Hippocratic AI splits its Polaris 6 voice model into four tiers to reach admin work

Hippocratic AI released Polaris 6 Ultra, the newest version of the language model behind its patient-facing voice agents, on 2 November 2026, and split the product into four tiers with new flexible pricing¹. The Menlo Park company, which has raised $444 million in total, is pitching the release at health systems that want to match the cost and speed of an AI agent to the job it does¹.

What changed

Polaris 6.0 now comes in four versions¹:

• Ultra, the full set of all 31 supervisors, for clinical conversations with patients¹;
• Pro, for sub-clinical workloads¹;
• Flash, for high-volume sub-clinical work¹;
• Nano, for administrative tasks¹.

The company says the system has grown to 7.4 trillion parameters, from 5.0 trillion in Polaris 5¹. A 671B primary model sits inside a group of 30 specialist supervisors and 142 production monitors that check each turn of a conversation in parallel, covering areas such as medications, consent, privacy and crisis escalation¹. Despite the larger system, Hippocratic AI reports a 1.3 times end-to-end speed gain over Polaris 5, and says the model reaches its first sentence in 0.28 seconds, against 0.71 to 0.94 seconds for the four fastest commercial providers on identical prompts¹.

The release also adds a planning step that runs while the previous sentence is still being spoken, so the agent can judge a patient's mood before choosing what to say next¹. Every plan is logged in a Safety Portal next to the transcript for clinicians to audit¹. On 2,778 clinician-reviewed production calls, the company says misheard medications or symptoms fell from 1.19 to 0.20 per 100 calls, and words invented from background noise fell from 24% to 0.6%¹. Its in-house voice model pronounced 90% of 629 tested drug names correctly, compared with 75% for the leading commercial text-to-speech product¹.

Chief executive Munjal Shah said Polaris 6.0 "was built from more than 323 million patient conversations"¹. The company separately states that its agents have completed more than 250 million patient voice interactions across more than 300 live clinical use cases¹.

What it means for the market

The tiering is the commercial news. By offering lighter versions for sub-clinical, high-volume and administrative work, Hippocratic AI is reaching for tasks beyond clinical patient calls, while keeping the fully supervised version for those conversations¹. The company ties the change to deeper EHR integration and says Polaris can now meet the needs of any healthcare client¹. Co-founder Subhabrata Mukherjee framed the rivalry as a trade-off that rivals accept: "speech-to-speech models are fast and shallow, reasoning models are capable and slow"¹.

Hippocratic AI also says its contracts allow it to take responsibility for its agents, and that the agents do not diagnose or prescribe¹. For health system buyers weighing liability, that is a term to test in negotiation. The release gives no prices for any tier, no customer names and no revenue, so the effect on what health systems pay cannot yet be judged¹.

Caveats

The performance and safety figures are the company's own, from internal benchmarks and its own clinician review of production calls, and none has been published or checked independently¹. The benchmark gains, including an 11 point lead over the best frontier model, use Hippocratic AI's internal healthcare-intelligence measure¹. The company says the minor-harm rate in clinician output testing remains 0.00%¹.