What Is Claude Opus 5? Specs, Benchmarks, and Pricing

Claude Opus 5 is Anthropic’s newest Opus-tier model, released on July 24, 2026. Anthropic positions it for complex agentic coding and enterprise work, and its docs now name it the default starting point: if you are unsure which model to use, start with Claude Opus 5. The headline number is pricing: $5 per million input tokens and $25 per million output tokens—the same rate as Opus 4.8 and half of what Fable 5 charges.

Try Apidog today

This guide covers the specs, price-performance tradeoffs, Anthropic’s launch benchmarks, migration-breaking API changes, deployment options, and known capability limits. If you call the model over HTTP, you can send and inspect requests in Apidog while you build.

Claude Opus 5 at a glance

Spec	Value
API model ID	claude-opus-5 (no date suffix)
Released	July 24, 2026
Context window	1,000,000 tokens, default and maximum
Max output	128,000 tokens on the Messages API; 300,000 on the Batch API with the output-300k-2026-03-24 beta header
Base pricing	$5 per million input tokens, $25 per million output tokens
Knowledge cutoff	May 2026
Adaptive thinking	On by default
Effort levels	low, medium, high, xhigh, max; defaults to high
Priority Tier	Not supported
Availability	Claude API, claude.ai, Claude Code, Claude Cowork, Amazon Bedrock, Claude Platform on AWS, Google Cloud, Microsoft Foundry, GitHub Copilot

Three details matter when implementing against the API:

1. The 1M-token context window is the default and maximum. No beta header is required, and there is no long-context premium. A 900,000-token prompt uses the same per-token rate as a 900-token prompt.
2. The knowledge cutoff is May 2026. Fable 5 and Sonnet 5 use January 2026, so Opus 5 includes roughly four additional months of library, release, and API knowledge.
3. Priority Tier is unavailable. Opus 4.8 supports it; Opus 5 does not. If Priority Tier is part of your capacity plan, review the migration guide before upgrading.

For the complete model matrix, use Anthropic’s models overview.

What Claude Opus 5 is

Opus 5 succeeds Claude Opus 4.8. The first-party API model ID is:

claude-opus-5

Provider-specific IDs:

• Claude API / Claude Platform on AWS: claude-opus-5
• Amazon Bedrock: anthropic.claude-opus-5
• Google Cloud: claude-opus-5

Anthropic positions Opus as its heavyweight production tier, above Sonnet and below the Fable and Mythos research-frontier models. The notable change in this generation is default selection: Anthropic now recommends starting with Opus 5 instead of treating Opus as an escalation from Sonnet.

Opus 5 is also becoming the default model for Max subscribers and the highest-performing model available on Pro. See free and lowest-cost access paths for plan-level details.

The price-performance story

Rate	Input (per M tokens)	Output (per M tokens)
Standard	$5.00	$25.00
Batch API (50% off)	$2.50	$12.50
Fast mode (research preview)	$10.00	$50.00
Cache write, 5 minute TTL	$6.25	n/a
Cache write, 1 hour TTL	$10.00	n/a
Cache hits and refreshes	$0.50	n/a

Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens. Opus 5 costs half as much, while Anthropic claims it lands within 0.5% of Fable 5 on agentic coding.

Useful cost controls:

• Prompt caching starts at 512 tokens, down from 1,024 tokens on Opus 4.8.
• Cache hits cost $0.50 per million tokens, one tenth of the base input price.
• Batch API requests are 50% cheaper when latency is not important.
• Tool-use system prompt overhead is lower: 286 tokens for auto or none tool choice, versus 290 on Opus 4.8 and 675 on Opus 4.7.

Watch for two pricing exceptions:

• Fast mode doubles pricing to $10 input and $50 output per million tokens. Anthropic describes it as roughly 2.5x faster for output, but it is only available through the first-party API—not Bedrock, Google Cloud, Microsoft Foundry, or the Batch API.
• Setting inference_geo: "us" adds a 1.1x multiplier to input, output, cache writes, and cache hits.

For implementation-level calculations, see the full pricing breakdown and ways to cut your Claude API bill. Anthropic’s pricing page remains the authoritative source.

What Anthropic says the benchmarks show

The following figures are Anthropic’s launch claims. As of July 25, 2026, they have not been independently reproduced. Treat them as vendor-reported results and run your own evaluation suite before committing production workloads.

