Claude Opus 4.5

A Distinctive Pre-Adaptive-Thinking Opus

Anthropic released Claude Opus 4.5 on November 24, 2025. The pinned Claude API model ID is claude-opus-4-5-20251101, while the provider also exposes the alias claude-opus-4-5.

The model provides a 200,000-token context window and supports up to 64,000 output tokens. It accepts text and images and produces text.

Its most distinctive technical characteristic is the reasoning interface. Opus 4.5 uses manual extended thinking rather than adaptive thinking. Developers explicitly enable thinking and set a budget_tokens value that controls how much reasoning capacity is available.

At the same time, Opus 4.5 supports Anthropic’s effort control. This makes it unusual in the Claude lineup: Anthropic describes it as the only model that is extended-thinking-only while also supporting effort.

Anthropic lists Opus 4.5 as Active (legacy) and recommends migrating to Claude Opus 5.5.

• Model ID: claude-opus-4-5-20251101.
• Claude API alias: claude-opus-4-5.
• Released November 24, 2025.
• 200K context window.
• 64K maximum output.
• Text and image input with text output.
• Extended thinking with manual budget_tokens.
• Active legacy status.

Effort and Budget Solve Different Problems

Anthropic’s current documentation distinguishes the two controls.

budget_tokens determines the available reasoning depth for extended thinking. It is a concrete token budget.

effort is broader. It influences how much work Claude puts into the whole response, including reasoning, tool use, and answer depth. Anthropic lists high as the default effort for Opus 4.5.

That means the two settings should not be treated as duplicates. A workload can use the same thinking budget at different effort levels and still produce different behavior.

Opus 4.5 is especially useful as an evaluation baseline because it makes these two dimensions explicit. Measure whether higher effort or a larger thinking budget improves the metric that actually matters: task completion, code correctness, analysis quality, tool efficiency, or reviewer acceptance.

When moving to Opus 4.6 or later, this dual-control pattern changes. Adaptive thinking replaces the need to manually set a reasoning token budget.

Reasoning Tokens Are Part of Output Economics

Anthropic lists 5-minute cache writes at $6.25 per million tokens, 1-hour cache writes at $10 per million tokens, and cache reads at $0.50 per million.

These rates are identical to several later legacy Opus releases, but the workload economics can still differ because Opus 4.5 manually allocates thinking tokens inside the output envelope.

For extended-thinking requests, reasoning tokens contribute to output consumption. A benchmark that compares only the visible final answer can therefore underestimate the cost of a reasoning-heavy request.

Prompt caching remains valuable for stable prefixes such as system instructions, large documents, repository context, schemas, or recurring policies.

Batch API workloads receive a 50% discount on input and output, which is useful for offline evaluations or large non-interactive jobs.

• Input: $5.00 per 1M tokens.
• Output: $25.00 per 1M tokens.
• 5-minute cache write: $6.25 per 1M tokens.
• 1-hour cache write: $10.00 per 1M tokens.
• Cache read: $0.50 per 1M tokens.
• Batch API: 50% discount on input and output.

Snapshot Identity and Release Date Are Separate Concepts

Anthropic exposes both a pinned snapshot ID and an alias.

The pinned ID is claude-opus-4-5-20251101. The alias is claude-opus-4-5.

A pinned snapshot gives an application an explicit version target rather than relying only on an alias. That is useful for reproducible evaluations, regression testing, and production systems that want to know exactly which model identifier was requested.

The 20251101 suffix should not be interpreted as the public release date. Anthropic’s model page lists November 24, 2025 as the release date.

Cloud-platform identifiers also differ. Amazon Bedrock uses anthropic.claude-opus-4-5-20251101-v1:0, while Google Cloud uses claude-opus-4-5@20251101.

This Is a Reasoning-Architecture Migration

The biggest change is not the model ID. It is the thinking contract.

Opus 4.5 uses thinking: {"type": "enabled", "budget_tokens": ...}. Opus 5.5 rejects that format. Thinking is always on, and applications control depth with effort instead of setting a manual token budget.

The context window increases from 200K to 1M tokens, and standard maximum output grows from 64K to 128K. Standard pricing falls from $5/$25 to $4/$20 per million input/output tokens.

Several request formats also change. Remove assistant prefills, incompatible sampling parameters, legacy beta headers, and forced tool selection that current Opus does not accept.

Do not mechanically convert a historical budget_tokens value into an effort level. Anthropic recommends testing effort levels on your own evaluations because the controls operate differently.