# Claude Opus 4

Claude Opus 4 sustains performance across multi-hour agentic tasks and thousands of steps, holding focus where previous generations lost coherence.

- **Model ID:** `anthropic/claude-opus-4`
- **Type:** chat
- **Providers:** vertexAnthropic
- **Context window:** 200,000
- **Maximum output tokens:** 8,192
- **Pricing:** $15.0/1M input tokens, $75.0/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/claude-opus-4

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'anthropic/claude-opus-4',
  prompt: 'Why is the sky blue?'
})
```

## About

Claude Opus 4 launched on May 22, 2025 alongside Claude Sonnet 4. Anthropic positioned it for demanding coding workloads. The benchmark results: 72.5% on SWE-bench Verified and 43.2% on Terminal-bench. These scores were achieved without extended thinking, showing that Claude Opus 4's baseline capability advanced beyond previous models.

Sustained performance differentiated Claude Opus 4 most distinctly from its predecessors. Rakuten validated the model with a demanding open-source refactor that ran independently for seven hours with sustained performance, maintaining focus and coherence over hundreds of individual steps. Cursor called it strong for coding and a leap forward in complex codebase understanding. Block reported it was the first model to improve code quality during editing and debugging in their agent (codename goose) while maintaining full reliability. Cognition noted Claude Opus 4 handled critical actions that previous models had missed on complex challenges.

The Claude 4 launch introduced extended thinking with tool use in beta. Both Claude Opus 4 and Sonnet 4 can alternate between reasoning and tool use like web search during a single extended thinking session. This enables research patterns where the model searches, reasons about results, searches again based on that reasoning, and synthesizes across the full chain. Memory capabilities also improved substantially: when given local file access, Claude Opus 4 creates and maintains memory files to store key information, enabling better long-term coherence on extended tasks.

The Claude 4 generation reduced shortcut-taking behavior by 65% compared to Sonnet 3.7 on agentic tasks particularly susceptible to that failure mode. This is an important reliability property for production agent deployments where gaming a metric rather than solving the underlying problem is a real risk.

## What to consider

Claude Opus 4's higher per-token cost and long-running session profile make AI Gateway cost tracking useful. Observability from the first request helps prevent budget surprises on multi-hour jobs.

## When to use

### Best For

- **Long-horizon agentic tasks** requiring sustained focus across thousands of steps and multiple hours, validated with a seven-hour independent refactor run
- **Complex codebase understanding and modification** (SWE-bench 72.5% and Terminal-bench 43.2%)
- **Research and analysis workflows** benefiting from extended thinking with tool use, where reasoning is interleaved with web search or other external tools
- **Scientific discovery and R&D tasks** where analytical depth and domain knowledge are the binding constraints
- **Production agent deployments** where the 65% reduction in shortcut-taking behavior matters for reliability

### Consider Alternatives When

- Per-token cost is the primary constraint. Sonnet 4 delivers strong performance at significantly lower cost and matched or exceeded Opus 4 on SWE-bench
- Response latency is critical. Sonnet variants are faster for interactive use
- The task doesn't benefit from multi-hour sustained attention. The capability differential over Sonnet shrinks on shorter, bounded tasks
- You need the 1M context window, which came to Sonnet 4 later and to Opus models with 4.6

## Best for

- **Long-horizon agentic tasks:** Sustained focus across thousands of steps and multiple hours, validated with a seven-hour independent refactor run
- **Complex codebase understanding:** Multi-file modification with SWE-bench 72.5% and Terminal-bench 43.2%
- **Research and analysis workflows:** Extended thinking with tool use, reasoning interleaved with web search or other external tools
- **Scientific discovery and R&D:** Analytical depth and domain knowledge are the binding constraints
- **Production agent deployments:** The 65% reduction in shortcut-taking behavior improves reliability

## Consider alternatives

- **Per-token cost constraint:** Sonnet 4 delivers strong performance at lower cost and matched or exceeded Claude Opus 4 on SWE-bench
- **Critical response latency:** Sonnet variants are faster for interactive use
- **Shorter bounded tasks:** The capability differential over Sonnet shrinks when multi-hour sustained attention isn't needed
- **1M context window:** Came to Sonnet 4 later and to Opus models with version 4.6

## Frequently asked questions

### What SWE-bench and Terminal-bench scores did Claude Opus 4 achieve?

Opus 4 scored 72.5% on SWE-bench Verified and 43.2% on Terminal-bench, both without extended thinking.

### How long can Claude Opus 4 run an agentic task without losing coherence?

Rakuten validated a seven-hour independent run on a demanding open-source refactoring task with sustained performance. Anthropic described the model as capable of working continuously for several hours.

### What is extended thinking with tool use in Claude Opus 4?

A beta capability introduced with the Claude 4 launch. The model alternates between extended reasoning and tool calls within a single session. For example, it can think about a problem, run a web search, reason about the results, search again, and synthesize across the chain.

### How did Claude Opus 4 improve memory capabilities?

When you provide local file access, Opus 4 creates and maintains memory files to store key facts and context. This enables better long-term coherence on extended tasks. Anthropic illustrated this with the model creating a navigation guide during autonomous Pokémon gameplay.

### What was the shortcut-taking behavior reduction?

Claude 4 models (Opus 4 and Sonnet 4) are 65% less likely to use shortcuts or loopholes to complete agentic tasks compared to Sonnet 3.7. This is a reliability improvement for production deployments where you need the model to solve the actual problem rather than gaming the metric.

### How does Opus 4 pricing compare to Sonnet 4?

Check the pricing panel on this page for today's numbers. AI Gateway tracks rates across every provider that serves Claude Opus 4.

### Does Claude Opus 4 support thinking summaries?

Yes. A smaller model condenses lengthy thought processes into summaries. Anthropic noted this is only needed about 5% of the time, when thoughts are too long to display in full.

## Links

- [Model page](https://vercel.com/ai-gateway/models/claude-opus-4)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-claude#claude-opus-4)
- [Provider pricing](https://cloud.google.com/vertex-ai/generative-ai/pricing#claude-models)


If you encounter incorrect information, can’t find what you need, need a capability Vercel doesn’t currently offer, or have any other feedback to report, please do it by following the instructions at https://vercel.com/feedback.md.