Anthropic Claude 4.8 Opus Documentation

Claude 4.8 Opus introduces a 1M-token context window alongside elite multimodal perception and reasoning. Built for long-horizon agentic work and memory-driven automation, it maintains structural coherence and logical depth across exceptionally long outputs and extended execution windows.

Key Capabilities

1,000,000 Token Context Window: Natively houses massive code repositories, complete architectural frameworks, or multi-year financial ledgers in a single prompt.
Highly Autonomous Agents: Engineered for long-running asynchronous pipelines that require minimal human oversight and adaptive decision-making.
Multimodal Comprehension: Seamlessly processes text, high-resolution imagery, and dense documentation/file inputs simultaneously.
End-to-End Project Orchestration: Exceptional at multi-stage debugging, parsing vast log histories, and coordinating systemic multi-file codebases.
Coherence Over Extended Sessions: Retains acute context and execution quality across long-horizon sessions without conversational drift or identity degradation.

Request Parameters

To interact with this model via an OpenAI-compatible gateway or OpenRouter endpoint:

Parameter	Type	Required	Description
`model`	`string`	Yes	Use `"claude-opus-4-8"` or `"anthropic/claude-4.8-opus"`.
`messages`	`array`	Yes	Array of standard conversation objects (role/content) supporting multi-modal payloads.
`max_tokens`	`integer`	No	Limits generation length. Supports massively expanded output targets.
`temperature`	`float`	No	Recommended: `0.0` for mission-critical logic; `0.5` for creative analysis.

Thinking Mode Configuration

claude-opus-4-8 supports Adaptive Thinking — the model dynamically decides when and how much to think. The legacy "type": "enabled" + budget_tokens parameters are removed on this model (will return a 400 error).

Important Constraints

If thinking is included, temperature, top_p, and top_k are not supported (will return 400).
max_tokens should be set generously (e.g., ≥ 64000) when using "effort": "xhigh" or "max".
Thinking content is omitted by default. Add "display": "summarized" to see the model's reasoning.

Effort Levels

Value	Description
`"low"`	Fast; minimal thinking for simple tasks
`"medium"`	Balanced speed and depth
`"high"`	Deep reasoning — recommended default
`"xhigh"`	Best for coding and agentic tasks
`"max"`	Maximum reasoning for the hardest problems

Request Example

json Copy

{
  "model": "claude-opus-4-8",
  "max_tokens": 64000,
  "thinking": {
    "type": "adaptive",
    "display": "summarized"
  },
  "output_config": {
    "effort": "xhigh"
  },
  "messages": [
    {
      "role": "user",
      "content": "Your question here"
    }
  ]
}

Note: "type": "enabled" and budget_tokens are fully removed on this model and will return a 400 error. Use "type": "adaptive" + output_config.effort instead.

claude-opus-4-8

Input

Output