Claude Opus 5 is now available: capabilities, pricing, and comparison with Opus 4.8, Sonnet 5, and Fable 5

On 24 July 2026, Anthropic released Claude Opus 5, a new model in its top-tier Opus family, designed primarily for complex agentic coding, long multi-stage processes, and demanding enterprise work.^[1]

The model’s official name is Claude Opus 5, not “Claude Opus 5.0”. The API identifier is:

claude-opus-5

Opus 5 does not replace Claude Fable 5 as Anthropic’s most capable publicly available model. Anthropic instead presents it as a model that approaches Fable 5’s capabilities at a lower price and with moderate, rather than the slowest, latency characteristics.^[2]

The most important news for current Opus 4.8 users is that the base price has not increased:

• USD 5 per million input tokens,
• USD 25 per million output tokens.^[3]

At the same time, Opus 5 introduces adaptive reasoning enabled by default, a new highest max effort level, a context window of up to one million tokens, up to 128,000 tokens of standard output, and new mechanisms intended for long-running agents.^[2][4]

The shortest conclusion: Claude Opus 5 is primarily a major upgrade for teams that assign complete tasks to a model rather than isolated code fragments. Not every application will benefit from migrating, however, because Sonnet 5 remains considerably cheaper and Fable 5 still occupies the top position in Anthropic’s range for maximum capability.

The information and documentation were verified on 24 July 2026.

TL;DR

Question	Answer
When was it released?	24 July 2026
What is the official name?	Claude Opus 5
What is the API identifier?	claude-opus-5
How much does it cost?	USD 5 / MTok input and USD 25 / MTok output
Has the price increased compared with Opus 4.8?	No
How large is the context window?	1 million tokens
What is the standard maximum output?	128,000 tokens
Is reasoning enabled by default?	Yes
Which effort levels are supported?	low, medium, high, xhigh, max
Does it replace Fable 5 as the strongest model?	No
Is Opus 4.8 disappearing?	No, it remains available
Where is it available?	Claude, Claude Code, API, AWS, Google Cloud, and Microsoft Foundry
Is Fast mode available?	Yes, as a research preview
Are all benchmarks independent?	No, most launch-day results were published by Anthropic or its partners

1. What is Claude Opus 5?

Anthropic describes Opus 5 as a model for:

• complex agentic coding,
• work on large repositories,
• long tool-use loops,
• solving multi-stage problems,
• document and data analysis,
• office work,
• coordinating multiple agents,
• building and refining interfaces and visual materials.^[4]

Opus 5 sits between Sonnet 5 and Fable 5:

Sonnet 5
A faster and less expensive model for work at scale

Opus 5
A model for difficult agentic and enterprise tasks

Fable 5
Anthropic's highest publicly available capability

Anthropic recommends starting with Opus 5 when an application performs complex agentic coding or enterprise work. Fable 5 should be selected when the highest available capability is required regardless of its higher cost.^[2]

2. Key technical specifications

Parameter	Claude Opus 5
API ID	claude-opus-5
Context	1,000,000 tokens
Standard maximum output	128,000 tokens
Reasoning	Adaptive, enabled by default
Effort levels	low, medium, high, xhigh, max
Input price	USD 5 / MTok
Output price	USD 25 / MTok
Batch API price	USD 2.50 / MTok input, USD 12.50 / MTok output
Fast mode	USD 10 / MTok input, USD 50 / MTok output
Reliable knowledge cutoff	May 2026
Training data cutoff	May 2026
Image input	Yes
Text output	Yes

The context, pricing, and cutoff data come from Anthropic’s official model overview and pricing documentation.^[2][3]

One million tokens is the default

In Opus 5, the one-million-token context window is both the default and maximum value. There is no separate smaller context variant of the model.^[4]

This does not mean every piece of information in a very long prompt will be used with identical accuracy. Maximum context capacity is not a guarantee of perfect retrieval, reasoning, or recall of every detail.

Standard output up to 128,000 tokens

Standard Messages API calls can generate up to 128,000 tokens. Anthropic also documents experimental output of up to 300,000 tokens for the Message Batches API when the relevant beta header is used.^[5]

This is intended for use cases such as:

• very long documentation,
• large-scale data extraction,
• extensive code scaffolding,
• book-length drafts,
• long reasoning processes.

