Claude Opus 5's 1M Context Window: What “Default” Really Means

Claude Opus 5AnthropicContext WindowClaude CodeClaude APIPrompt CachingToken CostAI Agent

The short answer is yes: short-context and long-context Claude Opus 5 are the same model. Anthropic does not publish a separate “long” model ID. The API model is claude-opus-5, and its 1,000,000-token context window is both the default and the maximum.

But that answer needs four important qualifications:

  1. A 1M window is a capacity, not an instruction to fill it.
  2. Product access and subscription usage limits are separate from model capacity.
  3. Prompt caching can reduce the cost of repeated context, but it does not reduce the number of tokens in the context window.
  4. Long-context performance is better than before, but irrelevant or stale context can still reduce accuracy.

There is also one material correction to the original claim: the current Anthropic help center says Opus 5 supports a 1M context window in Claude chat on all paid plans. The older “Claude.ai is typically limited to about 200k” distinction is no longer accurate for Opus 5.

Verified Facts at a Glance

ItemCurrent Opus 5 specification
API model IDclaude-opus-5
Context window1M tokens, default and maximum
Separate short-context modelNo
Maximum output128k tokens
Base API price$5 / million input tokens, $25 / million output tokens
Long-context surchargeNone
Extended thinkingEnabled by default
API platformsClaude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry

These figures come from Anthropic’s Opus 5 model documentation and current pricing page. Prices and product policies can change, so check those pages before making a production budget.

What “1M by Default” Actually Means

“Default” does not mean every request automatically sends, processes, or bills one million tokens. It means there is no smaller standard Opus 5 context variant and no beta header required to unlock the larger window.

If a request contains 18,000 input tokens, you pay for roughly 18,000 input tokens, not the model’s full capacity. If the conversation later grows to 300,000 tokens, that larger active context is processed according to the API’s token accounting and caching rules.

The window also contains more than the text you typed. Anthropic’s context window guide says it can include:

  • the system prompt;
  • every retained user and assistant message;
  • tool definitions and tool results;
  • attached images and documents;
  • extended-thinking tokens; and
  • the new output being generated.

That last point matters. Opus 5 can produce up to 128k output tokens, but max_tokens covers both internal thinking and the visible answer when thinking is enabled. Do not pack the input to the final token and assume the model still has unlimited room to reason and answer.

Where the 1M Window Is Available

The model is the same, but the surrounding product decides how you access it and how usage is managed.

Surface1M context statusPractical caveat
Claude APIAvailable by defaultPay per processed token; rate and spend limits still apply
Amazon Bedrock, Vertex AI, Microsoft FoundryAvailableProvider-specific quotas and configuration still apply
Claude paid chatAvailable for Opus 5 on all paid plansChat may automatically summarize earlier messages near the limit
Claude Code on Pro, Max, Team, EnterpriseAvailable for Opus 5Pro users must enable usage credits to use 1M with Opus models

The current Claude paid-plan context guide explicitly lists 1M for Opus 5 in paid chat and Claude Code. This is why a fixed “API gets 1M, web gets 200k” comparison is misleading today.

A subscription’s usage limit is still a different constraint. You can have access to a 1M-capable model and reach a rolling session or weekly usage cap before filling that context window. Capacity answers “how much can one active request consider?” Usage limits answer “how much service can this account consume over time?”

The Same Model at Two Different Operating Points

A useful way to think about short and long context is not “small Opus versus large Opus.” It is the same engine with a different working set.

Lean contextVery large context
Faster and cheaper requestsMore cross-file and cross-document evidence in one request
Less irrelevant material competing for attentionBetter for relationships spread across a large corpus
Easier to debug why the model answered a certain wayHigher latency and input cost
Best for focused edits and quick questionsGreater risk of stale, duplicated, or conflicting instructions

Anthropic says Opus 5 maintains instruction following, tool use, and reasoning more consistently across long contexts than earlier models. That is a real improvement. It is not a promise that one million uncurated tokens are as useful as ten thousand carefully selected ones.

Anthropic’s own context window guidance still warns about “context rot”: as context grows, recall and accuracy can degrade. The practical goal is therefore not to maximize token count. It is to maximize the amount of relevant, current, non-duplicated evidence available to the model.

What 1M Tokens Cost in Practice

At the base Opus 5 input rate of $5 per million tokens, input-only examples are simple:

Input processed in one requestBase input cost
20,000 tokens$0.10
50,000 tokens$0.25
200,000 tokens$1.00
1,000,000 tokens$5.00

Output is billed separately at $25 per million tokens. A request with 1M uncached input tokens and 4,000 output tokens would therefore cost about $5.10 before any platform-specific charges: $5.00 input plus $0.10 output.

