GPT-Rosalind vs Grok 4.7: A Health Benchmark Is Not a Biology Gate

Grok 4.7 ships with a number that looks, at a glance, like the answer to GPT-Rosalind: 56.7% on HealthBench Professional, up from 48.5% for Grok 4.6. That is a clinical-knowledge benchmark, reported by SpaceXAI in the launch table for a model released on 21 September 2026, and it is the closest thing to a medical evaluation in either company's material. GPT-Rosalind, on general sale since 5 October 2026 after an April 2026 introduction, publishes nothing comparable — its numbers are chemistry and wet-lab benchmarks, and you need a trusted-access review to run them yourself. So the honest question is not which model scores higher on health; it is what a health benchmark measures versus what a biology gate protects, and why one of these models costs a third as much per token as the other while being available to anybody. Benchmark disputes and access-gate changes that follow a comparison like this one are recorded with their dates and sources in the AI Incident Archive, alongside the rest of the model and vendor incidents we follow.

The overlap that turns out to be narrow

HealthBench Professional asks a model questions of the kind a clinician would answer. It rewards recall, calibration and the ability to hold a clinical frame. It does not require the model to design a perturbation experiment, read a wet-lab protocol, or reason about a structure-activity relationship. GPT-Rosalind's published suite is the other kind of test: MedChemBench covers potency and ADME prediction and retrosynthesis; GeneBench covers long-horizon genomics; LabWorkBench covers linking perturbations to outcomes in real protocols. Different instruments, different targets.

• What is being measured — Grok 4.7: clinical knowledge and judgement, 56.7% on HealthBench Professional. GPT-Rosalind: chemistry and wet-lab reasoning, 27.5% on MedChemBench and 63.2% on LabWorkBench.

• Who ran it — Grok 4.7: SpaceXAI, against its own previous model. GPT-Rosalind: O​penAI, against GPT-5.5.

• Independent reproduction — None for either benchmark.

• What neither number tells you — Whether the model can be trusted with a result that goes into a filing. That is a question about failure modes, not scores.

Put the two tables side by side and there is no row they share. That is not a gap in the reporting; it is the shape of the market. A general model is measured on breadth, a specialist on depth, and the vendors pick the tests that show their own strength.

The envelope, which is where these two actually differ

Specifications are the least glamorous part of a comparison and the part that decides whether a workflow is possible at all. Here the two models diverge more than the benchmark conversation suggests.

• Context window — Grok 4.7: 500,000 tokens, half of GPT-Rosalind's 1,000,000.

• Max output — Grok 4.7: 450,000 tokens, and SpaceXAI's own documentation states there is no text output limit, which does not agree with the 450K figure on our model card. We flag the disagreement rather than pick a side; the vendor is the authority on its own model. GPT-Rosalind: not separately published as an output ceiling.

• Input — Grok 4.7: text and images, JPG or PNG up to 20 MiB each, no documented cap on the number of images, any order. GPT-Rosalind: scientific file inputs through ChatGPT, Codex and the API; no general image-input claim published.

• Reasoning control — Grok 4.7: reasoning_effort at low, medium, high or xhigh, with the published benchmark figures quoted at xhigh. GPT-Rosalind: a reasoning-effort-style control in the API, with tool use at the core of the model.

• Tools — Grok 4.7: function calling, web search, X search and code execution, all server-side. GPT-Rosalind: a life-sciences tool plugin with more than fifty tools inside ChatGPT and Codex.

• Knowledge cutoff — Grok 4.7: May 2026, with no realtime awareness by default. GPT-Rosalind: not published.

The context gap is the one that bites in this domain. A 1M-token window is not a luxury for genomics work; it is what lets a model hold a whole atlas or a full protocol set in one pass. Grok 4.7's 500K is generous by general standards and half of what the specialist offers. Meanwhile the 450K output ceiling is far beyond anything a biology answer needs and exactly what a long agent trajectory or a large refactor wants.

The price gap is the widest in this set

Grok 4.7 is the cheapest model in any of these comparisons and GPT-Rosalind is the most expensive, which makes the arithmetic blunt.

• Input — Grok 4.7 $2.00 per 1M vs GPT-Rosalind $5.00. The specialist costs 2.5× as much.

• Output — Grok 4.7 $6.00 vs GPT-Rosalind $25.00. A 4.2× gap, and the largest single-number difference in this article.

• Cached input — Grok 4.7 $0.50 vs GPT-Rosalind $0.50. Identical, which is the one line where the specialist is not more expensive.

• Long prompts — Grok 4.7 doubles to $4.00 / $12.00 with $1.00 cached once a prompt reaches 200K tokens. GPT-Rosalind publishes no long-context surcharge.

