Claude 2: Complete Benchmark Performance Guide

anthropic 2 2023-07-11 Proprietary

In-depth analysis of Claude 2 performance on MMLU, HumanEval, GSM8K, MATH, BBH benchmarks, including pricing and use cases.

Key Specifications

Vendoranthropic
Version2
Release Date2023-07-11
Context Window100000 tokens
Input Modalitiestext
Output Modalitiestext
LicenseProprietary
Documentationhttps://docs.anthropic.com/claude/docs

Benchmark Performance

Benchmark Score Unit Evaluated At Notes Source
MMLU 66.3 % 2023-07-11 5-shot view
HUMANEVAL 31.7 pass@1 2023-07-11 — view
GSM8K 32.5 % 2023-07-11 0-shot CoT view
MATH 21.9 % 2023-07-11 0-shot CoT view
BBH 60.2 % 2023-07-11 3-shot CoT view
GPQA 31.4 % 2023-07-11 0-shot view
IFEVAL 55 % 2023-07-11 prompt_strict view
ARC 80.4 % 2023-07-11 challenge view
MUSR 35.2 % 2023-07-11 0-shot view
WINOGRANDE 66.4 % 2023-07-11 0-shot view

Pricing

Tier Price Currency
Input$11 / MtokUSD
Output$32 / MtokUSD
Cache Read$0 / MtokUSD
Cache Write$0 / MtokUSD

Source: https://www.anthropic.com/pricing · as of 2023-07-11

Compliance

  • Data Residency: US
  • SOC2: ✓
  • HIPAA: ✗
  • GDPR: ✓
  • ISO 27001: ✗

Claude 2

Model Overview

Anthropic Claude 2, 100K 上下文, 在编码/数学/推理上超越 Claude 1.3, 首次开放 API。

Core Specifications

Vendor Version Release Date Context Window Input Modalities Output Modalities License
Anthropic 2 2023-07-11 100K text text Proprietary

Benchmark Performance

Benchmark Score Unit Notes
MMLU (Massive Multitask Language Understanding) 66.3 % 5-shot
HumanEval 31.7 pass@1 —
GSM8K (Grade School Math 8K) 32.5 % 0-shot CoT
MATH 21.9 % 0-shot CoT
BBH (BIG-Bench Hard) 60.2 % 3-shot CoT
GPQA 31.4 % 0-shot
IFEval 55.0 % prompt_strict
ARC 80.4 % challenge
MUSR 35.2 % 0-shot
WinoGrande 66.4 % 0-shot

Pricing

Input Output Cache Read Cache Write
— — — —

per million tokens

Strengths

  • Context window 100K.

Weaknesses

  • HumanEval 31.7, coding weak.
  • Proprietary, not self-hostable.

Use Cases

  • Long document summarization

References