |
| Specialized table rows | 3 | 2 |
| Footnote | Names gpt-rosalind-research | Identical text |
| “Cyber models” section | Present | Present, unchanged |
When gpt-rosalind-discovery Appeared: An 85-Minute Window
The coverage dates this story to 10 October. The page changed the evening before, and two independent machine records pin it to a window of less than an hour and a half.
The timeline, to the minute
The Internet Archive captured the pricing page at 18:12 UTC on 9 October, and that capture has the old layout with one Rosalind row. At 19:37 UTC a bot account that keeps LiteLLM’s model cost map in sync with provider price lists opened pull request 45653, titled “add openai gpt-rosalind-discovery from the pricing page”. So the row went live somewhere between those two times.
| Time (UTC) | Event | Record |
|---|---|---|
| 9 Oct, 18:12 | Pricing page captured with one Rosalind row | Wayback Machine |
| 9 Oct, 19:37 | Sync bot opens a pull request adding the new id | LiteLLM PR 45653 |
| 9 Oct, 19:51 | Pull request merged into the cost map | LiteLLM commit bbc8d66 |
| 9 Oct, 20:44 | openai-python 3.28.0 released, model list unchanged | openai/openai-python |
| 9 Oct, 22:02 | Price tracker logs it as a “new_model” | LLMTracker changelog |
| 10 Oct, 05:44 | First public post, with prices | @aitrackerbot on X |
| 10 Oct, 06:20 | TestingCatalog quote-post with a screenshot | @testingcatalog on X |
| 10 Oct, 07:46 to 07:49 | First written news items | PANews, TokenPost |
| 10 Oct, 11:39 | First long analysis | OrcaRouter blog |
Machines saw it first
Every early sighting of gpt-rosalind-discovery came from automation. LiteLLM’s bot diffs provider price pages and opens pull requests when a row appears. Its description says it fetched the raw Markdown version of the page “with a cache buster on 2026-10-09”, and it documents each field against the cell it came from. LLMTracker, a GitHub project that rebuilds its data from LiteLLM and OpenRouter every few hours, logged the id at 22:02 UTC. The account that broke the news publicly, @aitrackerbot, describes itself as an “Automated AI model release tracker”.
The bot’s pull request also recorded the most useful technical detail anyone has published about gpt-rosalind-discovery. “The list API does not emit it (nor gpt-rosalind-research), since both sit behind the trusted-access program.” In other words, the public models endpoint shows neither Rosalind id to an ordinary API key, so a price page diff was the only way an outsider could have spotted the new one.
One coincidence makes the cost map worth a look. In LiteLLM’s JSON file, the entry immediately above gpt-rosalind-discovery is the Microsoft Decision Models price row, added the same day, which we covered in Microsoft Launches Decision-1 and Google Rolls Out Gemini Thinking Levels. Both were picked up by the same kind of automated diff, hours before people wrote about either.
How long the row went unnoticed
The chart counts hours from the 19:37 UTC pull request, the earliest confirmed sighting, to each later one. The arithmetic is plain subtraction of the timestamps in the table above.
By the time this was written, TestingCatalog’s post had drawn 285 likes and 22,243 views, and the original bot post 20,607 views. TestingCatalog asked the question most readers will have: “Should it be specifically a drug discovery model?” Nothing OpenAI has published answers it yet.
What OpenAI Has Not Published About gpt-rosalind-discovery
The absences are as informative as the row. We checked every surface where OpenAI normally documents a new model, and the research model’s own rollout gives a baseline for comparison.
No model page and no changelog entry
The model pages at /api/docs/models/gpt-rosalind-discovery, /api/docs/models/gpt-rosalind-research and /api/docs/models/gpt-rosalind all return 404, so even the model that has been on sale since September has no page of its own. The developer changelog’s most recent entry is dated 8 October, and it carries no line for gpt-rosalind-discovery. The research model got one, dated 8 September: “GPT-Rosalind (gpt-rosalind-research) is now generally available through the trusted-access program.”
