Generated title card headlined 'GPT-6.2 Astra Tops Hospital-Bench' with the subtitle '62.4% mean score, 14 of 48 tasks completed cleanly' and three evidence chips reading 'Top of 20 models', '320-bed hospital, 13 systems, 131 tools' and 'Claude Opus 5.5 leads full passes at 35.4%', over a footer noting the result was published 20 November 2026 by a YC-backed evaluator and that GPT-6.2 Astra shipped 17 November 2026.
Guides & Insights

GPT-6.2 Astra Tops Hospital-Bench at 62.4%: 14 of 48 Tasks Clean

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

A twenty-model evaluation posted to X on 20 November 2026 puts GPT-6.2 Astra first on Hospital-bench with a mean score of 62.4% and a full-pass rate of 29.2% — fourteen of the forty-eight tasks completed cleanly. Claude Fable 5.2 is second at 60.3%, G​emini 4 Neon third at 59.8%, and Claude Opus 5.5 fifth on score at 55.6% while leading the entire field on full passes at 35.4%, which is seventeen of forty-eight. Gemini 4 Argon finishes ninth at 46.9% mean and 16.7% pass. Those are the numbers. What the thread does not hand you is the thing that decides whether they mean anything: the name of the startup that ran the evaluation, and a methodology page you can read.

That gap is worth naming before anything else, because it is the difference between a leaderboard and a benchmark. The post describes a Y Combinator-backed company and links only to its own attached graphic; the shortener resolves back to the image, not to a paper, a GitHub repository or a results page. There is no task set to inspect, no scoring code to read, no human baseline to compare against. Everything below is therefore a read of one published table and the sentence printed above it, and it should be quoted that way.

None of that makes the result a launch. O​penAI shipped GPT-6.2 Astra on 17 November 2026, three days before the thread went up, and this is another independent board the model has appeared on since. What is new on 20 November is the board itself — Hospital-bench had not surfaced publicly before, and a twenty-model field with a published top ten is a large amount of signal to arrive in one image. The reason it deserves a careful read rather than a victory lap is that the two columns on it disagree about who won, and the disagreement is the most useful thing on the page.

What Hospital-bench actually measures

The methodology line in the graphic is a single sentence: "Measuring how well AI agents run the administrative operations of a 320-bed hospital across 13 connected systems." The thread fills in a little more. Agents are handed 131 tools spread across those thirteen systems and forty-eight tasks drawn from the back office of a hospital rather than its clinical side — prior authorisations, denied claims, records requests. No diagnosis, no triage, no imaging. The work being measured is the work that generates the paperwork a hospital runs on and the revenue-cycle fights it loses.

That choice of domain does more to explain the shape of the board than any claim about general intelligence. Hospital administration is a tool-selection problem before it is a reasoning problem. A prior authorisation is a lookup against a payer's formulary rules, a form, a submission, a follow-up and an appeal, and each of those steps lives in a different system with a different interface. With 131 tools available and forty-eight multi-step tasks, an agent has to retrieve the right subset, keep the case state straight across turns, and avoid the specific failure that defines this work: doing eleven steps correctly and then submitting to the wrong queue. An eval built like that will separate models on reliability long before it separates them on cleverness, which is exactly what the two columns show.

It also means the scores are not portable to anything else. A model that leads Hospital-bench has demonstrated something about tool-grounded administrative workflows at scale, and nothing at all about whether it can write code, pass a legal exam or hold a conversation. Read the board as a statement about agents operating a large tool surface under procedural rules. That is a real and under-served category of evaluation, and it is narrower than "runs a hospital."

The board as posted, top to bottom

Here is the full published top ten, mean score first and full-pass rate second:

• GPT-6.2 Astra — 62.4% mean, 29.2% pass (14 of 48)

• Claude Fable 5.2 — 60.3% mean, 29.2% pass (14 of 48)

• G​emini 4 Neon — 59.8% mean, 31.3% pass (15 of 48)

• GPT-6.2 Sol — 57.6% mean, 25.0% pass (12 of 48)

• Claude Opus 5.5 — 55.6% mean, 35.4% pass (17 of 48)

• Muse Wonder — 53.6% mean, 27.1% pass (13 of 48)

• GPT-6.1 Sol — 50.0% mean, 12.5% pass (6 of 48)

• Claude Sonnet 5.5 — 47.7% mean, 29.2% pass (14 of 48)

• Gemini 4 Argon — 46.9% mean, 16.7% pass (8 of 48)

• GPT-6 Astra — 46.7% mean, 8.3% pass (4 of 48)

