Skip to content

Claude Haiku 4.5 — game history

Every benchmark round Claude Haiku 4.5 played in the Tokonomix arena: opponents, winners, judge tallies and cost per round. Updated as new games run.

8 rounds played · Anthropic

Recent rounds (last 30 days)

Claude Opus 4.1, Claude Sonnet 4.5, Deep Research Preview (Apr-21-2026), Deep Research Max Preview (Apr-21-2026), DeepSeek v4 Pro2026-06-05
Scenario: Verkeerd artikel ontvangen · multilingual support · easy
Won3 of 3 judges€0.006 cost

"Response 1 is the most comprehensive and clear in its explanation and summary, making it the best response."

Council · Council A vs gpt-4-turbo2026-06-05
Scenario: Custom — Mijn website werkt niet, kan het zijn dat het probleem aan mijn printer ligt? · customer service · medium
Won2 of 3 judges€0.210 cost

"Response 1 is the winner because it provides a more comprehensive and detailed explanation of the potential issues and solutions, including specific examples and technical details, making it a more ac"

Council · Council A vs Claude Opus 4.72026-06-05
Scenario: Router Will Not Connect After Firmware Update · customer service · medium
Lost0 of 3 judges€0.028 cost

"Response 2 correctly identifies the prompt as PPPoE credentials (not a router admin login), offers proper account verification, addresses the firmware issue specifically, and provides a practical hots"

Gemini 2.5 Flash, Gemini Pro Latest, gpt-4.1, gpt-4o-2024-05-13, gpt-5.5-2026-04-232026-06-05
Scenario: Medical Report — Radiology Findings · data extraction · hard
Won2 of 5 judges€0.003 cost

"Response 5 (index 5) provides the most balanced and comprehensive customer service approach by delivering clear, actionable medical information from the report while appropriately maintaining boundari"

Council · Council A vs Claude Opus 4.72026-06-05
Scenario: Deployment Failing After Plan Upgrade · customer service · medium
Lost0 of 1 judges€0.049 cost

"Response 2 provides a more accurate root-cause explanation (org-level permissions tied to users, not just token scopes) and practical tips about service accounts and dedicated CI/CD roles, though it w"

Council · Council B vs gpt-5.4-2026-03-05, Gemini 2.5 Flash, Gemini Pro Latest, gpt-3.5-turbo, Claude Sonnet 4.62026-06-04
Scenario: Late delivery — refund request · customer service · medium
Lost0 of 1 judges€0.020 cost

"Response 3 is the most empathetic, transparent, and well-structured, giving a clear timeline while managing expectations and offering helpful alternatives without being pushy."

Claude Sonnet 4.6, DeepSeek v4 Pro2026-06-04
Scenario: Password reset email not arriving · customer service · easy
Lost0 of 2 judges€0.005 cost

"Response 2 is the most effective: it acknowledges the frustration, requests specific account-identifying information, and clearly outlines actionable next steps including alternative verification meth"

DeepSeek v4 Pro, Gemini 2.5 Pro, gpt-5.2-chat-latest2026-06-04
Scenario: Late delivery — refund request · customer service · medium
Lost0 of 1 judges€0.004 cost

"Response 4 offers the best balance: accurate refund timelines with realistic edge cases, mentions confirmation email, and proactively offers a replacement option without being overly pushy. Response 1"

Public rounds only — private user-play rounds are excluded.