• Regional billing — Grok 4.7 has a US endpoint at a 1.1× multiplier, which works out to $2.20 / $6.60 below 200K.

The output ratio is where the decision actually lives. A discovery workflow produces a lot of output: candidate lists, protocol drafts, code, long rationales. At $25 against $6 per million, a run that emits ten million output tokens costs $250 on the specialist and $60 on the general model. Whether that $190 buys you anything is the only question worth asking, and no published benchmark answers it.

What the independent evaluator adds

Grok 4.7 has a real independent record, and it is more informative than the vendor's table because it includes the costs nobody advertises.

• Intelligence — 46 on the Artificial Analysis Intelligence Index, v4.3.2, which is 21st of 211 models on that board and well above the index median of 26. GPT-Rosalind is not on the board at all; specialist models are out of scope.

• Cost per index run — $3.74, with the evaluator recording a 75% cache discount and an input/output rate matching the vendor's own $2.00 / $6.00. Its comment is that the model is reasonably priced against peers of similar price.

• Verbosity — 240M output tokens across the index, against a median of 88M. Artificial Analysis calls it "very verbose", and that is the single most useful number here for budgeting: a model that thinks in longer chains costs more per task at any fixed token rate.

• Speed — 82.6 output tokens per second, 83rd of 211, which the evaluator nonetheless labels slower than average.

The verbosity figure is the one that should change your mental arithmetic. Grok 4.7 at $6.00 output is not 4.2× cheaper than GPT-Rosalind at $25.00 per task if it emits nearly three times as many tokens getting there. On a per-task basis the real gap is narrower than the rate card says, and nobody publishes a Rosalind verbosity number to compare against — because Rosalind is not on the board.

Why the gate exists, and why Grok does not have one

This is the asymmetry a benchmark table cannot express. Grok 4.7 is sold to anyone with a billing account, with a 500K context, image input and server-side code execution. GPT-Rosalind is sold only to organisations that clear a review, and the review exists because O​penAI judged the model's biology capability to be the kind of thing that should not be available on demand.

Read that as a statement about capability, not about safety theatre. A lab that gates a model has decided the model is good enough at something consequential to be worth gating. SpaceXAI has made no equivalent judgment about Grok 4.7, and its published strengths — coding, agentic tool loops, long-running software work — are not the kind of thing that gets a gate.

The practical consequence: if your work is computational and adjacent to biology rather than inside it, Grok 4.7 is available now, cheap, and well suited to the shape of the task. If your work is the biology itself, the model you want is behind a form, and no amount of HealthBench Professional closes that. The gate is the product in that case, not an obstacle to it.

Routing the pair, and the one place we have to be careful

Grok 4.7 is on OrcaRouter at SpaceXAI's list price — $2.00 input, $6.00 output and $0.50 cached input per million tokens, with 0% markup, and the 200K threshold carried through as a pricing tier so a long-prompt request is billed honestly instead of surprising you. GPT-Rosalind is not on OrcaRouter and not on any third-party router; it runs through O​penAI's trusted-access programme directly. We are not going to imply otherwise, and there is no route to configure that would change it.

What that leaves is the pattern this whole set of comparisons keeps arriving at. The computational half of a research platform runs on Grok 4.7 and the other 200+ models behind one key, with automatic failover across providers so a capacity problem moves the request rather than failing the run. The biology half stays on a direct key. The routing DSL is where you make the boundary explicit: a named route for the specialist that fails loudly when it is unavailable, a default route for everything else, and no silent promotion of a general-model answer into a specialist slot.

Model fusion is worth knowing about for the boundary cases. A panel that asks both a cheap general model and a specialist the same question, then compares, is a way to catch the request that was misclassified — the one that looked like plumbing and turned out to need domain judgement. It is also the honest way to use a model you do not fully trust yet.

What would change the answer

Three developments, none of them exotic. SpaceXAI could publish a domain evaluation with the specificity of Rosalind's chemistry suite — HealthBench Professional is a start and not close. O​penAI could extend the October academic plan to include API access, which would put free specialist capacity in exactly the labs that currently cannot afford it. Or an independent harness could put both models through the same scientific task set, which would settle a question that two vendor tables have carefully avoided.

Until one of those arrives, the split is not close. If the request is code, data or prose, Grok 4.7 is a third of the output price with a bigger tool surface and an independent score, and choosing the specialist would be an expensive way to feel careful. If the request is the science, pay the 4.2× and accept the review, because the only model in this comparison that has been tuned for it is the one that made you fill in a form.