Meta’s Muse Spark beats every rival on professional healthcare tasks

Signals Inbox·July 13, 2026·MedTech

Meta says Muse Spark 1.1 now leads its HealthBench Professional comparison, a benchmark built from real tasks clinicians bring to AI. The tension is that this is not a medical approval, and the lead depends on which rivals are included: Meta compared it with GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro, and Grok 4.2, while newer public results already put GPT-5.6 Sol slightly ahead.

Q1What did Meta actually report?

In its official evaluation report, Meta says Muse Spark 1.1 performs strongly across agentic, coding, reasoning, multimodal, and health tasks. For HealthBench Professional, Meta used the 525-example benchmark and a GPT-5.4 grader, then compared Muse against a selected group of rival models.

Q2What does HealthBench Professional test?

It is not a medical-school multiple-choice exam. It tests the kind of work clinicians actually ask AI to help with: care consultation, writing and documentation, and medical research. The examples come from real clinician-style chats, and each answer is scored against detailed rubrics reviewed by several physicians.

Q3How hard is the benchmark?

The benchmark has 525 cases selected from more than 15,000 candidates. Hard cases for recent OpenAI models were made about 3.5 times more common, and roughly one-third of the set involves doctors deliberately trying to expose model failures. That makes it more meaningful than a simple medical trivia test.

Q4Did Muse Spark really beat every rival?

Not literally every model available. Meta's published comparison used Gemini 3.1 Pro, GPT-5.4, Claude Opus 4.6, and Grok 4.2. A newer public leaderboard snapshot placed GPT-5.6 Sol at 60.5% and Muse Spark 1.1 at 59.3%. So the safe version is that Muse topped Meta's comparison set, not that it holds an uncontested global lead.

Q5Why does this matter now?

Meta is no longer pitching Muse only as a cheap general model. It is showing specialist strength in a high-value workflow where hospitals, doctors, and health software companies already spend money. If that performance holds outside benchmarks, Meta could compete for clinical copilots, documentation tools, and medical research assistants while undercutting more expensive frontier models.

Q6Does this make Muse safe for patient care?

No. A benchmark win is evidence of capability, not clinical approval. It does not prove lower hallucination rates in every specialty, safe use with live patient records, or better outcomes in hospitals. The next real test is deployment: independent evaluations, specialist oversight, privacy controls, error rates, and whether clinicians trust it after months of daily use.