Also note that Opus models from Claude 4.7 onward use a newer tokenizer. Anthropic says the same text can produce roughly 30% more tokens than with earlier models. Recount a real prompt instead of estimating migration cost from an older model’s token total.

For API applications, use Anthropic’s free Token Count API before sending unusually large requests. It can count messages, tools, images, and documents using the target model’s tokenizer.

Prompt Caching: Powerful, but Often Misunderstood

Prompt caching is valuable when many requests reuse the same large prefix: a codebase snapshot, policy library, product catalog, or long system prompt.

For Opus 5, current cache pricing is:

Cache operationCost per million tokens
Standard input$5.00
5-minute cache write$6.25
1-hour cache write$10.00
Cache read$0.50

A cache hit therefore makes the reused portion cost 10% of standard input—a 90% reduction. However, the first write costs more than ordinary input, so caching pays off only when that prefix is reused.

Most importantly, cached tokens still count toward the 1M context window. Caching changes billing and latency for repeated prefixes; it does not create a 10M context window or remove cached text from the model’s active context. Anthropic’s prompt caching documentation explains the cache hierarchy and lifetime in detail.

A Practical Context-Budget Workflow

Use this process before feeding Opus 5 a very large repository or document stack.

1. Define the decision first

State the output you need: a migration plan, a list of conflicting contract clauses, a dependency map, or an implementation patch. A precise task makes it easier to exclude unrelated material.

2. Count the real request

Count the complete payload, not just source documents. Include tool schemas, conversation history, images, and the system prompt. Leave room for thinking and output.

3. Separate stable context from changing context

Put stable instructions and reference material in a reusable prefix. Put the current task, recent observations, and volatile data near the end. This structure also improves cache reuse.

4. Remove stale and duplicate evidence

Do not include three generated summaries of the same file, obsolete tool output, and the full original file unless the task needs all four. Conflicting copies make the model’s job harder.

5. Retrieve details on demand

For an enormous codebase, start with the repository map, key interfaces, tests, and files directly connected to the task. Let tools fetch additional files when a dependency is discovered. A 1M window is a useful ceiling, not a replacement for retrieval.

6. Compact long-running sessions

For agent workflows that accumulate many turns and tool results, summarize completed work, preserve decisions and unresolved questions, then discard detail that is no longer operationally useful. Anthropic also provides server-side compaction for supported API workflows.

7. Measure quality, cost, and latency together

Test a representative task with lean, medium, and large context. Compare correctness, missed evidence, response time, and token cost. The best context size is the smallest one that reliably contains the evidence the task needs.

When the Full 1M Window Is Worth It

Use very large context when the answer genuinely depends on relationships across a large corpus:

  • a cross-cutting refactor across hundreds of files;
  • dependency analysis spanning multiple services;
  • a legal or compliance review across many related documents;
  • a long-horizon agent that must preserve decisions and tool history;
  • a research synthesis where omissions are more costly than added latency; or
  • a migration in which old and new implementations must be compared together.

Even then, organize the input. Add a manifest, label document boundaries, distinguish authoritative sources from notes, and identify the task’s priority files.

When Lean Context Is Better

Keep context small for:

  • a single-file bug fix;
  • a focused code review;
  • rewriting one section of a document;
  • a quick factual or formatting task;
  • an interactive experience where latency matters; or
  • repeated production requests where every unnecessary token multiplies cost.

If Claude can fetch a missing file in seconds, sending an entire repository on every request is usually wasteful.

Decision Checklist

Before sending a huge prompt, ask:

  • Does the answer require evidence from most of this material?
  • Can retrieval provide the relevant pieces when needed?
  • Are any instructions duplicated, stale, or contradictory?
  • Have I reserved room for thinking and output?
  • Will the stable prefix be reused enough to justify caching?
  • Am I measuring subscription usage limits separately from context capacity?

Final Answer

Claude Opus 5 does not have separate short-context and 1M-context versions. claude-opus-5 natively supports a 1M-token window, with no special long-context model ID or surcharge.

But “supports 1M” is not the same as “should always receive 1M.” Product access, account usage limits, request cost, cache behavior, latency, and context quality remain separate operational decisions. Use the full window when a task truly needs broad cross-document reasoning. For focused work, a smaller, cleaner context is usually faster, cheaper, and easier to trust.

Official Sources

Related