That changelog date sits three days before the 11 September update note that OpenAI appended to its June announcement, which says GPT-Rosalind “is coming out of research preview” and that “published pricing will take effect on October 5, 2026”. The two records disagree by three days on when general availability began. It is a small discrepancy, but it shows that OpenAI’s Rosalind documentation is updated piecemeal rather than all at once.
Not in the SDKs
OpenAI generates its official libraries from an OpenAPI specification, and the research model’s arrival there is precisely dated. On 22 September at 19:10 UTC the openai-openapi repository recorded “Add new model gpt-rosalind-research to available model list”. At 20:41 UTC the Python, Node, Go, Java and Ruby libraries each received a commit titled “add GPT-Rosalind research model”, and openai-python 3.19.0 shipped shortly after midnight.
No equivalent commit exists for gpt-rosalind-discovery. The Python library’s model type on its main branch, released as 3.28.0 at 20:44 UTC on 9 October, an hour after the price row was spotted, lists gpt-5.6-cyber and gpt-rosalind-research and nothing newer from either family. The same library’s “access_programs” parameter, described as “Domain-specific access programs to use for this request”, has exactly one key today: cyber.
Not in ChatGPT, the Help Center or the product page
ChatGPT’s credit table, where every row we checked equals the API dollar price multiplied by 25, lists “GPT-Rosalind-Research” at 125, 12.5 and 625 credits per million tokens. It has no Discovery row. The Help Center article on GPT-Rosalind reads “Updated: 29 days ago”, which puts its last revision around the September release, and describes one model in the ChatGPT picker under “Legacy models”. The openai.com Rosalind page describes Rosalind Workbench with two modes, “Explore mode” and “Research mode”, and does not use the word discovery for a product or a model.
| Surface | gpt-rosalind-research | gpt-rosalind-discovery |
|---|---|---|
| API pricing page | Listed, billing from 5 Oct | Listed, no billing date |
| Models catalogue | Listed | Absent |
| Model doc page | 404 | 404 |
| Changelog | 8 Sep entry | None |
| OpenAPI spec and SDKs | Added 22 Sep | Absent |
| Public models endpoint | Not emitted | Not emitted |
| ChatGPT credit table | 125 / 12.5 / 625 | Absent |
| Help Center article | Described | Absent |
| openai.com/rosalind | Powers “Research mode” | Absent |
What the gap says about the rollout
For the research model, the order ran announcement in April, model rebuild in June, update note and changelog in September, SDKs on 22 September and billing on 5 October. gpt-rosalind-discovery has started at the end of that sequence, with a price and nothing before it. A billing row is the piece of plumbing that invoicing, quotas and usage dashboards need in place before anyone can be charged, so seeding it early is plausible. It is also the cheapest thing to delete if plans change.
The GPT-Rosalind Story So Far
The new row only makes sense against the family it joins. OpenAI’s life sciences programme has moved quickly since April, and each step changed who could use it and on what terms.
April to October in dates
| Date (2026) | What happened |
|---|---|
| 16 Apr | GPT-Rosalind launched as a research preview for qualified US Enterprise customers |
| 29 May | Rosalind Biodefense launched for selected government and allied partners |
| 3 Jun | Model rebuilt on GPT-5.5 (“Introducing GPT-Rosalind-5.5”); access opened to eligible organisations globally |
| 28 Aug | Rosalind Workbench introduced on the developer blog |
| 8 Sep | Changelog: gpt-rosalind-research generally available |
| 11 Sep | Update note: out of research preview, pricing from 5 October |
| 22 Sep | gpt-rosalind-research added to the OpenAPI spec and all SDKs |
| 5 Oct | Billing for gpt-rosalind-research begins |
| 9 Oct | gpt-rosalind-discovery appears on the pricing page |
From the April launch to the new row is 176 days. From general availability in the changelog to the new row is 31 days, and from the first invoice date to the new row is four. Whatever gpt-rosalind-discovery turns out to be, it arrived after the family had begun charging money, not while it was still a free preview.
What the current model measured
The June rebuild is the last time OpenAI published numbers for a Rosalind model, and they set the bar any successor will be measured against. On MedChemBench GPT-Rosalind scored 27.5 percent against GPT-5.5’s 25.1 percent while using 7.2 percent fewer tokens. On GeneBench it scored 21.6 percent against 20.4 percent with 31 percent fewer tokens. On LabWorkBench, built on proprietary wet-lab protocols, it scored 63.2 percent against 55.8 percent with 5.3 percent fewer tokens.
