Anthropic retired frontier model
Claude Opus 4.1
Anthropic’s August 2025 Opus upgrade for agentic tasks, real-world coding, multi-file refactoring, reasoning, research, and data analysis—now retired from the Claude API and preserved mainly as a legacy migration and compatibility baseline.
- Context window
- 200K
- tokens
- Max output
- 32K
- tokens
- Input
- $15.00
- historical · per 1M
- Cache read
- $1.50
- historical · per 1M
- Output
- $75.00
- historical · per 1M
01 / Overview
What Claude Opus 4.1 Was
Claude Opus 4.1 was Anthropic’s August 2025 upgrade to Claude Opus 4, focused on stronger agentic tasks, real-world coding, reasoning, research, and data analysis.
A Focused Upgrade Rather Than a New Model Generation
Anthropic released Claude Opus 4.1 on August 5, 2025.
The release kept the same premium pricing as Claude Opus 4 but improved several areas that mattered directly to software-engineering and research workflows.
Anthropic highlighted stronger agentic task performance, improved real-world coding, better reasoning, stronger in-depth research, better data analysis, more precise detail tracking, improved agentic search, and stronger multi-file code refactoring.
The model ID on the Claude API was:
claude-opus-4-1-20250805
Opus 4.1 used the dated model-ID convention that Anthropic employed before the 4.6 generation moved to dateless IDs.
The model later entered deprecation on June 5, 2026 and was retired from the Claude API on August 5, 2026.
Anthropic’s recommended replacement is Claude Opus 4.8.
That makes Opus 4.1 most relevant today as a historical coding benchmark, regression baseline, legacy integration reference, migration source model, and compatibility target on partner platforms where it remains available.
- Provider
- Anthropic
- Family
- Claude Opus 4.1
- Claude API ID
- claude-opus-4-1-20250805
- Released
- Aug 5, 2025
- Claude API status
- Retired
- Reliable knowledge cutoff
- Jan 2025
- Input modalities
- Text, Image
- Output modality
- Text
02 / Coding & refactoring
Opus 4.1 Was Built Around Real-World Coding Improvement
The most distinctive launch story of Claude Opus 4.1 was coding precision: Anthropic emphasized stronger agentic software work rather than a broad change to price, context size, or model tier.
74.5% on SWE-bench Verified at Launch
Anthropic reported a 74.5% score on SWE-bench Verified for Claude Opus 4.1.
More important for production use, the launch announcement highlighted external feedback about the kinds of code changes the model improved.
GitHub reported gains across most capabilities relative to Opus 4, with particularly notable improvements in multi-file code refactoring.
Rakuten highlighted the model’s ability to identify exact corrections in large codebases without introducing unnecessary edits or new bugs.
That precision is a useful historical differentiator.
Many coding agents fail not because they cannot generate code, but because they change too many files, rewrite unrelated logic, overlook dependencies, introduce regressions, stop before validation, or repeatedly revisit already solved parts of the task.
Opus 4.1 was specifically positioned as an improvement in this type of practical engineering behavior.
For historical evaluation, measure the entire coding trajectory: files touched, tests passed, unnecessary diffs, tool calls, repeated reads, repair iterations, human corrections, and final acceptance.
- 01
Locate the real change
Identify the relevant files and code paths inside a larger repository rather than editing broadly.
- 02
Refactor across files
Coordinate changes that span multiple modules while preserving surrounding behavior.
- 03
Avoid collateral edits
Minimize unnecessary modifications that increase review effort and regression risk.
- 04
Validate the trajectory
Evaluate tests, repair loops, changed files, and human corrections rather than judging a coding agent from one generated patch.
03 / Research & analysis
Research and Data Analysis Were the Other Major Upgrade Areas
Anthropic also described Claude Opus 4.1 as stronger at in-depth research and data analysis, especially when the task required tracking details across an agentic search process.
Better Detail Tracking Matters in Long Multi-Step Work
A research agent may collect many pieces of evidence before forming a conclusion.
