Claude Opus 5 Benchmarks Explained: How Fast, Powerful, and Affordable Is Anthropic’s New AI Model?

Overall Intelligence

Benchmark: Artificial Analysis Intelligence Index

What it measures:
A collection of evaluations covering advanced reasoning, mathematics, scientific knowledge, instruction following and problem-solving.

Top-model comparison:
Claude Opus 5 Max: 61
Claude Fable 5 Max: 60
GPT-5.6 Sol Max: 59
Kimi K3: 57
Claude Opus 4.8 Max: 56

What it means:
Opus 5 narrowly leads the overall Intelligence Index, although its one-point advantage over Fable 5 means the two models should be considered broadly competitive rather than separated by a large capability gap.

Professional Knowledge Work

Benchmark: GDPval-AA v2

What it measures:
The ability to research information, analyze source material, make professional judgments and produce complete workplace deliverables.

Top-model comparison:
Claude Opus 5 Max: 1,861 Elo
Claude Fable 5 Max: 1,747 Elo
GPT-5.6 Sol Max: More than 100 Elo points behind Opus 5

What it means:
Opus 5 established the highest reported GDPval-AA v2 result, finishing 114 Elo points ahead of Fable 5 and more than 100 points ahead of GPT-5.6 Sol.

The result suggests that Opus 5’s largest advantage may appear during complicated professional assignments rather than simple question-and-answer tasks.

Reports, Presentations and Spreadsheets

Benchmark: AA-Briefcase

What it measures:
Realistic agent-based work involving thousands of source files and deliverables such as reports, presentations, research documents and spreadsheets.

Top-model comparison:
Claude Opus 5 Max: 1,720 Elo
Claude Fable 5 Max: 1,574 Elo
GPT-5.6 Sol Max: 1,505 Elo

What it means:
Opus 5 finished 146 Elo points ahead of Fable 5 and 215 points ahead of GPT-5.6 Sol on this professional-work benchmark.

Its advantage was driven primarily by objective correctness and analytical quality. GPT-5.6 Sol remained highly competitive in the visual presentation quality of completed deliverables.

Agentic Coding

Benchmark: Artificial Analysis Coding Agent Index

What it measures:
The ability of an AI coding agent to understand repositories, navigate development tools, modify multiple files and independently complete software-engineering assignments.

Top-model comparison:
Claude Opus 5 XHigh with Claude Code: Joint first place
GPT-5.6 family: Among the closest leading competitors
Claude Fable 5: Comparable frontier-level coding performance

What it means:
Opus 5 is positioned among the strongest available coding models rather than simply outperforming older Claude releases.

It also recorded the highest result on SWE-Atlas-QnA, one of the evaluations included in the Coding Agent Index.

Terminal and Command-Line Tasks

Benchmark: Terminal-Bench v2.1

What it measures:
The ability to complete difficult technical assignments through a command-line terminal, including navigating files, executing tools and resolving errors.

Top-model comparison:
Claude Opus 5 Max: 89%
GPT-5.6 Sol XHigh: Approximately level with Opus 5

What it means:
Opus 5 is one of the leading terminal-use models, although the result does not establish a decisive advantage over GPT-5.6 Sol.

This benchmark is particularly relevant for coding agents, system-administration assistants and AI tools that operate development environments.

Expert-Level Academic Reasoning

Benchmark: Humanity’s Last Exam

What it measures:
Extremely difficult expert-level questions covering mathematics, science, the humanities and other specialized academic subjects.

Top-model comparison:
Claude Opus 5 Max: Approximately 53%
Claude Fable 5 Max: Approximately level with Opus 5

What it means:
Opus 5 reached Fable-level performance but did not establish a clear lead on this benchmark.

The result shows that Opus 5 is competitive with the largest frontier models in scientific and academic reasoning, but it does not dominate every type of intellectual task.

Factual Knowledge and Hallucinations

Benchmark: AA-Omniscience

What it measures:
Factual knowledge, answer accuracy and whether a model attempts to answer questions when it lacks reliable information.

Top-model comparison:
Claude Fable 5: Outperformed Opus 5
Claude Opus 5: Improved over Opus 4.8 but remained behind Fable 5

What it means:
Opus 5 is not the strongest model in every category. Fable 5 retains an advantage in factual knowledge, while Opus 5 may be more willing to provide an answer when it is uncertain.

Artificial Analysis measured a 50% hallucination rate for Opus 5 on this specific evaluation. That figure applies only to the AA-Omniscience testing methodology and should not be interpreted as the model’s hallucination rate across all everyday uses.

Performance per Dollar

What it measures:
How much benchmark performance a model delivers relative to the cost of running each task.

Average Intelligence Index task cost:
Claude Opus 5 Max: $2.03
Claude Fable 5 Max: $2.75
Claude Opus 4.8 Max: $1.80
Claude Sonnet 5 Max: $1.53

What it means:
Opus 5 is not the cheapest leading model, but it delivers approximately Fable-level intelligence at a substantially lower average cost per benchmark task.

Its strongest selling point is therefore not raw benchmark dominance alone. It is the combination of frontier-level performance, professional agent capabilities and a lower operating cost than Anthropic’s larger Fable 5 model.