The gains are real but modest, and most of the product page’s “per token” headline figures come from token savings rather than accuracy. That matters for gpt-rosalind-discovery. If OpenAI is preparing a second model at the same price, it will need either a different job or a clear lead on these benchmarks to justify the choice a buyer will face.
Who can use the family today
The Help Center article sets the terms. GPT-Rosalind is “available globally to eligible organizations with Enterprise or Business agreements through our trusted-access program”. API use is for “approved internal research tools, workflows, and applications” and is “not available for customer-facing products or external commercial applications at this time”. None of that has been extended to gpt-rosalind-discovery in writing, but the footnote’s promise that eligible organisations “will continue to get access to the latest GPT-Rosalind models as they’re released” suggests the same gate would apply.
Reading the Price of gpt-rosalind-discovery
Identical prices are the only hard data, so it is worth being careful about what they can and cannot support. The useful comparisons are OpenAI’s mainline models and the other trusted-access family on the same page.
Against OpenAI’s mainline models
Both Rosalind models charge the same input price as GPT-5.5, the model the family was rebuilt on in June, and $5.00 less per million output tokens, a 16.7 percent saving on output. Since then OpenAI’s mainline has moved on. GPT-6 Sol and GPT-6.1 Sol list at $2.00 in and $10.00 out, so a Rosalind call now costs 2.5 times a GPT-6 Sol call on both input and output. GPT-6 Astra, the top model, charges $10.00 and $50.00. For the speed tiers above these, see OpenAI Rolls Out GPT-6.1 Sol Ultrafast on ChatGPT Work, Codex, and API.
| API model (standard, short context) | Input per 1M | Cached input per 1M | Output per 1M |
|---|---|---|---|
| gpt-5.6-cyber | $12.50 | $1.25 | $75.00 |
| gpt-6-astra | $10.00 | $1.00 | $50.00 |
| gpt-5.5 | $5.00 | $0.50 | $30.00 |
| gpt-rosalind-research | $5.00 | $0.50 | $25.00 |
| gpt-rosalind-discovery | $5.00 | $0.50 | $25.00 |
| gpt-5.6-sol | $4.00 | $0.40 | $20.00 |
| gpt-6-sol | $2.00 | $0.20 | $10.00 |
| gpt-6.1-sol | $2.00 | $0.10 | $10.00 |
Output is where long literature syntheses and multi-step analyses spend their money, so the chart plots that column only, with gpt-5.6-cyber as the full bar.
The Daybreak precedent
The section that now sits directly above Rosalind on the pricing page is the closest thing to a template for what OpenAI may be doing. Daybreak is its trusted-access programme for cybersecurity work, and the Help Center describes tiers within it. Daybreak Blue gives reduced refusals on mainline models. Daybreak Red adds GPT-5.5-Cyber. A further step, “Daybreak Red with additional model approval”, adds GPT-5.6-Cyber, “Requires additional model-specific approval”.
Here is the detail that matters for gpt-rosalind-discovery. The Markdown version of the pricing page lists gpt-5.5-cyber at $12.50 in, $1.25 cached and $75.00 out, and gpt-5.6-cyber at exactly the same three prices, plus a $15.625 cache-write rate. So OpenAI already sells two models in one trusted-access family at identical list prices, separated by generation and by approval level rather than by cost. The Rosalind rows now have the same shape. Teams in security will know these tiers from OpenAI Launches GPT-6 Astra, Its First Model to Cross a Critical Cybersecurity Threshold.
