Grok 4.6 Tops MedAgentBench: What 95.9% Clinical Accuracy Actually Means

The short version

Medical Sphere ran the standard MedAgentBench suite — 300 cases, 3 runs — and published the leaderboard on August 17, 2026. Grok 4.6 came out on top at 95.9% clinical agent task accuracy.

The rest of the field:

| Model | Clinical agent task accuracy |

| --- | --- |

| Grok 4.6 (xAI) | 95.9% |

| GPT-5.6 Sol (OpenAI) | 94.7% |

| Grok 4.5 (xAI) | 93.4% |

| Gemini 3.7 Flash (Google) | 92.7% |

| Muse Spark 1.1 (Meta) | 92.2% |

Everything in that table is inside a 3.7-point band. That is the story most coverage will miss.

What MedAgentBench actually measures

MedAgentBench is an agent benchmark, not a trivia quiz. It scores whether a model can carry out clinical tasks end to end — reading a chart, calling the right tool, pulling the right record, and producing an action — rather than whether it can recall a fact about a drug interaction.

That is a meaningfully harder test than the old multiple-choice medical exams, and a meaningfully better proxy for what an AI agent does inside a real clinical workflow.

It is still not a proxy for safety, liability, or bedside judgment.

Why the 3.7-point spread matters more than first place

When five frontier models cluster between 92.2% and 95.9%, model choice stops being the deciding variable in whether your deployment works.

At 300 cases, a 1.2-point gap between first and second is roughly three or four cases. Run the suite again next month with a new checkpoint and the order can shuffle. Treat the leaderboard as "these models are all in the same tier," not "one of them is uniquely qualified."

What actually decides outcomes at this point:

The 4.1% is the design brief

A 95.9% score means roughly one in 25 clinical agent tasks is wrong. In a marketing workflow, that is noise. In a clinical workflow, that is the entire reason you need a human in the loop and a confidence threshold that routes uncertain cases to a person.

The right framing for any regulated deployment is not "the model is accurate enough to run alone." It is "the model is accurate enough that a human reviewing its output is faster than a human doing the work from scratch." That is a real, large ROI — and it is a very different claim.

What this means if you are deploying agents

Whether or not you work in healthcare, the pattern generalizes.

1. Pick a benchmark that matches your task shape. Agent benchmarks beat knowledge benchmarks when you are building agents.

2. Assume the leaderboard rotates. Build a model-agnostic layer so swapping Grok for GPT for Gemini is a config change, not a rewrite.

3. Instrument your own eval set. Twenty of your real tasks, scored weekly, will tell you more than any public leaderboard.

4. Design for the failure rate, not the success rate. Where does a wrong answer go, who sees it, and how fast can it be undone?

5. Keep regulated data on rails. Access scoping, logging, and retention rules come before capability.

Where 88 Labs sits on this

We build agents for businesses that have to be right, not just fast. That means the same thing every time: narrow the scope, wire the tools properly, define the escalation path, and log everything. The model on top of that stack is replaceable — and this benchmark is a good reminder of exactly how replaceable.

If you want to see what a scoped, audited agent looks like on your own workflow, we build the first one as a demo.

FAQ

What is MedAgentBench?

An agent-oriented clinical benchmark that scores whether a model can complete real clinical tasks end to end, rather than answer medical exam questions. This run used the standard suite: 300 cases across 3 runs.

Which model scored highest?

Grok 4.6 at 95.9% clinical agent task accuracy, per Medical Sphere's independent eval published August 17, 2026.

How big is the gap to second place?

1.2 points over GPT-5.6 Sol (94.7%). At 300 cases, that is a handful of cases — close enough that ordering can change between checkpoints.

Does a 95.9% score mean an AI can practice medicine?

No. The benchmark measures task completion, not safety, liability, or clinical judgment. Roughly one task in 25 is still wrong, which is why human review and escalation rules are mandatory in any real deployment.

Should I choose my agent model based on this leaderboard?

Use it to confirm a model is in the top tier, then decide on tool integration, latency, cost, and your own task-specific evals. All five models here are close enough that the surrounding system matters more.