Opus 4.8

Name: Opus 4.8
Author: Anthropic

Anthropic · 2026-05-28

Opus 4.8 was released as Anthropic's most capable general-access model, improving software engineering, agentic tool use, and knowledge work over Opus 4.7. It supports optional adaptive thinking and is a permitted fallback target for refused Fable 5 requests.

Benchmark scores

Benchmark	Score
A-Fantasia	50.0
A-Fantasia (Valid Attempts with Retries)	34.7
A-Fantasia Backwards Spelling	28.0
A-Fantasia Chess	31.0
A-Fantasia Chess (Valid Attempts with Retries)	30.0
A-Fantasia Cube (Valid Attempts with Retries)	57.0
A-Fantasia Cube Rotation	91.0
A-Fantasia Spelling (Valid Attempts with Retries)	17.0
AA-AnalystAgent (pass@1)	63.5
AA-AnalystAgent (pass@5)	78.8
AA-AnalystAgent (pass^5)	45.0
AA-Briefcase Elo	1354
AA-Briefcase v1.1 Elo	1320
AA-LCR	67.7
AA-Omniscience Index	28.8
ActiveVision	2.4
ActiveVision with tools	24.7
AdvancedIF	73.7
AdvancedIF (rubric mean score)	93.5
Agent Red Teaming k=100 (CAIS)	47.0
Agents' Last Exam (ALE)	27.0
Agents' Last Exam ALE-CLI Pass Rate	25.7
Agents' Last Exam ALE-CLI Score	44.6
Agents' Last Exam Full-Spectrum Pass Rate	23.6
Agents' Last Exam Full-Spectrum Score	43.1
Agents' Last Exam Last-Exam Pass Rate	2.6
Agents' Last Exam Last-Exam Score	14.4
Agents' Last Exam Near-Term Pass Rate	43.3
Agents' Last Exam Near-Term Score	64.0
Agents' Last Exam Overall Score	45.2