# Claude Opus 4 (Thinking) ![Organization](https://img.shields.io/badge/Org-Anthropic-d97757) ![Context](https://img.shields.io/badge/Context-200K%20tokens-1f6feb) ![Pricing](https://img.shields.io/badge/Price-%2415%20%2F%20%2475%20per%201M-2da44e) ![License](https://img.shields.io/badge/License-Proprietary-d29922) ![Modality](https://img.shields.io/badge/Modality-Text%20%2B%20Vision-8957e5) ![Released](https://img.shields.io/badge/Released-May%202025-555) > **Quick answer:** Claude Opus 4 with extended thinking enabled is Anthropic's May 2025 frontier model at maximum reasoning depth. It scores 56.6% on LiveCodeBench and 10.72% on Humanity's Last Exam with thinking on. At $15/$75 per 1M tokens with a 200K context window, it targets the hardest coding and multi-step reasoning tasks. ## At a Glance **Where Claude Opus 4 (Thinking) leads** - 56.6% LiveCodeBench — strong competitive programming with extended reasoning - 10.72% Humanity's Last Exam — best in the Claude 4 family on expert tasks - Extended thinking: visible chain-of-thought reasoning for debugging and auditability - Computer use capability - 200K context window **Where it lags** - $15/$75 per 1M tokens — highest cost in the Claude 4 family - 10.72% HLE trails o3 (20.32%), Gemini 2.5 Pro (21.64%), and GPT-5.6 Sol - No open weights **Best for:** Hardest multi-step reasoning tasks, research-grade code generation, and production pipelines where auditability of reasoning steps is required. ## What Claude Opus 4 (Thinking) Is Claude Opus 4 is Anthropic's May 2025 flagship model — the highest-capability tier in the Claude 4 family. The "thinking" configuration enables extended thinking mode, which gives the model a scratchpad for internal reasoning before producing its response. This improves accuracy on multi-step problems at the cost of higher latency and more output tokens. The model's 56.6% LiveCodeBench score reflects strong coding capability when reasoning is applied. Combined with the 10.72% HLE score, it represents Anthropic's strongest general reasoning capability at the Opus 4 generation. For cost-sensitive workloads, Claude Sonnet 4 (Thinking) provides similar LiveCodeBench performance (55.9%) at significantly lower cost ($3/$15 per 1M). Claude Opus 4 (Thinking) is the correct choice when maximum accuracy on complex tasks is the primary requirement regardless of cost. ## Specifications | Field | Value | |---|---| | Organization | Anthropic | | Parameters | Undisclosed | | Context window | 200,000 tokens | | Max output | 32,000 tokens (including thinking tokens) | | License | Proprietary (API only) | | Release date | May 22, 2025 | | Knowledge cutoff | March 2025 | | Thinking mode | Extended (visible reasoning) | | Modality | Text + Vision (multimodal) | ## Pricing | | Input (per 1M tokens) | Output (per 1M tokens) | |---|---|---| | Anthropic API | $15.00 | $75.00 | Thinking tokens billed as output tokens. Prompt caching available. Pricing per [Anthropic pricing page](https://www.anthropic.com/pricing). ## Context Window Claude Opus 4 (Thinking) has a **200,000-token context window** — roughly **150 pages** of text in a single request. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [LiveCodeBench](/benchmarks/livecodebench) | 56.6% | Benchgen evaluation | 2025-07 | | [Humanity's Last Exam](/benchmarks/humanitys-last-exam) | 10.72% | Benchgen evaluation | 2025-07 | ## Claude Opus 4 (Thinking) vs Alternatives | Model | LiveCodeBench | HLE | Thinking | Price (in/out per 1M) | |---|---|---|---|---| | **Claude Opus 4 (Thinking)** | 56.6% | 10.72% | Yes | $15 / $75 | | [Claude Sonnet 4 (Thinking)](/models/anthropic/claude-sonnet-4-thinking) | 55.9% | 7.76% | Yes | $3 / $15 | | [o3](/models/openai/o3) | 75.8% | 20.32% | Yes | $10 / $40 | | [o4-mini (high)](/models/openai/o4-mini-high) | 80.2% | 18.08% | Yes | $1.10 / $4.40 | | [Gemini 2.5 Pro](/models/google/gemini-2-5-pro) | 73.6% | 21.64% | Yes | $1.25 / $10 | Claude Opus 4 Thinking's LiveCodeBench (56.6%) is significantly below o3 (75.8%) and o4-mini-high (80.2%). For coding-intensive workloads, OpenAI o-series or Gemini 2.5 Pro provide better results. Claude Opus 4 Thinking is differentiated by Anthropic's safety properties and visible chain-of-thought reasoning. ## Related Models & Benchmarks - Compare: [Claude Sonnet 4 (Thinking)](/models/anthropic/claude-sonnet-4-thinking) · [o3](/models/openai/o3) · [Gemini 2.5 Pro](/models/google/gemini-2-5-pro) - Benchmarks: [LiveCodeBench](/benchmarks/livecodebench) · [Humanity's Last Exam](/benchmarks/humanitys-last-exam) ## Frequently Asked Questions

What is Claude Opus 4 Thinking?

Claude Opus 4 (Thinking) is Anthropic's May 2025 flagship model with extended thinking mode enabled, scoring 56.6% LiveCodeBench and 10.72% HLE. Priced at $15/$75 per 1M tokens.

What is extended thinking in Claude Opus 4?

Extended thinking enables a visible reasoning scratchpad before the final response. The model uses this to break down complex problems step-by-step. Thinking tokens are billed as output tokens.

Should I use Claude Opus 4 or Sonnet 4 Thinking?

Claude Sonnet 4 (Thinking) scores 55.9% LiveCodeBench (vs 56.6%) and 7.76% HLE (vs 10.72%) at $3/$15 per 1M instead of $15/$75. For most workflows, Claude Sonnet 4 Thinking provides nearly identical coding performance at 80% lower cost. Use Opus 4 Thinking for the most demanding accuracy-critical tasks.

--- *Specs and scores from Anthropic's official Claude 4 announcement (May 2025) and Benchgen evaluations. Pricing cited to the Anthropic pricing page. Last updated 2026-07-24.*

Claude Opus 4 (Thinking) — Evaluations

Evaluations