Generating such a large response can take more than an hour and is not available through the ordinary synchronous Messages API.^[5]

3. Thinking is now enabled by default

This is one of the most important migration changes.

In Opus 4.8, a request without a thinking field ran without additional reasoning. In Opus 5, the same request starts adaptive reasoning. The model decides when and how long it should analyse the problem.^[4]

A basic request may look like this:

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-5",
    max_tokens=16_000,
    output_config={"effort": "high"},
    messages=[
        {
            "role": "user",
            "content": "Analyse the project architecture and propose a migration plan."
        }
    ],
)

print(response.content)

You do not need to add:

thinking={"type": "adaptive"}

because this is the default behaviour. Explicitly setting this value remains supported.

Five effort levels

Opus 5 supports the full scale:

low
medium
high
xhigh
max

max is the highest test-time compute level. It does not require a beta header.^[4]

Higher effort can improve quality on difficult tasks, but it usually also means:

• greater token consumption,
• higher cost,
• longer latency,
• more tool operations,
• longer work performed by subagents.

For this reason, max should not be set automatically for every request.

4. An important change: when can thinking be disabled?

Opus 5 permits thinking to be disabled only with effort set to:

• low,
• medium,
• high.

The combination:

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

or:

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

returns an HTTP 400 error.^[4]

This is a breaking change compared with Opus 4.8.

Anthropic also warns that, when thinking is disabled, Opus 5 may occasionally:

• write a tool call as ordinary text instead of a tool_use block,
• place internal XML markers in the visible response.

Anthropic recommends leaving thinking enabled and controlling cost through a lower effort setting unless the integration requires different behaviour.^[4]

5. What has actually improved compared with Opus 4.8?

According to Anthropic’s documentation, the largest changes concern:

1. deep reasoning,
2. agentic coding,
3. long-running tasks,
4. quality scaling with effort,
5. code review and bug finding,
6. visual analysis,
7. long-context work,
8. documents, spreadsheets, and presentations,
9. multi-agent coordination.^[4]

Longer tasks and less unfinished work

Anthropic states that Opus 5 maintains the goal better during long tool-use loops and more often completes:

• features spanning multiple files,
• large refactors,
• end-to-end tasks,
• work requiring testing and verification.

Anthropic highlights reduced use of placeholders and fewer unfinished fragments.^[4]

Automatic verification

Opus 5 more often checks its own results. The migration documentation even suggests removing repeated instructions such as:

Be sure to verify your work at the end.
Use a separate subagent for review.

Keeping such instructions may cause excessive verification, additional tool calls, and unnecessary cost.^[4]

More delegation to subagents

In multi-agent environments, the model more often splits tasks and delegates them to subagents. This can be useful for complex projects, but it requires:

• budget limits,
• monitoring the number of calls,
• controlling tool permissions,
• protection against concurrent editing of the same files.

6. Benchmarks: what did Anthropic announce?

At launch, Anthropic announced that Opus 5 achieved its strongest results in coding and knowledge-work evaluations, including Frontier-Bench and GDPval-AA.^[1]

The proper context is essential:

• some evaluations are internal,
• the vendor or a partner selected the configuration and harness,
• results may depend on effort,
• cost per task depends on token usage,
• not every benchmark has a public, easily reproducible dataset,
• these are not independent quality certifications.

Frontier-Bench v0.1

Anthropic states that Opus 5:

• ranks first on Frontier-Bench v0.1,
• scores more than twice as high as Opus 4.8,
• does so at a lower cost per task.^[1]

In a footnote, Anthropic explains that this was an internal run of Frontier-Bench v0.1 using the mini-SWE-agent harness, a GKE backend, and the mean of five trials per task. When safety-classifier refusals occurred, Opus 4.8 was used as a fallback model.^[1]

This methodological limitation should be reported together with the result.

CursorBench 3.2

At max effort, Opus 5 reportedly finished within 0.5 percentage points of Fable 5’s best result at roughly half the cost per task.^[1]

This can be useful for assessing the price-to-quality relationship, but it comes from tests published by Anthropic and its partner rather than a neutral certifying institution.

ARC-AGI 3

Anthropic reports that Opus 5’s result was roughly three times higher than the next-best model in the table.^[1]

Without a complete description of every configuration, this should not be translated into “three times more intelligent”. The benchmark measures a specific type of novel problem solving, not general intelligence.

Zapier AutomationBench

According to Anthropic, Opus 5’s pass rate was approximately 1.5 times higher than the next model at a similar cost per task. Anthropic also states that even Opus 5’s lowest effort level passed more tasks than the other models in the comparison.^[1]

OSWorld 2.0

On the computer-use benchmark, Opus 5 reportedly achieved the best performance-to-cost ratio and exceeded Fable 5’s strongest score at slightly more than one-third of the cost.^[1]

Scientific research

Anthropic reports improvements over Opus 4.8 in all of its life-sciences evaluations, including:

• 10.2 percentage points in an internal test of inferring molecular structures from spectroscopic data,
• 7.7 percentage points in tasks predicting how protein-sequence changes affect function.^[1]

These are vendor-internal results. They should not be presented as independent clinical or scientific validation.

7. Opus 5 vs Opus 4.8

Feature	Opus 4.8	Opus 5
Base price	USD 5 / 25 per MTok	USD 5 / 25 per MTok
Context	Up to 1M tokens	1M tokens by default
Maximum output	128k	128k
Thinking without configuration	Disabled	Adaptively enabled
Highest effort	Lower range	max
Minimum prompt-cache length	1,024 tokens	512 tokens
Tool changes during a conversation	No new capability	Beta
Automatic fallback by refusal category	No new mode	Beta
Fast mode price	USD 10 / 50	USD 10 / 50
Status	Still active	Latest Opus

Opus 4.8 has not been retired. According to Anthropic’s official schedule, both models remain active, and Opus 5 is not scheduled for retirement before 24 July 2027.^[6]

When should you stay with Opus 4.8?

Remaining on the previous model may make sense when:

• the existing prompt and agent are precisely tuned,
• the team has not completed regression testing,
• the application assumes thinking is disabled unless explicitly enabled,
• the process requires a stable, familiar response style,
• Opus 5 classifiers block legitimate cybersecurity workflows,
• migration does not produce a measurable quality improvement.

8. Opus 5 vs Sonnet 5

Feature	Sonnet 5	Opus 5
Profile	Speed and intelligence at scale	Difficult agentic and enterprise tasks
Standard price	USD 3 / 15 per MTok	USD 5 / 25 per MTok
Sonnet promotional price	USD 2 / 10 through 31 August 2026	No equivalent promotion
Context	1M	1M
Maximum output	128k	128k
Thinking	Adaptive	Adaptive
Anthropic latency profile	Fast	Moderate
Highest effort	Lower than Opus’s full range	max

Sonnet 5 will be the better choice for:

• high traffic,
• frequent short tasks,
• automation where cost takes priority,
• standard code generation,
• classification and extraction,
• applications requiring lower latency.

Opus 5 makes more sense when one incorrect or unfinished task costs more than the additional tokens.

9. Opus 5 vs Fable 5

Feature	Opus 5	Fable 5
Position in range	Difficult agentic coding and enterprise	Highest publicly available capability
Price	USD 5 / 25	USD 10 / 50
Context	1M	1M
Maximum output	128k	128k
Anthropic latency profile	Moderate	Slower
Thinking	Adaptive	Always enabled
Data-retention requirement	No special requirement for ordinary access	30 days required in the API
Cyber safeguards	Less restrictive	Considerably stronger classifiers

Anthropic presents Opus 5 as approaching Fable 5’s intelligence at half its price.^[1]

This does not mean Opus 5 always matches Fable 5. Anthropic’s official overview still identifies Fable 5 as its most capable broadly available model.^[2]

Where can Opus 5 be more practical?

• everyday programming work,
• code review,
• large refactors,
• enterprise workflows,
• document analysis,
• work requiring fewer refusals in legitimate cybersecurity,
• use cases where Fable’s cost is too high.

10. New features for agents and the API

Changing tools during a conversation

In beta, tools can be added or removed between turns without invalidating prompt cache.^[4]

The required header is:

mid-conversation-tool-changes-2026-07-01

This can be useful when an agent:

1. starts with analysis tools,
2. later receives repository access,
3. finishes with only testing and deployment tools.

It allows permissions to be limited instead of exposing every tool from the start.

Automatic fallbacks

The fallbacks parameter received a "default" mode in which Anthropic selects a fallback model by refusal category.^[4]

The feature is in beta and requires the header:

server-side-fallback-2026-07-01

In Anthropic applications, some blocked cybersecurity requests may be routed to Opus 4.8. A similar mechanism can be enabled in the API.^[1]

Shorter prompts can use cache

The minimum prompt length eligible for cache storage fell from:

1,024 tokens in Opus 4.8
to
512 tokens in Opus 5

This may reduce the cost of repeated system prompts and medium-length agent instructions.^[4]

11. Fast mode

Opus 5 offers Fast mode as a research preview.

Anthropic states that the model runs about 2.5 times faster in this mode than in standard mode.^[1]

Pricing is:

• USD 10 per million input tokens,
• USD 50 per million output tokens.^[3]

That is twice the base token price.

Fast mode is available through the Claude API and through usage credits in Claude Code. The documentation says it is not currently available through Amazon Bedrock, Google Cloud, or Microsoft Foundry.^[4]

It is worth using only when lower latency has real business value, such as:

• interactive pair programming,
• incident response,
• on-demand code review,
• urgent analytical workflows,
• an agent operating in a human-in-the-loop process.

12. Security and cybersecurity limitations

Anthropic states that Opus 5 does not move the frontier in the highest-risk dual-use capabilities and remains behind Mythos 5 in biology and offensive cybersecurity.^[1]

Anthropic intentionally did not train Opus 5 directly on cyber tasks. The model nevertheless improved at vulnerability discovery as a side effect of broader capability gains.

According to Anthropic:

• Opus 5 approaches Mythos 5 in identifying vulnerabilities,
• it remains substantially weaker than Mythos 5 in creating working exploits,
• it uses safety classifiers similar to Opus 4.8,
• it received stronger safeguards for a narrow range of cyber tasks.^[1]

What may be blocked?

Anthropic describes blocking, among other things:

• binary vulnerability scanning,
• some penetration-testing tasks,
• exploit generation.

At the same time, the model is intended to continue permitting beneficial uses such as finding vulnerabilities in source code.^[1]

Anthropic estimates that Opus 5 classifiers will intervene approximately 85% less often than Fable 5 safeguards. This is Anthropic’s forecast, not an independent measurement.^[1]

Verified organisations and researchers may use the Cyber Verification Program to obtain access to a variant with fewer safety restrictions.

13. What does Anthropic say about alignment?

In Anthropic’s automated behavioural audit, Opus 5 scored 2.3 for overall misaligned behaviour. Anthropic describes this as the lowest score among its recent models.^[1]

According to Anthropic, the model:

• follows the Claude Constitution better,
• behaves deceptively less often,
• is harder to persuade into misuse,
• less often performs risky actions with difficult-to-reverse consequences.

This is a vendor-run audit. The result does not mean the model is infallible, completely safe, or resistant to every prompt injection.

14. Limitations of Claude Opus 5

It does not know everything after May 2026

The model’s reliable knowledge cutoff and training-data cutoff are May 2026.^[2]

Later information requires:

• web search,
• current databases,
• tools,
• documents supplied in context.

One million tokens does not guarantee perfect memory

A large context window increases the amount of available material, but it does not eliminate:

• omitted details,
• misinterpretation,
• conflicting instructions,
• degradation in poorly organised context,
• distraction from irrelevant documents.

Default thinking makes cost less predictable

The same prompt can use a different number of tokens depending on the problem and effort setting. An application should measure:

• input tokens,
• output tokens,
• cache hits,
• response time,
• number of tool calls,
• total task cost rather than only per-request cost.

Longer responses may require prompt adjustments

The documentation warns that default responses and documents may be longer than in Opus 4.8. The model also reports work progress more often.^[4]

Tasks requiring a concise result should use explicit format limits.

A benchmark is not production

A strong benchmark result does not guarantee:

• correctness in a private framework,
• knowledge of internal APIs,
• compliance with company standards,
• absence of regressions,
• stable cost,
• good performance with a specific agent harness.

15. How should you migrate safely from Opus 4.8?

Step 1: do not swap the model without testing

Prepare a set of real tasks:

• bug fixing,
• refactoring,
• code review,
• test creation,
• document analysis,
• tool use,
• a long-horizon task.

Step 2: update the identifier

Replace:

model = "claude-opus-4-8"

with:

model = "claude-opus-5"

Step 3: review the thinking configuration

If the previous integration did not send a thinking field, the migrated model will start using adaptive reasoning.

If thinking must remain disabled, do not set effort above high.

Step 4: raise max_tokens for difficult tasks

Thinking and the final response share the same limit. A max_tokens value that is too low may end work earlier than in Opus 4.8.^[4]

Step 5: remove redundant verification instructions

Check whether the prompt demands repeated review that Opus 5 already performs by itself.

Step 6: measure the cost of a complete task

Compare:

cost of a completed task
time to a correct result
number of iterations
number of tool calls
number of manual corrections

Do not compare only the price per million tokens.

Step 7: retain a fallback

Opus 4.8 remains active. It can serve as:

• a fallback for classifier refusals,
• regression protection,
• a stable model for existing prompts,
• an alternative for specific cybersecurity workflows.

16. Is it worth switching to Claude Opus 5?

Yes, when:

• the agent completes tasks spanning many files,
• the model should work for a long time without continuous supervision,
• code review and subtle bug detection matter,
• you perform large refactors,
• the model uses many tools and subagents,
• you create documents, spreadsheets, or presentations,
• current Opus 4.8 often finishes only part of a task,
• Fable 5 is too expensive.

Not necessarily, when:

• tasks are short and repetitive,
• Sonnet 5 already provides sufficient quality,
• latency matters more than maximum accuracy,
• prompt regression testing has not been performed,
• the application requires thinking disabled at the highest effort,
• cybersecurity workflows are blocked by new classifiers,
• the quality gain does not compensate for additional token use.

17. Verdict

Claude Opus 5 is not simply “Opus 4.8 with a new number”.

The most important changes concern how work is performed:

• thinking is enabled by default,
• max effort is available,
• the model checks its own output more often,
• it handles long, multi-file tasks better,
• it coordinates subagents more readily,
• it adds mechanisms for dynamic tools and automatic fallbacks,
• it keeps the Opus 4.8 price.

Based on Anthropic’s information, Opus 5 may offer a strong capability-to-cost ratio for difficult programming and enterprise tasks. It is not an unconditional best choice:

• Sonnet 5 remains more economical,
• Fable 5 remains Anthropic’s most capable broadly available model,
• most launch-day benchmarks came from Anthropic or its partners,
• every migration requires internal quality, cost, and security testing.

The most sensible strategy is not to replace every model with Opus 5. It is routing:

Sonnet 5
for fast, high-volume tasks

Opus 5
for difficult, agentic, multi-stage processes

Fable 5
for the hardest cases where maximum capability justifies the cost

18. Pre-deployment checklist

API

• □ The model ID has been changed to claude-opus-5.
• □ The SDK version has been checked.
• □ max_tokens has been verified.
• □ Default thinking has been accounted for.
• □ Thinking has not been disabled with xhigh or max.
• □ HTTP 400 handling has been tested.
• □ A fallback has been added.
• □ Prompt cache has been measured.
• □ Region and cloud provider have been checked.

Quality

• □ An internal evaluation set has been prepared.
• □ Opus 5 has been compared with Opus 4.8.
• □ Opus 5 has been compared with Sonnet 5.
• □ Correct task completion, not response style, has been measured.
• □ Large repositories have been tested.
• □ Code review has been tested.
• □ Frontend and image analysis have been tested.
• □ Long documents have been tested.
• □ Repeatability across several runs has been tested.

Cost

• □ Input tokens have been measured.
• □ Output and thinking tokens have been measured.
• □ Cache hits have been measured.
• □ Tool calls have been counted.
• □ Total task cost has been measured.
• □ Standard mode and Fast mode have been compared.
• □ low, medium, high, xhigh, and max have been compared.
• □ Subagent budget limits have been set.

Security

• □ Tool permissions have been reviewed.
• □ Destructive operations have been restricted.
• □ Production deployments and changes require approval.
• □ Prompt injection has been tested.
• □ Cybersecurity-classifier behaviour has been tested.
• □ Agent logging and auditability have been verified.
• □ The model’s self-verification has not been treated as an independent audit.