Two pricing philosophies
The comparison also shows a difference. GPT-5.6-Cyber charges 3.125 times its base model’s input price and 3.75 times its output price ($12.50 against $4.00, $75.00 against $20.00), a premium for a riskier capability. Rosalind charges less than its base model on output. OpenAI prices its cybersecurity models up and its life sciences models down, and gpt-rosalind-discovery so far follows the life sciences pattern.
| Feature | Daybreak (cyber) | Rosalind (life sciences) |
|---|---|---|
| Pricing page heading | “Our latest Daybreak models.” | “Our latest Rosalind models.” |
| Specialised models listed | gpt-5.5-cyber, gpt-5.6-cyber | gpt-rosalind-research, gpt-rosalind-discovery |
| Same price within family | Yes, $12.50 / $1.25 / $75.00 | Yes, $5.00 / $0.50 / $25.00 |
| Price against base model | 3.75 times output | 0.83 times output |
| Access tiers documented | Standard, Blue, Red, Red plus model approval | One trusted-access programme |
| API parameter | access_programs.cyber | None yet |
| ChatGPT credit rows | Daybreak Blue, Daybreak Red | GPT-Rosalind-Research only |
A worked job
Take a target-prioritisation run that reads 200,000 tokens of papers, database records and internal notes and writes a 50,000-token report. At list prices it costs $2.25 on either Rosalind model ($1.00 in, $1.25 out). The same job costs $2.50 on GPT-5.5, $1.80 on GPT-5.6 Sol, $0.90 on GPT-6 Sol, $4.50 on GPT-6 Astra and $6.25 on GPT-5.6-Cyber. For a team choosing between the two Rosalind ids, the price will never be the deciding factor, because it is the same. For practical ways to keep that per-run figure down, see Five Keys to Controlling AI Token Costs.
What gpt-rosalind-discovery Could Be: Four Readings
None of the following is confirmed. Each reading is set against the evidence above, and each comes with the observation that would settle it. A name, three prices and a missing doc page establish that something exists and very little else.
A second access tier, Daybreak style
The pricing page now presents Rosalind exactly as it presents Daybreak, and Daybreak’s two specialised models share a price while sitting behind different approvals. On that reading gpt-rosalind-discovery would be the more capable or less restricted Rosalind model, offered only to organisations that pass a further review, much as GPT-5.6-Cyber needs “additional model-specific approval”. The SDK’s access_programs object, described as domain-specific but holding only a cyber key, has room for a life sciences equivalent. The confirming signal would be a second access level in the Help Center article or a new key in that parameter.
A model for a discovery workflow
The research model’s id matches Rosalind Workbench’s “Research mode”, which the product page describes as using GPT-Rosalind “to work on more complex biological questions”. A discovery model could serve a later stage, such as target identification, screening or candidate generation, where the work is less about synthesising literature and more about proposing molecules or experiments. OpenAI’s own copy already reaches for this language. The June post describes “core drug-discovery domains such as medicinal chemistry and genomics”, and the product page has a section headed “Choose promising biological targets”. A third Workbench mode would confirm this reading.
The next GPT-Rosalind generation
GPT-Rosalind was rebuilt on GPT-5.5 in June. OpenAI’s mainline has since moved to GPT-5.6, GPT-6 and GPT-6.1, and the Daybreak family shows OpenAI shipping a newer specialised model (GPT-5.6-Cyber) next to an older one at the same price. A Rosalind model built on a newer base could get a new suffix rather than a version number. The footnote’s “latest GPT-Rosalind models as they’re released” is consistent with that. Against it, OpenAI has so far versioned Rosalind publicly (“GPT-Rosalind-5.5”) rather than renaming it.
A staging row that never ships
The plainest reading is that someone created a billing row ahead of a launch, by copying the existing one and changing the id. That would explain the identical prices and the singular footnote. Staging rows usually precede launches, but they also precede cancellations and renames. If gpt-rosalind-discovery disappears from the table, this was plumbing. If its price ever moves away from $5.00 and $25.00, it is a different system.
| Reading | Supports it | Cuts against it | Would confirm it |
|---|---|---|---|
| Second access tier | Daybreak layout, same-price cyber pair | No second tier in the Help Center | New approval level or API key |
| Discovery workflow model | Research id matches Research mode | Workbench has two modes, not three | A third Workbench mode |
| Next generation | Base model is two generations old | Rosalind versioned as 5.5 so far | Model page naming a new base |
| Staging row | Copied prices, singular footnote | Page was rebuilt around it | Row removed or renamed |
The page rebuild is the strongest single argument against the staging-row reading. Somebody moved Rosalind into its own section and gave it a heading in the plural, “Our latest Rosalind models”, on the same day. That is more deliberate than a forgotten test row.
