PayerBench

A τ²-bench domain for health-plan member services: a written policy, a synthetic plan database, 15 agent tools, 3 caller tools and 104 simulated calls. An LLM plays the member (or a provider office) with a hidden goal; the agent under test follows the policy and calls tools; the call is scored on the final state of the database, on facts the agent must state, and, for 32 tasks, on written assertions checked by an LLM judge.

Code to run it: github.com/knotkeep/payerbench.

Subsets

subset	rows	one row is
tasks	104	a simulated call: persona, what the caller knows, instructions, reference actions, required facts, judge assertions
results	1,080	one simulation from the leaderboard runs, with its reward, failure reasons and full transcript
leaderboard	8	pass^1 to pass^4 for one agent on one tier

Plain files: db.json (the starting database), policy.md (the agent's policy), split_tasks.json (smoke 33, core 73, hard 31, full 104).

Nested fields (reference_actions, initial_state, task_json, messages) are JSON strings. task_json is the task exactly as τ²-bench loads it.

Leaderboard (October 2026, 4 trials per task)

pass^k is the share of tasks solved in all k of k attempts.

agent	tier	tasks	pass^1	pass^4	agent cost
Claude Haiku 5.5	hard	31	0.992	0.968	$0.62
Claude Sonnet 5.5	hard	31	0.984	0.935	$12.10
Gemma 4 31B 8-bit (local)	full	104	0.925	0.875	$0
Gemma 4 31B 8-bit (local)	hard	31	0.798	0.774	$0
Qwen3.8-Flash-Next 4-bit (local)	full	104	0.858	0.673	$0
Qwen3.8-Flash-Next 4-bit (local)	hard	31	0.887	0.710	$0

Agents ran as τ²'s standard llm_agent. Every agent faced the same simulated caller (Gemma 4 31B 8-bit, temperature 0.7) and the same judge (Gemma 4 31B 8-bit, temperature 0). Local models were served by oMLX on a Mac Studio; Claude models ran over the Anthropic API with default settings.

Scoring

• Database: τ² replays the task's reference actions on a fresh copy of the database and compares hashes with the agent's final database. Identity verification is a recorded write, so verifying the wrong person fails the task. Free text is stored as codes and addresses and phones are normalized, so the hash measures what was done, not how it was worded.
• Required facts: each string in communicate_info must appear in an agent message (case and commas ignored). Only language-neutral tokens are used: amounts as digits, plan names, IDs.
• Written assertions: an LLM judge checks nl_assertions where neither of the above can see the failure (a refusal, a disclosure, a false approval).

Caveats

• Saturated for frontier models: v1 does not separate Claude Haiku 5.5 from Sonnet 5.5.
• One simulated caller and one judge, both Gemma 4 31B. Results with a different caller or judge are a new baseline.
• The policy and tasks were written by the benchmark's author; the hard tier was designed after seeing where Gemma succeeded.
• Four trials per task; a single flip moves a hard-tier row by 3 to 12 points.

Data and license

All member, claim, prior-authorization and provider records are synthetic, generated with fixed seeds; no real person's data is included. MIT license. τ²-bench is MIT-licensed by Sierra Research.