The difficulty is not only finding information. The model must preserve which source supported which claim, conflicting evidence, dates, entities, constraints, intermediate calculations, unresolved questions, and the original research goal.
Anthropic specifically called out improvements in detail tracking and agentic search.
That positioning makes Opus 4.1 historically relevant for technical research, due diligence, data investigation, document comparison, repository research, multi-source synthesis, and iterative analysis with tools.
Its 200K context window was substantial for the 2025 generation, but modern Opus models later expanded the working set dramatically.
For comparisons with newer models, preserve the same source packet and acceptance criteria so improvements are not confused with access to more context.
- 01
Collect
Gather evidence from files, tools, documents, or external research steps.
- 02
Track
Keep entities, constraints, numbers, and source relationships consistent across a multi-step investigation.
- 03
Analyze
Combine evidence with reasoning rather than reducing the task to retrieval alone.
- 04
Synthesize
Produce a final conclusion that reflects the evidence accumulated across the complete research trajectory.
04 / Pricing
Claude Opus 4.1 Pricing
Claude Opus 4.1 kept the premium Claude Opus 4 price: $15 per million input tokens and $75 per million output tokens.
A Much More Expensive Opus Era
Historical Claude Platform pricing lists:
- input: $15.00 per 1M tokens;
- output: $75.00 per 1M tokens;
- 5-minute cache write: $18.75 per 1M tokens;
- 1-hour cache write: $30.00 per 1M tokens;
- cache read: $1.50 per 1M tokens.
Message Batches received a 50% discount on standard input and output pricing.
Those numbers matter for migration.
Later Opus generations substantially reduced the base token price while also expanding context and output capacity.
For example, Opus 4.8 uses $5 / $25 standard pricing.
That makes Opus 4.1 useful as an example of how much frontier-model economics changed between the first Claude 4 Opus releases and later 4.x models.
For a legacy agent, do not compare only one request. Include thinking tokens, repeated tool calls, cache writes, cache reads, failed attempts, long outputs, human intervention, and fallback models.
1M tokens · USD
- Input
- $15.00
- 5m cache write
- $18.75
- 1h cache write
- $30.00
- Cache read
- $1.50
- Output
- $75.00
Example: 20K input + 4K output
- Input cost
- $0.3000
- Output cost
- $0.3000
- Estimated uncached total
- $0.6000
05 / Context
200K Context and 32K Output
Claude Opus 4.1 used a 200,000-token context window and supported up to 32,000 output tokens, placing it firmly in the pre-million-context generation of Claude.
Large for 2025, Small Relative to Later Opus Models
A 200K context can still contain substantial working material: source code, long documents, images, conversation history, tool results, system instructions, and retrieved evidence.
For many 2025-era agent workflows, that was enough to hold a meaningful repository slice or research packet.
But later Opus generations moved to 1M context windows and 128K standard output.
That creates an important migration difference.
A workflow that used aggressive retrieval, summarization, or context pruning for Opus 4.1 may no longer need the same strategy on a newer model.
At the same time, preserving disciplined retrieval can still improve cost and signal quality.
Use migration testing to distinguish between context-management logic that was required by the old limit and logic that remains useful regardless of the model’s maximum window.
Context window
200,000
Max output
32,000
Opus 4.1 belongs to the 200K / 32K generation. Later Opus models expanded both input context and output capacity substantially.
06 / Extended thinking
Opus 4.1 Used Manual Extended Thinking
Claude Opus 4.1 predates adaptive thinking: reasoning had to be enabled through the older thinking.type: "enabled" API with an explicit budget_tokens value.
Reasoning Depth Was Controlled With a Token Budget
For Claude 4.5 and earlier thinking-capable models, Anthropic’s extended-thinking API uses:
thinking: {"type": "enabled", "budget_tokens": N}
The minimum budget_tokens value is 1,024.
The budget must normally be lower than max_tokens, because thinking tokens and final response tokens share the model’s output allowance.
A larger budget can improve difficult reasoning, but it also increases latency and billed output.
Anthropic recommends beginning with a task-appropriate budget and increasing it only when evaluation shows a useful quality gain.
Opus 4.1 also supported interleaved thinking for tool workflows through the interleaved-thinking-2025-05-14 beta header.
That allowed the model to reason between tool calls within one assistant turn.
However, its conversation-state behavior differs from later Opus models.
Opus 4.1 preserves only the latest assistant turn’s thinking blocks by default. Opus 4.5 and later Opus generations preserve earlier thinking blocks across turns.
Adaptive thinking and effort are not part of the Opus 4.1 reasoning contract.
Simple request
Leave extended thinking disabled when additional reasoning does not improve acceptance enough to justify latency and output cost.
Complex coding or research
Enable manual extended thinking and tune budget_tokens against quality, latency, tool behavior, and total output cost.
07 / Tools
Tool Use, Forced Tool Choice, and Interleaved Reasoning
Claude Opus 4.1 supported agentic tool workflows under the older Claude 4 tool contract, including automatic selection, forced tool choice, and optional interleaved thinking.
The Old Tool Contract Is Part of the Migration Surface
Anthropic’s current tool documentation still records Opus 4.1 behavior for tool_choice: auto, tool_choice: none, tool_choice: any, and forced named tools.
Forced tool selection is useful when the application knows that a specific action must occur.
Automatic selection is more appropriate when the model should reason about whether a tool is needed.
Opus 4.1 can also interleave thinking with tool use when the historical beta header is enabled.
These details matter because current Opus models have changed both reasoning and tool orchestration.
A legacy application may depend on a named tool always being called, exact tool-system-prompt overhead, old computer-use declarations, manual thinking budgets, or thinking blocks appearing around tool calls.
Migration tests should cover the complete agent loop rather than only a prompt-to-text response.
- Supported
Automatic tool choice
Let the model decide whether and when a tool should be invoked.
- Supported
Forced tool choice
The Opus 4.1 generation supports any-tool and named-tool forcing.
- Supported
Interleaved thinking
Manual extended thinking can reason between tool calls when the historical beta header is enabled.
- Supported
Image input
Use images alongside text in multimodal analysis and tool workflows.
- Supported
Streaming
Receive generated content progressively through the Messages API.
- Supported
Batch processing
Message Batches reduce standard token rates by 50% for asynchronous work.
08 / Caching
Prompt Caching Reduced a $15 Input Rate to $1.50 on Cache Hits
Prompt caching was especially valuable on Claude Opus 4.1 because its standard input price was high and repeated stable context could otherwise dominate agent cost.
Cache Reads Cost One Tenth of Standard Input
Historical rates were:
- standard input: $15.00/MTok;
- 5-minute cache write: $18.75/MTok;
- 1-hour cache write: $30.00/MTok;
- cache read: $1.50/MTok.
The minimum cacheable prompt length for Opus 4.1 is 1,024 tokens.
Typical cached prefixes included system instructions, tool definitions, repository context, policy documents, long reference material, and stable research packets.
Thinking adds a legacy detail.
On Opus 4.1 and earlier Opus models, previous thinking blocks are removed from context beyond the most recent assistant turn.
That differs from later Opus generations, where more thinking state is preserved and billed as input.
When replaying historical cost models, reproduce the actual conversation structure rather than applying current Opus token-accounting assumptions to old traces.
- 01
Cache stable prefixes
Reuse expensive system prompts, tools, repositories, and reference documents rather than paying the $15/MTok base input rate every turn.
- 02
Measure cache hits
A cache marker is useful only when request prefixes remain stable enough to produce actual reads.
- 03
Respect the 1,024-token minimum
Shorter prefixes are processed normally even if marked for caching.
- 04
Use historical thinking semantics
Do not assume later Opus thinking-block preservation rules when reconstructing Opus 4.1 token usage.