What Life Sciences Teams Should Do Now
For most organisations the answer is nothing urgent, because gpt-rosalind-discovery cannot be called without trusted access and may not be callable even with it. For teams already inside the programme, or applying, a few steps are worth taking this week.
If you already have trusted access
Ask your OpenAI account team directly whether gpt-rosalind-discovery is enabled for your organisation, and what it is for. Your own API key may list it even though public keys do not, so check your organisation’s model list rather than relying on screenshots. Billing for the research model started on 5 October, so reconcile your first invoice against your usage dashboard now, while the volumes are small and any surprises are cheap.
If you are applying
The eligibility terms have not changed: an Enterprise or Business agreement, approved users, and internal research use only. Write your application around the work rather than a model name. A request that describes the scientific workflow, the governance around it and the data it will touch will carry over to whichever Rosalind model OpenAI assigns. A clear AI strategy that already names owners, approval steps and evaluation criteria makes that application faster to write and easier to approve.
Do not hard-code either id
Treat Rosalind model ids as configuration, not code. The family has already changed its base model once, moved on the pricing page once and gained a second id without notice. Route calls through one setting your team can change, log which model served each response, and keep an evaluation set of your own questions so you can compare gpt-rosalind-discovery against the research model on your data the day it becomes available.
Watch the right pages
The most reliable early signals are the ones that caught this row: the pricing page’s Markdown version, LiteLLM’s cost map, the OpenAPI repository’s commit log, and the “access_programs” types in the SDKs. A model page going live, a changelog entry or a new Help Center tier would turn a price into a product. Removal of the row would end the story.
A note for UK organisations
The footnote’s 10 percent uplift for regional processing applies to “models released on or after March 5, 2026, that are eligible for data residency”. OpenAI has not said whether either Rosalind model is eligible, and its data residency guide does not name either one in its regional tables. If your research data must stay in the UK or Europe, get that answer in writing before you design a workflow around either model. That applies with extra force to cybersecurity and data protection reviews of anything touching patient or trial data.
gpt-rosalind-discovery FAQ
Is gpt-rosalind-discovery available?
Not publicly. It appears on OpenAI’s API pricing page, but it has no announcement, documentation page, changelog entry or SDK entry, and the public models endpoint does not list it. Access to the Rosalind family requires approval through OpenAI’s trusted-access programme.
How much does gpt-rosalind-discovery cost?
The pricing page lists $5.00 per million input tokens, $0.50 per million cached input tokens and $25.00 per million output tokens, identical to gpt-rosalind-research. No billing start date has been published for it.
When did gpt-rosalind-discovery first appear?
Between 18:12 and 19:37 UTC on 9 October 2026. A Wayback Machine capture at 18:12 does not show it, and an automated LiteLLM pull request citing the page was opened at 19:37. The first public post followed at 05:44 UTC on 10 October.
Is gpt-rosalind-discovery a drug discovery model?
Nobody outside OpenAI knows. The name fits OpenAI’s own drug-discovery language, but the company has not described the model’s purpose, and the Rosalind product page still lists only Explore and Research modes.
Can I use it through a gateway such as LiteLLM?
LiteLLM has added a price row so it can account for calls, but a gateway cannot grant access OpenAI has not approved. You need trusted access on your own OpenAI organisation first.
References and Further Reading
Wayback Machine capture of the pricing page, 9 October 2026, 18:12 UTC
LiteLLM pull request 45653: add openai gpt-rosalind-discovery from the pricing page
AI Tracker Bot on X, 10 October 2026
TestingCatalog on X, 10 October 2026
OrcaRouter: GPT-Rosalind Discovery, Unannounced Second Rosalind SKU
TokenPost: OpenAI Pricing Page Lists Unannounced Model
OpenAI Help Center: GPT-Rosalind for Life Sciences Research
OpenAI: Introducing GPT-Rosalind-5.5
OpenAI Help Center: Daybreak, Trusted Access for Cyber
OpenAI: Use Daybreak in the Responses API
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
