Anthropic unveiled Claude Opus 5 on July 24, 2026. While the name might suggest a simple version upgrade, the core of this announcement is the shift in focus from an era where “only the most expensive, top-tier models were powerful” to “high-performance agent models deployable in daily work.”
Opus 5 is not a model aimed at the absolute top-tier, long-term autonomous tasks like Claude Fable 5. Instead, it is positioned as a model that offers performance close to Fable 5 for practical tasks such as coding, knowledge work, document processing, search, computer usage, and long-term tool use, but at a lower cost. Anthropic explains that Opus 5 has been significantly improved over Claude Opus 4.8 and provides frontier-level intelligence at half the price of Fable 5.
This article summarizes the key points to observe in the Opus 5 announcement. Rather than just listing benchmark scores, let’s look at why Opus 5 is important for developers and enterprise users, what tasks it’s suited for, and how cautiously we should interpret the numbers.
One-Line Summary
Claude Opus 5 is closer to a “high-performance agent model for daily use.” While Fable 5 is the top-tier model for the most complex and long-term autonomous tasks, Opus 5 is designed to handle coding, knowledge work, document analysis, long-context tasks, and automation tasks at a more realistic cost.
Three changes are particularly noteworthy:
- It offers a 1M token context and 128k maximum output tokens.
- Thinking is enabled by default, and the depth of reasoning can be adjusted with the effort setting.
- The pricing remains the same as Opus 4.8: $5 per 1 million input tokens and $25 per 1 million output tokens.
This combination is quite significant. When a model becomes more powerful but the price remains the same, users don’t just get a “smarter model”; they can perform longer tasks, use more complex tools, and run more self-verification loops with the same budget.
Opus 5’s Target Position
Looking at Anthropic’s current Claude lineup clarifies the role of each model. The Claude Platform documentation recommends Opus 5 for starting complex agent coding and enterprise tasks, and Fable 5 if the highest available performance is required.
| Model | Position | API Model ID | Price |
| — | — | — | — |
| Claude Fable 5 | Highest long-term autonomous task performance | claude-fable-5 | Input $10 / Output $50 |
| Claude Opus 5 | Complex agent coding and enterprise tasks | claude-opus-5 | Input $5 / Output $25 |
| Claude Sonnet 5 | Balance of speed and intelligence | claude-sonnet-5 | Input $3 / Output $15 |
| Claude Haiku 4.5 | Fast processing and low cost | claude-haiku-4-5 | Input $1 / Output $5 |
The interesting point here is that Opus 5 is not the “most expensive model.” In the past, to get the best results, one invariably had to choose the top-tier model. However, Opus 5 is designed to handle complex tasks that practitioners encounter daily at a lower cost than Fable 5.
In other words, the significance of Opus 5 is less about “the highest-scoring model has arrived” and more about “the price point for high-performance models has come down enough to make them a default for daily work.”
What to Look for in the Benchmark Table
Comparing the benchmark tables shared by users and Anthropic’s official announcement, Opus 5 significantly outperforms Opus 4.8 in several areas. Notably, items like agentic terminal coding, knowledge work, ARC-AGI-3, OSWorld 2.0, and AutomationBench stand out.
Figure: Key benchmark comparison of Claude Opus 5, Fable 5, Opus 4.8, and GPT-5.6 Sol. Source: Anthropic Official Announcement
When interpreting the table, it’s better to look at “what types of tasks it has become stronger in” rather than just “how many it won.”
First, Opus 5 scored 43.3% in agentic terminal coding. In the same table, Fable 5 is presented at 33.7%, Opus 4.8 at 21.1%, and GPT-5.6 Sol at 34.4%. Terminal-based coding benchmarks assess not just simple coding problems, but the ability to use tools, observe failures, and retry. An increased score in this area means it’s closer to a “model that drives tasks to completion” rather than a “model that gets the answer right on the first try.”
Second, for knowledge work, it scored 1861 according to GDPval-AA v2. GDPval is a benchmark designed to evaluate tasks closer to economically valuable real-world work. More important than the number itself is that model evaluation is increasingly moving beyond math and coding problems to areas like reports, documents, business judgments, and analytical outputs.
Third, for ARC-AGI-3, a score of 30.2% was presented. The separate results page for the ARC Prize also states that as of July 24, 2026, Claude Opus 5 achieved the highest score of 30.2% in ARC-AGI-3. ARC-AGI type evaluations strongly aim to assess the ability to generalize and solve unfamiliar problems rather than merely recalling memorized knowledge from training data.
Fourth, for OSWorld 2.0, a score of 70.6% was presented. OSWorld evaluates computer-using agents performing long task flows in a real desktop environment. As agents increasingly work across browsers, terminals, document editors, and internal tools, such evaluations are likely to become more important.
However, it’s important not to forget that this table is based on the provider’s announcement materials. Benchmarks vary depending on conditions, harnesses, number of attempts, tool provision methods, inference effort, and cost calculation methods. Therefore, in practice, it’s necessary to verify separately whether “Opus 5 actually fails less in our specific task failure types” rather than just “Opus 5 outperformed in a certain public table.”
