Claude Opus 4 - AI model implementation reference

Benchmark & Evaluation

ModelRefs currently has partial, narrow benchmark coverage for Claude Opus 4. Treat the available benchmark evidence as one input to the decision, not a guarantee that Claude Opus 4 is the strongest option for your workload, and evaluate it on representative workloads before selecting it.

• Seven canonical provider-run records cover SWE-bench Verified, Terminal-Bench, GPQA Diamond, AIME 2025, tau-bench Retail, tau-bench Airline, and MMMU validation for the dated Claude Opus 4 launch snapshot.
• These scores are source-scoped and provider-reported. Several use extended thinking, tool scaffolds, increased step limits, or Anthropic-specific agent frameworks; none supplies a physical execution date or independent reproduction.

Implementation considerations

• Treat claude-opus-4-20250514 as a retired, immutable evaluation subject; inventory and migrate production calls rather than routing new traffic to it.
• For historical reproduction, preserve the exact extended-thinking budget, tool scaffold, prompt, sampling settings, and cloud channel because the launch scores are model-plus-harness results.
• Anthropic retired claude-opus-4-20250514 on its operated platforms on June 15, 2026; requests to a retired model fail, and Anthropic recommends claude-opus-4-8 as the replacement.
• Amazon Bedrock and Google Cloud control their own lifecycle schedules. Verify the exact partner model ID, region, retirement state, and data terms before assuming historical availability.

Architecture disclosure

• Anthropic documents Claude Opus 4 as a proprietary hosted hybrid-reasoning model with text and image input, text output, a 200,000-token context window, manual extended thinking, and tool use. It does not publish parameter count, layer design, tokenizer specification, serving topology, or model weights.
• Launch evaluations distinguish standard single-attempt results from parallel high-compute selection. ModelRefs admits only explicitly scoped rows and does not collapse the high-compute result into the standard score.

Prompts and code examples

Historical pinned Messages API request (retired snapshot)

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();
// Historical reproduction only: Anthropic retired this snapshot on 2026-06-15.
const message = await client.messages.create({
  model: "claude-opus-4-20250514",
  max_tokens: 4096,
  messages: [{
    role: "user",
    content: "Review this migration plan and identify the three highest-risk assumptions.",
  }],
});

console.log(message.content);

Provider implementation source

Training disclosure

• Anthropic does not publish a complete Claude Opus 4 training-data inventory, parameter count, training-compute ledger, or reproducible training recipe.
• Anthropic's privacy documentation describes broad categories of model-training data and personal-data handling, but it does not identify the exact Claude Opus 4 corpus or provide a model-specific contamination audit.
• The reviewed launch, system-card, and retirement sources do not establish an exact training cutoff suitable for ModelRefs to present as a verified model-specific fact.

History and source-reviewed changelog

1. 2025-05-22 — Claude Opus 4 launched

   Anthropic released Claude Opus 4 and Claude Sonnet 4 with hybrid reasoning, extended thinking with tool use, and provider-run coding, reasoning, agentic, multilingual, and vision evaluations.

   Milestone source
2. 2026-04-14 — Claude Opus 4 deprecated

   Anthropic announced the retirement schedule for claude-opus-4-20250514 and directed developers to migrate to Claude Opus 4.8.

   Milestone source
3. 2026-06-15 — Claude Opus 4 retired on Anthropic-operated platforms

   Anthropic marks the dated Claude Opus 4 API model as retired; historical evidence remains valid only for the pinned snapshot and does not establish current availability.

   Milestone source

Risks and limitations

• Outputs can be incorrect or unsuitable for the intended task; use task-specific evaluation, grounding, and human review where consequences are material.
• API availability, model aliases, rate limits, data controls, regions, and prices are mutable and differ by product channel.

Source coverage

ModelRefs verified the Claude Opus 4 launch, exact API snapshot identity, system card, historical pricing, Messages API shape, extended-thinking controls, retirement record, and seven source-scoped provider evaluations. Anthropic does not publish the physical evaluation execution dates, so ModelRefs preserves them as undisclosed rather than substituting the May 22, 2025 publication date.

Connected ModelRefs evidence

• Anthropic provider profile
• Decision guide
• Decision guide
• Related workflow
• Related workflow
• Related workflow
• Benchmark evidence
• Benchmark evidence
• Benchmark evidence

Sources

• Introducing Claude 4 — Anthropic
• Claude 4 System Card — Anthropic
• Model deprecations — Anthropic
• Pricing — Anthropic
• Create a Message — Anthropic
• Extended thinking — Anthropic
• How do you use personal data in model training? — Anthropic

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Claude Opus 4 - AI model implementation reference.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path
• Compare Claude Opus 4 with Claude Sonnet 4
• Compare Claude Opus 4 with GPT-5