> ## Documentation Index
> Fetch the complete documentation index at: https://documentation.deepmask.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Opus 5

> Claude Opus 5 from Anthropic — the next-generation flagship. Up to 96.2% GPQA, 5M context, continuous reasoning, and native multi-agent orchestration.

Opus 5 is Anthropic's next-generation flagship within DeepMask and the new ceiling of the Opus line. It is a generational step beyond the Opus 4 series, built for sustained autonomous work, deep multi-document reasoning, and complex multi-agent systems — while keeping the accuracy and long-horizon coherence that demanding, high-stakes tasks require.

## About Opus 5

**Opus 5** (released August 26, 2026) replaces the discrete effort toggles of the Opus 4 series with **Continuous Reasoning**, which allocates thinking depth dynamically per step instead of per request. It reaches a GPQA Diamond score of 96.2%, handles context windows up to 5M tokens with persistent cross-session memory, and introduces native **multi-agent orchestration** — spinning up, coordinating, and load-balancing fleets of sub-agents without external frameworks. Combined with real-time self-verification carried over from Opus 4.8, it is the strongest option in DeepMask for mission-critical research, security, and large-scale engineering.

## Key Capabilities

<CardGroup cols={2}>
  <Card title="Continuous Reasoning" icon="infinity">
    Allocates thinking depth dynamically per step rather than per request — deep where a problem is hard, fast where it is routine, without manual effort settings.
  </Card>

  <Card title="Native Multi-Agent Orchestration" icon="diagram-project">
    Spins up, coordinates, and load-balances fleets of sub-agents natively — ideal for parallel research, security investigations, and large-scale refactors.
  </Card>

  <Card title="5M Context with Persistent Memory" icon="database">
    Reasons over up to 5M tokens and retains verified state across sessions, so long-running agents resume exactly where they left off.
  </Card>

  <Card title="Real-Time Self-Verification" icon="shield-check">
    Audits its own intermediate reasoning as it works and reasons natively over text, images, and structured data to catch errors before they propagate.
  </Card>
</CardGroup>

## Best For

Choose **Opus 5** for the most demanding work in your stack: multi-day autonomous engineering, enterprise-scale research across huge document sets, and multi-agent systems that coordinate many tools in parallel. It is the right pick when accuracy, long-horizon coherence, and self-directed effort matter more than cost or latency. For routine tasks, or where cost efficiency is the priority, Opus 4.5–4.8 or Sonnet 4.6 remain strong, lower-cost options.

## Use Cases

* **Multi-agent systems** — Orchestrate fleets of coordinated sub-agents for parallel research, QA, and tool use from a single prompt.
* **Enterprise-scale research** — Synthesize across 5M tokens of reports, filings, and technical material with persistent memory.
* **Autonomous engineering** — Plan, implement, test, and iterate on large codebases over multi-day sessions with resumable state.
* **Cybersecurity defense** — Run end-to-end vulnerability investigations with real-time self-verification of each step.
* **High-stakes analysis** — Legal, financial, and scientific discovery where edge-case reasoning and multi-step verification are mandatory.

## Specifications

| Specification | Opus 5 |
| - | - |
| Provider | Anthropic |
| Context Window | 5.0M tokens |
| Reasoning | Continuous (dynamic per-step) |
| GPQA Diamond | 96.2% |
| Latency (TTFT) | 0.12s |
| Throughput | 28 tokens/sec |
| Key use cases | Multi-agent orchestration, enterprise-scale research, autonomous engineering |

[Try Opus 5 in DeepMask →](https://chat.deepmask.io/)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.