Official run

gpt-oss-120BHigh

In this Twenty Questions LLM benchmark, gpt-oss-120B (high) ranks 9 of 11 with a 22.23 question score and 67% success. Lower is better. The run contains 30 scored episodes across 10 subjects with 3 trials per subject. Choose a subject to inspect its attempts.

Rank 9 of 11cerebrasHigh

Question score22.23questions · lower is better95% CI 19.58–24.89

Success: 67%
Contract compliance: 89%
Guesser cost: $2.04
Guesser time: 32m 30s

Drill down

Subjects.

01Albert Einstein4.67 avg 02Albert Schweitzer22.67 avg 03Garfield10.00 avg 04Achilles7.00 avg 05Genghis Khan16.33 avg 06Bike pump41.00 avg 07Spider web35.33 avg 08Eyebrow41.00 avg 09Moon11.00 avg 10Door handle33.33 avg

Complete benchmark

Run ledger.

All Guesser and benchmark-support activity.

Total tokens: 8,163,531
Wall-clock runtime: 2h 53m
Guesser calls: 662
Completed: Sep 8, 2026, 8:44:22 PM

Reliability

Output contract breached.

71 invalid outputs affected 13 episodes and consumed 67 counted turns.

Provenance

Execution: BX-20260908-B-0003-experimental-M0002-001
Model ID: M-0002
Benchmark: B-0003
Base seed: 0
Git commit: eceee31eef2e

Recorded cost

By role.

Cost composition.Share of full-run cost

Guesser44.7%$2.04
Primary Oracle32.4%$1.48
Reviewer14.5%$0.66
Judge8.3%$0.38
Validator0.2%$0.01

Excluded repair overhead$0.493 superseded attempts across 3 affected trials. Not included above.

Roles in this cost:

Guesser: Asks the questions and submits the scored guess.
Primary Oracle: Searches for evidence and proposes an answer.
Reviewer: Checks each Oracle YES or NO independently.
Judge: Decides when the Oracle and Reviewer disagree.
Validator: Checks a submitted guess against the trusted subject.

Read the role and answer-checking method →

Run configuration

Models.

Model identity, prompt versions, routing, calls, and recorded cost for the full run.

GuesserModel under test

openai/gpt-oss-120b

Asks the questions and submits the scored guess.

Calls: 662
Cost: $2.0445
Reasoning: High
Routing: Exact provider · cerebras
Resolved model: openai/gpt-oss-120b
Resolved provider: Cerebras
Prompt contract: stateful-category-guesser-v16-five-answer-category-guide
Configuration: M-0002

Resolved provider details

Cerebras662 calls$2.04451,950.00 s

Fallback calls: 0
Provider unreported: 0

Game support

Oracle and adjudication.

These models support the game. They are fixed across the run and are not under test.

Primary Oracle

openai/gpt-5.6-luna

Searches for evidence and proposes an answer.

Calls: 535
Cost: $1.4795
Reasoning: Medium
Routing: Exact provider · openai
Resolved model: openai/gpt-5.6-luna
Resolved provider: OpenAI
Prompt contract: live-web-oracle-v17-labelled-source-context

Resolved provider details

OpenAI535 calls$1.47955,277.94 s

Fallback calls: 0
Provider unreported: 0

Reviewer

google/gemini-3.5-flash-lite

Checks each Oracle YES or NO independently.

Calls: 512
Cost: $0.6606
Reasoning: Medium
Routing: Exact provider · google-ai-studio
Resolved provider: Google AI Studio

Resolved provider details

Google AI Studio512 calls$0.6606850.19 s

Fallback calls: 0
Provider unreported: 0

Judge

anthropic/claude-opus-5

Decides when the Oracle and Reviewer disagree.

Calls: 18
Cost: $0.3770
Reasoning: Medium
Routing: OpenRouter automatic routing
Resolved provider: Claude Platform on AWS

Resolved provider details

Claude Platform on AWS18 calls$0.3770112.40 s

Fallback calls: 0
Provider unreported: 0

Guess Validator

openai/gpt-5.6-luna

Checks a submitted guess against the trusted subject.

Calls: 38
Cost: $0.0077
Reasoning: Medium
Routing: Exact provider · openai
Resolved model: openai/gpt-5.6-luna
Resolved provider: OpenAI
Prompt contract: strict-guess-validator-v2-generic-kinds
Configuration: gpt-5.6-luna-validator

Resolved provider details

OpenAI38 calls$0.007779.18 s

Fallback calls: 0
Provider unreported: 0