The Hospital-bench leaderboard graphic attached to the 20 November 2026 X thread, headed 'Hospital-bench' with the methodology line 'Measuring how well AI agents run the administrative operations of a 320-bed hospital across 13 connected systems', labelled 'Top 10 of 20 models' with columns 'Mean score' and 'Pass rate'. Rows read GPT-6.2 Astra 62.4% / 29.2%, Claude Fable 5.2 60.3% / 29.2%, Gemini 4 Neon 59.8% / 31.3%, GPT-6.2 Sol 57.6% / 25.0%, Claude Opus 5.5 55.6% / 35.4%, Muse Wonder 53.6% / 27.1%, GPT-6.1 Sol 50.0% / 12.5%, Claude Sonnet 5.5 47.7% / 29.2%, Gemini 4 Argon 46.9% / 16.7% and GPT-6 Astra 46.7% / 8.3%, on an axis running 0% to 100%.

Two structural facts about that list before any interpretation. The first is that the field is twenty models deep and only ten are shown, so the bottom half is invisible — and in a domain where the pass rate is the interesting column, the models that fail everything are the ones that would tell you how hard the floor is. The second is that the whole spread from first to tenth is 15.7 points of mean score. On a benchmark with forty-eight tasks and no published variance, that is not a wide top. It is a cluster with a leader in it.

The generational ordering is the cleanest signal in the table and the one least likely to be an artifact. GPT-6.2 Astra at 62.4% sits 15.7 points above GPT-6 Astra at 46.7%, and the pass rate goes from 8.3% to 29.2% — a 3.5x jump in tasks completed cleanly. GPT-6.2 Sol at 57.6% sits 7.6 points above GPT-6.1 Sol at 50.0%, with its pass rate doubling from 12.5% to 25.0%. Whatever the benchmark is measuring, it is measuring something that moved between generations, in both O​penAI lines, in the same direction, at a magnitude well outside a rounding error.

Why the pass-rate column matters more

The mean score is the average across all forty-eight tasks with partial credit. The pass rate is the share of tasks the agent finished completely. Those are different measurements of different things, and the gap between them on this board is enormous: GPT-6.2 Astra averages 62.4% but completes 29.2% of tasks, which means the average task is being scored around two-thirds correct while fewer than one in three is being finished. That is the signature of an agent that gets most of a workflow right and then loses it at a step.

In most benchmarks the distinction is academic. In hospital administration it is the whole product. A prior authorisation submitted with the right codes to the wrong payer queue scores well on partial credit and produces a denied claim. A records request that assembles the correct document set and misses the disclosure-log entry is a compliance finding. The tasks in this domain are built so that the last step is the one that matters, which means the pass-rate column is closer to a measure of whether the agent can be trusted with the workflow than the mean score is, and it is the column almost nobody quoted.

It is also the column with the least statistical support. Forty-eight tasks is a small denominator. One task is 2.08 percentage points of pass rate, so the difference between Claude Fable 5.2 and GPT-6.2 Astra — both at 29.2%, both at fourteen tasks — is zero, and the difference between G​emini 4 Neon at fifteen tasks and the two leaders at fourteen is one task. Three models are separated by a single task out of forty-eight. Treat the top three as tied on completion and separated only on the mean, where the gaps are 2.1 and 0.5 points and equally fragile.

Opus 5.5 finishes first where it counts

Claude Opus 5.5 is the model the table was built to expose. It ranks fifth on mean score at 55.6% — 6.8 points behind the leader, which is not close — and first on full passes at 35.4%, seventeen of forty-eight, three tasks clear of the next model. No other model on the board inverts its rank like that. Claude Sonnet 5.5 does it mildly, eighth on score and tied third on passes. Gemini 4 Argon does the reverse, ninth on score and ninth on passes, which is at least consistent.

Three tasks is not a large margin, and it is worth saying so plainly: seventeen against fourteen out of forty-eight is a difference a single ambiguous task could erase, and no confidence interval is published. But the direction is consistent across the two A​nthropic models on the board, and it points at a real trade. Opus 5.5 appears to be spending more turns and more tool calls per task to reach a completed state, which raises its completion rate and caps its average. Astra reaches a higher average with fewer clean finishes. If you are buying an agent for a workflow where a half-completed claim is worse than a slow one, the fifth-place model is the better purchase, and no amount of mean-score ranking will tell you that.

There is a second reading, less flattering to the benchmark. A model can raise its pass rate by declining to attempt tasks it is likely to fail, or by terminating early with a partial answer the scorer counts as complete. Without the task-level output — which the thread does not publish — there is no way to distinguish "finishes more of what it starts" from "finishes less of what it is given." The table supports the first reading. It cannot rule out the second.