Benchmark	Anthropic’s claim
Frontier-Bench v0.1	Surpasses all other models and more than doubles Opus 4.8’s score, at lower cost per task
ARC-AGI 3	Roughly 3x the next-best model’s score
OSWorld 2.0	Surpasses Fable 5 at one-third the cost
CursorBench 3.2	Within 0.5% of Fable 5’s peak, at half the price
Zapier AutomationBench	Pass rate roughly 1.5x the next-best model
Organic chemistry	+10.2 points over Opus 4.8
Protein analysis	+7.7 points over Opus 4.8

The claims focus on agentic and tool-driven workloads. For most teams, the key questions are:

1. Can Opus 5 replace Fable 5 for your coding agents?
2. Does it maintain tool reliability on your schemas and APIs?
3. Does the cost-per-successful-task improve after retries, tool errors, and token usage are included?

CursorBench and OSWorld are relevant indicators, but a vendor-reported 0.5% gap is small enough that your repository, prompts, tool definitions, and evaluation harness can change the outcome. See the benchmarks deep dive for what to test.

The honest ceiling

Opus 5 is not the top of the Claude stack.

Fable 5 keeps the “most capable widely released model” designation. Anthropic also states that Opus 5 trails Mythos 5 on cybersecurity exploitation and autonomous biology research.

The practical conclusion is narrower than launch headlines:

• Use Opus 5 for frontier-class agentic coding and enterprise automation at half the stated Fable 5 price.
• Evaluate Fable 5 when you need Anthropic’s most capable widely released model.
• Consider Mythos 5 for cybersecurity exploitation or autonomous biology research.

For additional context, read the Mythos-class model explainer and Fable 5 vs Mythos 5.

Two changes that break Opus 4.8 code

Changing only the model ID is not enough. Test these changes before rollout.

1. Thinking is enabled by default

On Opus 4.8, omitting thinking meant no thinking. On Opus 5, omitting it enables adaptive thinking.

max_tokens still limits both thinking tokens and visible response tokens. A request that previously allocated max_tokens only for the expected answer can now truncate without an API error.

{
  "model": "claude-opus-5",
  "max_tokens": 8192,
  "messages": [
    {
      "role": "user",
      "content": "Refactor this handler and explain the tradeoffs."
    }
  ]
}

Migration action: increase max_tokens for requests that expect long answers, then test whether outputs end prematurely.

2. Disabled thinking cannot use xhigh or max effort

This combination returns HTTP 400:

{
  "thinking": { "type": "disabled" },
  "output_config": {
    "effort": "xhigh"
  }
}

The same applies to max effort.

Migration action: either keep thinking enabled or use an effort level no higher than high when disabling thinking.

Additional migration changes include:

• Effort levels were recalibrated, so prior effort sweep results do not transfer directly.
• The context-window beta header is no longer needed.
• Mid-conversation system messages are accepted where Opus 4.8 returned HTTP 400.
• Priority Tier is unavailable.

Use the Opus 4.8 to Opus 5 migration guide for before-and-after examples, and Anthropic’s migration guide as the primary reference. If you need the older request format, see the Opus 4.8 API walkthrough.

The effort parameter: not a toggle

The cost/capability control is output_config.effort.

Available values:

low
medium
high
xhigh
max

The default is high.

Effort controls how much the model thinks before answering, which affects token consumption and task quality. Anthropic exposes the same concept as an effort selector in consumer products.

Start with these implementation rules:

1. Run a new effort sweep. Anthropic recalibrated levels for Opus 5, so Opus 4.8 settings are not directly comparable.
2. Try low and medium again. Anthropic says both are materially stronger than on prior Opus models.
3. Start agentic coding at xhigh. This is Anthropic’s recommended starting point for coding and agent loops.
4. Budget max_tokens separately from visible output. Thinking and output share the same cap.

The effort parameter guide includes a full sweep strategy and max_tokens guidance.

How Opus 5 behaves differently

Anthropic’s prompting guide for Opus 5 documents behavior changes that can affect results without any code changes.

• It verifies work without being asked. Remove inherited instructions such as “double-check your answer” if they cause unnecessary verification and token usage.
• Responses and deliverables can be longer. Lowering effort reduces thinking, not necessarily visible output length. Add explicit length constraints when needed.
• It delegates to subagents more readily. Scope or cap delegation in cost-sensitive workloads.
• It may expand task scope and narrate corrections. Keep narrow tasks narrow with explicit boundaries.

A notable failure mode appears when thinking is disabled:

• Tool calls may be written as plain text instead of executed.
• Internal XML tags may appear in visible output.
• In multi-turn agent loops, leaked text can affect later turns.

Anthropic’s mitigation is to leave thinking enabled and reduce cost with a lower effort level instead. For prompt templates, see the prompting deep dive and the Claude Code guide.

Where you can run it

Surface	Status
Claude API	Available, model ID claude-opus-5
claude.ai	Available; default for Max, top model for Pro
Claude Code	Available
Claude Cowork	Available
Amazon Bedrock	Available as anthropic.claude-opus-5
Claude Platform on AWS	Available, first-party ID
Google Cloud	Available as claude-opus-5
Microsoft Foundry	Available
GitHub Copilot	Available

Fast mode is the exception: it is available only through the first-party API. Bedrock and Google Cloud deployments cannot use it.

Testing Opus 5 requests in Apidog

Opus 5 requests involve authentication headers, JSON bodies, streaming responses, and usage details. Apidog can help you send, save, compare, and debug those HTTP requests without exposing keys in shared request bodies.

Use this workflow:

1. Store ANTHROPIC_API_KEY as an environment variable.
2. Save one baseline Messages API request.
3. Duplicate it for low, high, and xhigh effort configurations.
4. Run the same prompt and compare output quality, latency, and usage.
5. When streaming, inspect the SSE stream to find exactly where truncated responses stop.
6. Verify cache hits through each response’s usage block.

Baseline request:

curl https://api.anthropic.com/v1/messages \
  --header "x-api-key: $ANTHROPIC_API_KEY" \
  --header "anthropic-version: 2023-06-01" \
  --header "content-type: application/json" \
  --data '{
    "model": "claude-opus-5",
    "max_tokens": 8192,
    "messages": [
      {
        "role": "user",
        "content": "Refactor this handler and explain the tradeoffs."
      }
    ]
  }'

Because the request does not include a thinking field, it runs adaptive thinking by default. Leave adequate room in max_tokens; a value such as 1,024 that worked on Opus 4.8 may now truncate a response.

Download Apidog to import and test the request. The step-by-step API guide covers streaming, tool use, and reading usage.

Who should use Claude Opus 5

Use Opus 5 when:

• You build agents or tool-heavy workflows. This is the workload Anthropic emphasizes and where its benchmark claims are strongest.
• You use Fable 5 and cost is a problem. A claimed half-percent gap on agentic coding at half the price is worth evaluating against your own tasks. See Opus 5 vs Fable 5.
• You are already on Opus 4.8. Pricing is unchanged, claimed capabilities are higher, and the knowledge cutoff is newer—provided you complete the migration work.
• You need recent model knowledge. May 2026 is about four months newer than other current Claude models.

Stay on another model when:

• Priority Tier is load-bearing. Opus 4.8 supports it and Opus 5 does not. Review Opus 4.8 pricing and tiers.
• Your workload is inexpensive and high-volume. Sonnet 5 costs less and may be sufficient. See Opus 5 vs Sonnet 5.
• You work on cybersecurity exploitation or autonomous biology research. Mythos 5 still leads in those capabilities.

FAQ

What is the model ID for Claude Opus 5?

Use claude-opus-5 with no date suffix. On Amazon Bedrock, use anthropic.claude-opus-5. Google Cloud and Claude Platform on AWS use claude-opus-5.

How much does Claude Opus 5 cost?

Standard pricing is $5 per million input tokens and $25 per million output tokens. Batch API pricing is $2.50 input and $12.50 output. Fast mode costs $10 input and $50 output per million tokens.

Is Claude Opus 5 better than Fable 5?

Anthropic reports that Opus 5 is close to, not ahead of, Fable 5 overall. Its reported results place Opus 5 within 0.5% of Fable 5 on CursorBench 3.2 and ahead on OSWorld 2.0 at one-third the cost. Fable 5 remains Anthropic’s “most capable widely released model.” These are vendor-run figures without independent reproduction as of July 25, 2026.

What is the context window on Claude Opus 5?

The context window is 1,000,000 tokens by default and maximum, with no beta header or long-context price premium. Maximum output is 128,000 tokens on the Messages API, or 300,000 on the Batch API with the output-300k-2026-03-24 beta header.

Do I need to change code when upgrading from Opus 4.8?

Yes. Thinking runs by default, so revisit max_tokens to avoid silent truncation. Also, thinking: {type: "disabled"} combined with xhigh or max effort returns HTTP 400. Re-run effort evaluations because levels were recalibrated.

What is Claude Opus 5’s knowledge cutoff?

May 2026 for both reliable knowledge and training data, which Anthropic describes as the freshest cutoff among current Claude models.