09 / Retirement
Claude Opus 4.1 Is Retired on the Claude API
Anthropic deprecated Claude Opus 4.1 on June 5, 2026 and retired it from the Claude API on August 5, 2026—exactly one year after the model’s release.
Retired Is Stronger Than Deprecated
Anthropic’s model lifecycle distinguishes active models, deprecated models that are still callable before retirement, and retired models that no longer accept inference requests on that platform.
For the Claude API, claude-opus-4-1-20250805 is now retired.
Requests naming the retired Claude API model fail.
Anthropic’s official replacement is:
claude-opus-4-8
Partner-operated platforms have their own lifecycle schedules.
Anthropic’s current documentation still identifies Opus 4.1 as available outside the Claude API in some partner contexts, including Amazon Bedrock and Google Cloud.
On Amazon Bedrock, the documented legacy base model ID is:
anthropic.claude-opus-4-1-20250805-v1:0
Current Bedrock documentation lists a US inference-profile path for the model.
Therefore, “retired” should always be interpreted together with the platform: retired on the Claude API, while partner availability can differ.
- 01
Released · Aug 5, 2025
Anthropic launched Opus 4.1 as an upgrade to Opus 4 for agentic work, coding, reasoning, research, and data analysis.
- 02
Deprecated · Jun 5, 2026
Anthropic notified Claude API developers that Opus 4.1 would be retired two months later.
- 03
Retired · Aug 5, 2026
Claude API inference requests to claude-opus-4-1-20250805 no longer succeed.
- 04
Replacement · Opus 4.8
Anthropic names Claude Opus 4.8 as the official replacement for Opus 4.1.
10 / Migration
Migrating From Claude Opus 4.1 to Opus 4.8
Anthropic recommends Claude Opus 4.8 as the replacement for Opus 4.1, but the generations differ enough that migration should be treated as a behavior and integration change—not a simple model-ID swap.
The Replacement Changes Reasoning, Capacity, Vision, and Economics
The model ID changes from:
claude-opus-4-1-20250805
to:
claude-opus-4-8
Several major differences need retesting.
Reasoning
Opus 4.1 uses manual extended thinking with budget_tokens.
Opus 4.8 uses adaptive thinking and effort controls. The old manual reasoning contract is not the correct migration target.
Context
Opus 4.1 provides 200K context.
Opus 4.8 provides 1M context.
Output
Opus 4.1 uses a 32K standard output ceiling.
Opus 4.8 supports 128K standard output.
Pricing
Opus 4.1 historically cost $15 input / $75 output per million tokens.
Opus 4.8 costs $5 / $25.
Vision and agent behavior
Later Opus models introduce newer image processing, different thinking-state preservation, updated tool behavior, newer tokenization, and additional agent capabilities.
Migration testing should replay multi-file refactoring, repository debugging, long research, image-heavy tasks, manual-thinking prompts, tool loops, forced-tool workflows, cached conversations, and large outputs.
- 01
Replace manual thinking
Move from thinking enabled plus budget_tokens to the newer adaptive-thinking and effort contract.
- 02
Re-baseline context
Test whether old retrieval, summarization, and pruning logic is still necessary with the 1M context window.
- 03
Replay agent workflows
Retest tool selection, thinking blocks, image handling, cached histories, long outputs, and failure recovery.
- 04
Recalculate economics
Opus 4.8’s $5/$25 standard rates are one third of Opus 4.1’s historical $15/$75 rates before workflow-level differences.
11 / 4.1 vs 4.8
Claude Opus 4.1 vs Claude Opus 4.8
Opus 4.1 and Opus 4.8 belong to the same Claude 4 family, but their APIs and production economics represent two very different stages of the Opus platform.
Opus 4.1 Is Now Primarily a Historical and Compatibility Baseline
Claude Opus 4.1:
- released August 5, 2025;
- retired on the Claude API;
- 200K context;
- 32K output;
- $15 / $75 historical pricing;
- manual
budget_tokensthinking; - dated model ID;
- earlier thinking-state behavior.
Claude Opus 4.8:
- active legacy model;
- 1M context;
- 128K standard output;
- $5 / $25 pricing;
- adaptive thinking with effort;
- dateless model ID;
- newer agent and multimodal behavior.
For a new workload, this is not a close model-selection decision.
Opus 4.1 is retired on Anthropic’s direct API.
Its remaining value is in historical benchmark comparisons, partner-platform compatibility, migration validation, reproducing old production behavior, and understanding how an older coding agent produced a result.
- 01
4.1 · historical frontier
A premium 2025 coding and reasoning flagship with 200K context, manual thinking budgets, and $15/$75 pricing.
- 02
4.8 · replacement
Anthropic’s official replacement with 1M context, 128K output, adaptive thinking, and lower $5/$25 pricing.
- 03
Do not compare price alone
Tokenizer, reasoning, tools, vision, context capacity, and agent behavior all changed across the generations.
- 04
Preserve regression cases
Use the same repository tasks, research prompts, and tool trajectories when documenting migration behavior.
12 / Evaluation
Claude Opus 4.1 Strengths and Limitations
Claude Opus 4.1 remains historically important because it improved practical coding precision and agentic research at the top of Anthropic’s 2025 model lineup, but today it is a retired Claude API model with an expensive legacy contract.
Historical strengths
Strong real-world coding
Anthropic launched Opus 4.1 specifically with improvements in agentic coding, multi-file refactoring, and precise large-codebase corrections.
Research detail tracking
The release improved in-depth research, data analysis, detail tracking, and agentic search.
Manual reasoning control
Extended thinking exposes an explicit budget_tokens control for workloads that needed predictable reasoning budgets.
Mature legacy tool support
Tool use, forced tool choice, interleaved thinking, prompt caching, vision, streaming, and batch processing support substantial legacy agent workflows.
What to consider
Retired on Claude API
Anthropic retired claude-opus-4-1-20250805 on August 5, 2026, so direct Claude API inference is no longer available.
Very high historical price
$15 input and $75 output per million tokens is far above later Opus generations.
Smaller context and output
200K context and 32K output are substantially below the 1M / 128K profile of later Opus models.
Legacy reasoning contract
Manual budget_tokens thinking and older thinking-block semantics require migration work when moving to modern adaptive-thinking models.
Preserve the legacy baseline
Compare Claude Opus 4.1 behavior with its replacements
Use historical prompts, coding regressions, multi-file refactors, extended-thinking configurations, cached contexts, and tool workflows to document how newer Opus models change quality, cost, context capacity, and integration behavior.
Start FreeClaude Opus 4.1 is retired on the Claude API. Anthropic’s official replacement is Claude Opus 4.8; partner-platform availability can differ, so verify the provider endpoint before attempting live requests.
Common Questions
What is Claude Opus 4.1?
Claude Opus 4.1 is Anthropic’s August 2025 upgrade to Claude Opus 4, focused on stronger agentic tasks, real-world coding, multi-file refactoring, reasoning, research, and data analysis.
What is the Claude Opus 4.1 model ID?
The Claude API model ID was claude-opus-4-1-20250805. It uses the dated model-ID format that Anthropic used before Claude 4.6 introduced dateless IDs.
When was Claude Opus 4.1 released?
Anthropic released Claude Opus 4.1 on August 5, 2025.
Is Claude Opus 4.1 deprecated?
It progressed beyond deprecation on the Claude API. Anthropic deprecated Opus 4.1 on June 5, 2026 and retired it from the Claude API on August 5, 2026.
Can I still call Claude Opus 4.1 through the Claude API?
No. Anthropic marks claude-opus-4-1-20250805 as retired on the Claude API, and inference requests naming a retired model fail.
Is Claude Opus 4.1 still available on Amazon Bedrock or Google Cloud?
Anthropic’s current documentation still lists Opus 4.1 as an exception to Claude API retirement on some partner-operated platforms, including Bedrock and Google Cloud. Availability and retirement schedules are platform-specific and should be checked with the provider before relying on the model.
What model replaces Claude Opus 4.1?
Anthropic’s official deprecation documentation recommends Claude Opus 4.8 as the replacement for Claude Opus 4.1.
What is the Claude Opus 4.1 context window?
Claude Opus 4.1 has a 200,000-token context window.
What is the maximum output of Claude Opus 4.1?
Claude Opus 4.1 supports up to 32,000 output tokens in the historical model configuration.
What input types does Claude Opus 4.1 support?
Claude Opus 4.1 accepts text and image input and produces text output.
What is Claude Opus 4.1’s knowledge cutoff?
Anthropic’s published Claude Opus 4.1 system prompt states a reliable knowledge cutoff at the end of January 2025.
How much did Claude Opus 4.1 cost?
Historical standard pricing was $15 per 1M input tokens and $75 per 1M output tokens.
How much did Claude Opus 4.1 prompt caching cost?
Historical pricing was $18.75 per 1M tokens for 5-minute cache writes, $30 for 1-hour cache writes, and $1.50 per 1M tokens for cache reads.
What was the minimum cacheable prompt for Claude Opus 4.1?
Anthropic documents a 1,024-token minimum cacheable prompt for Claude Opus 4.1 on platforms where it remains supported.
Did Claude Opus 4.1 support extended thinking?
Yes. Opus 4.1 supports the older manual extended-thinking API using thinking type enabled and an explicit budget_tokens value.
What is the minimum thinking budget for Claude Opus 4.1?
Anthropic’s manual extended-thinking API requires at least 1,024 budget_tokens.
Does Claude Opus 4.1 support adaptive thinking?
No. Opus 4.1 belongs to the earlier Claude 4 generation where manual extended thinking is the supported reasoning mode. Adaptive thinking arrived in later Claude generations.
Does Claude Opus 4.1 support interleaved thinking?
Yes for Messages API tool workflows when the historical interleaved-thinking-2025-05-14 beta header is used.
Does Claude Opus 4.1 preserve old thinking blocks across the whole conversation?
No. Anthropic documents that Opus 4.1 and earlier Opus models preserve only the latest assistant turn’s thinking blocks by default, while Opus 4.5 and later Opus models preserve earlier thinking blocks.
Can Claude Opus 4.1 force a specific tool?
Yes. Anthropic’s tool-use documentation records support for automatic, none, any-tool, and named-tool choice on Opus 4.1.
How good was Claude Opus 4.1 at coding?
Anthropic reported 74.5% on SWE-bench Verified at launch and highlighted improvements in multi-file refactoring and precise corrections inside large codebases.
How is Claude Opus 4.1 different from Claude Opus 4?
Opus 4.1 kept the same pricing and general generation profile as Opus 4 but improved agentic tasks, real-world coding, reasoning, research, data analysis, detail tracking, and multi-file code refactoring.
How is Claude Opus 4.1 different from Claude Opus 4.8?
Opus 4.1 is a retired Claude API model with 200K context, 32K output, $15/$75 historical pricing, and manual thinking budgets. Opus 4.8 is Anthropic’s official replacement with 1M context, 128K standard output, $5/$25 pricing, and adaptive thinking with effort controls.
Should I use Claude Opus 4.1 for a new application?
No for the Claude API because the model is retired there. Use a current Claude model for new development. Opus 4.1 remains useful for historical evaluation, regression testing, migration work, or partner-platform deployments that still support it.
Model information
Last updated
Release positioning, coding results, retirement dates, replacement guidance, historical pricing, prompt-caching rates, extended-thinking behavior, tool-use behavior, and platform availability on this page are based primarily on Anthropic’s official Claude Opus 4.1 announcement and Claude Platform documentation. The historical 32K output limit is retained as the published model configuration used by supported provider integrations.