What 131 tools cost per task

Hospital-bench is an expensive evaluation to run and an expensive one to run repeatedly, and the thread publishes no cost figure at all. A hundred and thirty-one tool definitions is a large system prompt before the agent takes a single action; forty-eight multi-step tasks across twenty models is a lot of input tokens and a lot of turns. Any per-task cost you compute from token prices will be a floor, not an estimate, because it will miss the retrieval, the retries and the failed tool calls that an agentic loop generates on its own.

• Mean score — GPT-6.2 Astra 62.4% vs Gemini 4 Argon 46.9%

• Full-pass rate — Astra 29.2% vs Argon 16.7%

• Input price per 1M — Argon $2.00 vs Astra $8.00

• Output price per 1M — Argon $10.00 vs Astra $40.00

• Tools in scope — 131, identical for every model on the board

• Cost per task — not published by the evaluator

A generated three-panel comparison card titled 'What the lead costs', with panels reading 'Mean score: Astra 62.4% vs Argon 46.9%', 'Full passes: Astra 14 of 48 vs Argon 8 of 48' and 'Price per 1M: Astra $8 in / $40 out vs Argon $2 in / $10 out', over a footer noting the evaluator publishes no cost per completed task.

Read those two blocks against each other and the interesting number is not the 15.5-point mean gap. It is that Argon costs a quarter of Astra in both directions and completes a little over half as many tasks cleanly, which puts the two models within a factor of two of each other on cost per completed task despite a 4x price difference. That is a much closer call than the leaderboard suggests, and it is the kind of calculation you can only do on your own workload, because the ratio depends entirely on how often your tasks terminate early.

This is where a routing layer earns its place rather than being decoration. OrcaRouter passes provider list price through at 0% markup, so the $8 and $40 above are what you pay per million tokens rather than a vendor rate with a platform margin stacked on top — and because the pass-through is live, a vendor price cut on either model shows up on your bill the same day it is announced. Automatic failover matters more here than on a chat workload: a forty-eight-task agentic campaign against 131 tools will hit a rate limit somewhere, and a route that moves to the second model instead of dying mid-run is worth more than a point of mean score.

Worth saying plainly: GPT-6.2 Astra is not on our catalogue as of 20 November. The O​penAI models we serve are GPT-6 Astra, which scores 46.7% and 8.3% on this board, and the GPT-6 Sol line, where GPT-6.2 Sol posts 57.6% and 25.0%. Gemini 4 Argon is routable at $2 and $10. Until Astra is available here, the model on our catalogue closest to the top of this table is GPT-6.2 Sol, and the cheapest model that beats the previous O​penAI generation is Argon.

What would change this read

Four things, roughly in order of how much they would move the picture. The first is a methodology page. A benchmark that publishes its task set, its scoring rules and its run logs can be argued with; one that publishes a top-ten image cannot. The single sentence above the table is doing an enormous amount of work, and every claim in this article is bounded by it.

The second is the missing ten models. Twenty were evaluated and ten were shown, and in a domain where partial credit inflates the mean, the models at the bottom are the ones that calibrate the task. If the eleventh-place model is at 45% mean and 6% pass, the board is a spread. If it is at 44% and 2%, the board is a cliff and the top ten are the only models that can do the work at all.

The third is a human baseline. Hospital-bench publishes no human number, and without one there is no answer to the only question a buyer actually has, which is whether 35.4% completion is good. A medical billing specialist with the same 131 tools and forty-eight tasks would give the table a scale. It has none.

The fourth is a second run. The model is four days old and the evaluation is a single campaign, and agentic benchmarks with long tool chains have more run-to-run variance than single-shot evals — which is exactly the variance that a 2.1-point gap between first and second cannot absorb. If Astra and Fable 5.2 land in the same order three times, the lead is real. Once, it is a leaderboard entry.

What survives all four caveats is narrower than the thread's framing and more useful than a dismissal. GPT-6.2 Astra is the best model published on a twenty-model agentic administration benchmark, by a margin that is inside the noise on completion and outside it on average. The generational jump is the part that is hard to explain away — 15.7 points of mean score and a 3.5x improvement in clean completions over GPT-6 Astra, on a board with no reason to favour one vendor. That is a real result about tool-grounded administrative work, published by an evaluator nobody can yet check, four days after the model shipped. Both halves of that sentence belong in the same breath.

A generated single-column scoreboard titled 'Hospital-bench, top of the board' with six rows reading 'GPT-6.2 Astra 62.4% mean / 29.2% pass', 'Claude Fable 5.2 60.3% / 29.2%', 'Gemini 4 Neon 59.8% / 31.3%', 'GPT-6.2 Sol 57.6% / 25.0%', 'Claude Opus 5.5 55.6% / 35.4%' and 'Gemini 4 Argon 46.9% / 16.7%', with column headers 'Mean score' and 'Pass rate' and the footer '48 tasks, 20 models evaluated, top 10 published'.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily