Claude Opus 5: Features, Pricing, API and Benchmarks

Anthropic released Claude Opus 5 on July 24, 2026 as its practical high-end model for complex coding, research, and enterprise work. It sits below Claude Fable 5 in maximum capability, but costs half as much through the API and keeps the same base price as Claude Opus 4.8.

For most developers, that positioning is the point. Claude Opus 5 is not meant to win every benchmark at any cost. It is designed to deliver much of Fable 5's intelligence at a price and speed that make sense for regular advanced work.

Quick Answer

Claude Opus 5 is Anthropic's recommended starting model for complex agentic coding and enterprise workflows. Its API model ID is claude-opus-5. It has a 1-million-token context window, up to 128,000 output tokens, adaptive thinking enabled by default, and standard API pricing of $5 per million input tokens and $25 per million output tokens.

Choose Opus 5 when a task needs sustained reasoning, tool use, codebase understanding, or careful verification. Choose Sonnet 5 for faster and cheaper everyday workloads, and Fable 5 only when the highest available capability justifies twice the token price.

Claude Opus 5 at a Glance

The dateless model ID is still a pinned snapshot, not a rolling alias. Anthropic's model documentation says newer dateless IDs identify fixed releases, so a future Opus model should not silently replace claude-opus-5.

What Is New in Claude Opus 5?

Better Long-Horizon Agentic Coding

The largest improvement is not merely code generation. Anthropic says Opus 5 is better at holding the thread across long, multi-step tasks, identifying root causes, using tools, and continuing until it has verified a result.

That matters for work such as:

• Debugging a failure spread across multiple services
• Migrating a large codebase to a new API or framework
• Reviewing a pull request for subtle, project-specific risks
• Building and validating a feature across backend and frontend code
• Running research that requires repeated searches, calculations, and checks
• Producing a document or presentation through several revision cycles

These are exactly the jobs where a model can appear intelligent in the first few minutes and still fail by losing constraints later. Opus 5 is designed to reduce that drift.

Adaptive Thinking Is On by Default

Claude Opus 5 decides when and how much to reason on each turn. You do not need to add a thinking field for the normal adaptive behavior.

The main control is now output_config.effort:

• low for easier tasks where cost and latency matter most
• medium for routine analysis and coding
• high as the default balance
• xhigh for harder multi-step work
• max for the deepest available reasoning

There is an important migration detail: max_tokens limits the model's total output, including internal thinking and visible response text. A value that worked for Opus 4.8 without thinking may be too small for the same task on Opus 5.

Stronger Self-Verification

Anthropic reports that Opus 5 checks its work more proactively and iterates instead of stopping at the first plausible answer. Early users described it testing browser layouts, finding underlying bug causes, building its own validation harnesses, and checking research through independent methods.

This can improve reliability, but it also changes prompting. Repeated instructions such as "always run a final verification subagent" can cause over-verification because the model already does more of that work by default. Test older system prompts rather than carrying every workaround forward.

Mid-Conversation Tool Changes

The Claude API now has a beta feature that lets an application add or remove tools between conversation turns without invalidating the existing prompt cache.

Use the mid-conversation-tool-changes-2026-07-01 beta header when testing it. This is useful for agents that discover they need a database tool, browser, code runner, or specialist integration only after the task has started.

Default Fallbacks for Refusals

Another beta feature lets developers request Anthropic's recommended fallback model by refusal category instead of maintaining a fixed fallback list.

The current beta header is server-side-fallback-2026-07-01. Applications should still record which model completed a request, because a fallback can affect capability, behavior, and cost expectations.

A Lower Prompt-Cache Minimum

Claude Opus 5 can cache prompts starting at 512 tokens, down from 1,024 tokens on Opus 4.8. Smaller system prompts and tool definitions that previously could not create a cache entry may now qualify without code changes.

At the published rate of $0.50 per million cache-read tokens, stable prompt prefixes can materially reduce the cost of long-running agents. The usual rule still applies: keep reusable instructions and tool definitions stable, and move frequently changing content later in the prompt.

Claude Opus 5 Benchmarks: What the Claims Mean

Anthropic's launch results position Opus 5 as a strong value model rather than simply a cheaper Fable 5.

According to Anthropic:

• Opus 5 more than doubled Opus 4.8's score on Frontier-Bench v0.1 at a lower cost per task.
• At maximum effort on CursorBench 3.2, it finished within 0.5% of Fable 5's peak score at half the cost per task.
• Its ARC-AGI 3 result was three times the next-best model in Anthropic's comparison.
• Its Zapier AutomationBench pass rate was about 1.5 times the next-best model at the same cost per task.
• On OSWorld 2.0, it exceeded the other compared models at any given cost.

These are launch benchmarks, not guarantees for every application. Anthropic also notes details such as harness choice, effort settings, repeated attempts, and safety fallbacks in its methodology. A production decision should therefore use a small evaluation set made from your own real tasks.

The useful conclusion is narrower: Opus 5 appears most compelling when task difficulty, duration, and verification requirements rise. A short extraction or rewrite may not benefit enough to justify Opus pricing.

Claude Fable 5 vs Opus 5 vs Sonnet 5 vs Haiku 4.5

Anthropic launched distinct tiers because one model cannot optimize capability, latency, and price at the same time.

• Fable 5 is for the hardest work where maximum capability matters more than cost.
• Opus 5 is the practical advanced model for demanding daily engineering and enterprise tasks.
• Sonnet 5 is the general-purpose choice when responsiveness and cost matter alongside strong reasoning.
• Haiku 4.5 is for faster, cheaper, high-volume work that does not require the deepest reasoning.

For many production systems, the best architecture is routing rather than choosing one model forever. Use a cheaper model for classification, extraction, and simple transformations, then escalate genuinely difficult cases to Opus 5.

Claude Opus 5 Pricing Explained

Standard Claude API pricing is:

This is the same standard token price as Claude Opus 4.8. For the previous model's behavior and migration context, see my Claude Opus 4.8 developer guide. Opus 5 uses Anthropic's newer tokenizer, however, and the pricing documentation says it can produce roughly 30% more tokens for the same text than older Claude generations. The actual change depends on the workload, so compare total task cost rather than rate-card numbers alone.

Fast mode is a separate research preview priced at $10 per million input tokens and $50 per million output tokens. Anthropic says it runs around 2.5 times the default speed. As of July 27, 2026, Opus 5 Fast mode is available only on the first-party Claude API, not Amazon Bedrock, Google Cloud, or Microsoft Foundry.

Fast mode is useful when latency blocks a person or another agent. Standard mode is usually the better value for background jobs, overnight processing, and tasks where waiting does not affect the product experience.

Claude Opus 5 API Example in Python

Install the current Anthropic SDK and set ANTHROPIC_API_KEY in your environment. A basic streaming request can use adaptive thinking without an explicit thinking field:

from anthropic import Anthropic

client = Anthropic()

with client.messages.stream(
    model="claude-opus-5",
    max_tokens=16000,
    output_config={"effort": "high"},
    messages=[
        {
            "role": "user",
            "content": (
                "Review this migration plan. Identify hidden dependencies, "
                "propose a safer sequence, and define verification checks."
            ),
        }
    ],
) as stream:
    response = stream.get_final_message()

print(response.content)

Use high as a baseline, then move effort down or up based on measured quality, latency, and cost. For xhigh or max, leave enough room in max_tokens for the model's reasoning, tool calls, and final response.

Migrating From Claude Opus 4.8

The model-name change is simple:

claude-opus-4-8 -> claude-opus-5

The behavioral migration deserves more care:

1. Update the model ID to claude-opus-5.
2. Review max_tokens because thinking now counts toward the same output limit.
3. Use adaptive thinking and control depth with output_config.effort.
4. Do not combine disabled thinking with xhigh or max; the API returns a 400 error.
5. Remove redundant prompts that force verification on every step, then evaluate the result.
6. Handle refusal and model_context_window_exceeded stop reasons explicitly.
7. Verify tool-call JSON with a standard parser and preserve trailing newlines in tool arguments.
8. Run representative development evaluations before routing production traffic.

If you are migrating from Opus 4.5 or older, Anthropic's migration guide lists additional breaking changes around assistant prefills, sampling parameters, thinking budgets, tool versions, and legacy beta headers.

Current Limitations and Caveats

Claude Opus 5 is not automatically the best choice for every request.

• The web-fetch tool available on Sonnet 5 is not available on Opus 5.
• Opus 5 does not currently support Priority Tier.
• Fast mode is currently limited to the first-party Claude API.
• Disabling thinking at xhigh or max returns a 400 error.
• Stronger self-verification can add output and latency when old prompts also demand repeated checking.
• A 1M-token context window does not remove the need for clean context, retrieval, or compaction strategy.
• Security classifiers can route certain requests to a fallback model when fallbacks are enabled.

These details are especially important for agents that depend on predictable tool availability, latency guarantees, or strict model identity.

Should You Upgrade to Claude Opus 5?

Upgrade from Opus 4.8 if your important tasks are long, difficult, tool-heavy, or expensive to get wrong. The price per token is unchanged, while Anthropic reports meaningful gains in coding, research, computer use, visual work, and self-verification.

Do not send every request to Opus 5 by default. Sonnet 5 is likely the more efficient choice for general application features, interactive assistants, and medium-complexity coding. Haiku remains useful for fast classification and extraction. Fable 5 makes sense when your own evaluations show that its capability advantage is worth twice the Opus rate.

A practical rollout is:

1. Build a test set from 20 to 50 real production tasks.
2. Compare Opus 4.8, Opus 5, and Sonnet 5 at the effort levels you would actually use.
3. Measure completion quality, retries, tool failures, latency, and total task cost.
4. Route only the task classes that benefit from Opus 5.
5. Monitor the completed-work cost, not only input and output token rates.

That approach turns a model launch into an engineering decision instead of a benchmark reaction.

Frequently Asked Questions

What is Claude Opus 5?

Claude Opus 5 is Anthropic's advanced model for complex agentic coding, research, and enterprise work. It offers a 1M-token context window, 128K maximum output, adaptive thinking, vision, tool use, and stronger long-horizon verification.

What is the Claude Opus 5 API model ID?

The Claude API model ID is claude-opus-5. Anthropic documents this dateless ID as a pinned model snapshot rather than an evergreen alias.

How much does Claude Opus 5 cost?

Standard Claude API pricing is $5 per million input tokens and $25 per million output tokens. Prompt cache reads cost $0.50 per million tokens. Fast mode costs $10 input and $50 output per million tokens.

Does Claude Opus 5 have a 1M context window?

Yes. Claude Opus 5 has a 1-million-token context window by default, and 1 million tokens is also its maximum context size. Its maximum output is 128,000 tokens.

Is Claude Opus 5 better than Claude Fable 5?

Fable 5 remains Anthropic's highest-capability generally available model. Opus 5 is positioned as the more practical advanced model: Anthropic says it approaches Fable performance on some coding evaluations at roughly half the API cost.

Is Claude Opus 5 better than Claude Sonnet 5?

Opus 5 is intended for harder, longer, and more agentic work. Sonnet 5 is faster and cheaper, making it a better default for many everyday product features. The right choice depends on task quality, latency, and total cost in your own evaluations.

Is Claude Opus 5 available in Claude Code?

Yes. Anthropic says Opus 5 is available across its platforms and through usage credits in Claude Code. It is also the new default model on Claude Max and the strongest model available on Claude Pro.

Should I replace Opus 4.8 with Opus 5 immediately?

Test it first. The standard token price is unchanged, but adaptive thinking is now on by default and can change output length, latency, and max_tokens requirements. Run a representative evaluation set before moving production traffic.

Related Reading

• Claude Fable 5 and Mythos 5: Is It Worth the Hype?
• Claude Opus 4.8 Developer Guide
• Why I Moved From Claude Code to Codex for Daily Development

Official Sources

• Anthropic: Introducing Claude Opus 5
• Claude Platform: What's new in Claude Opus 5
• Claude Platform: Models overview
• Claude Platform: Pricing
• Claude Platform: Migration guide

Information and pricing were checked against Anthropic's official pages on July 27, 2026.