Most Practical Change in Opus 5: Thinking Enabled by Default
The most significant change in the Claude Platform documentation is that thinking is enabled by default. In Opus 4.8, it ran without thinking unless thinking: {“type”: “adaptive”} was explicitly set. In Opus 5, the same requests will now run with thinking enabled by default.
This change has quite a practical impact on developers.
client = anthropic.Anthropic()
with client.messages.stream(
model="claude-opus-5",
max_tokens=64000,
output_config={"effort": "max"},
messages=[
{
"role": "user",
"content": "Explain why the sum of two even numbers is always even.",
}
],
) as stream:
response = stream.get_final_message()
print(response)The key in this example is output_config={“effort”: “max”}. Opus 5 offers low, medium, high, xhigh, and max effort levels, with high as the default. For complex tasks, you can increase the effort to enable deeper reasoning, while for simpler tasks, you can lower it to reduce costs and latency.
However, caution is needed when interpreting max_tokens. According to the documentation, max_tokens is the total output limit, combining the thinking process and the final response. If you migrate workloads that previously used Opus 4.8 without thinking, you might need to re-adjust the output limit for Opus 5, as thinking will consume some of the token budget.
There’s also another breaking change. To disable thinking in Opus 5, the effort must be high or lower. Setting thinking: {“type”: “disabled”} with xhigh or max effort will result in a 400 error. This is an easily overlooked point during migration.
What 1M Context and 128k Output Mean
Opus 5 provides a 1M token context window as both its default and maximum. The output is up to 128k tokens based on the synchronous Messages API. This goes beyond simply “being able to input many long documents.”
In practice, the following tasks become possible:
- Reading multiple files from a large codebase together to determine refactoring directions.
- Comparing long policy documents, contracts, security standards, and operating manuals all at once.
- Accumulating logs, test results, and error messages generated by an agent during a task without losing context.
- Revising document drafts, tables, slides, and code changes multiple times within a single session.
However, a long context is not free. Inputting too much increases costs, and the more unnecessary context there is, the more likely the model is to miss important signals. Therefore, Opus 5’s 1M context should be seen not as “let’s put everything in,” but as “an option to maintain truly long tasks without interruption.”
Closer to an Agent Model, Not Just a Coding Model
The recurring word in the Opus 5 announcement is “agentic” rather than “coding.” Anthropic explains that Opus 5 is strong at maintaining long tool-use loops, implementing multi-file features, performing large refactorings, and completing end-to-end functional tasks without placeholders.
This differs from general code generation capabilities. Code generation is closer to “create this function for me.” Agentic coding is closer to “find and fix bugs in this repository, test them, re-verify broken parts, and then summarize the reasons for the changes.”
What’s important in real development work isn’t the brilliance of the first answer. It’s the ability to read logs when failures occur midway, change hypotheses, re-run tests, and distinguish between causes and symptoms. Opus 5’s emphasis on “self-verification” and “long-horizon tasks” directly connects to this.
This also has implications for those who run blogs, create lecture materials, or set up cloud security labs. The model’s true capability shines not in writing that ends with a single prompt, but in multi-stage tasks like research, drafting, code verification, table organization, image placement, and upload verification.
Automatic Fallback and Mid-Conversation Tool Changes
Along with Opus 5, two beta features were also announced for the Claude Platform.
The first is mid-conversation tool changes. You can change the list of tools to use in the middle of a conversation while maintaining the prompt cache. Previously, it was common to continuously send a fixed list of tools at the start of a session. However, in actual agent tasks, a search tool might be needed initially, a code execution tool in the middle, and only a document creation tool at the end. If tools can be flexibly changed while preserving the cache, the cost and design of long tasks will be different.
The second is default fallbacks mode. Instead of users manually managing fallback model lists based on safety classifiers or rejection scenarios, Anthropic’s recommended fallbacks can be applied. Especially when a specific request in Opus 5 or Fable 5 triggers a safety policy, it’s possible to create a flow that redirects to an appropriate model instead of simply blocking it.
However, these features require using beta headers. This means they are less about being features to thoughtlessly enable in all production systems immediately, and more about being features for teams building their own agent platforms or those wanting fine-grained cost optimization to experiment with.
Safety: Stronger, but Some Areas Require More Caution
An interesting aspect of the Opus 5 announcement is that safety is as big a topic as performance. Anthropic explains that Opus 5 showed the lowest misaligned behavior score among recent models in its automated behavioral audits and is less likely to be susceptible to misuse.
At the same time, Anthropic states that Opus 5 does not push the frontier of dangerous dual-use capabilities further. Specifically, they explain that Mythos 5 is stronger in offensive cybersecurity and long-term autonomous biological research. From a security perspective, the important sentence is that “finding source code vulnerabilities is allowed, but areas like binary-based vulnerability scanning, penetration testing, and exploit generation are blocked.”
This is somewhat nuanced for those in security work. Opus 5 can be useful for static code reviews, explaining vulnerability patterns, and defensive analysis. However, in high-risk scenarios like automated penetration testing or exploit generation, safeguards are likely to intervene.
Anthropic also explains that Opus 5’s cyber classifier is less restrictive than Fable 5’s, and flagged requests in Opus 5 will default to fallback to Opus 4.8 in Claude.ai, Claude Code, and Claude Cowork. This means Opus 5 is less a “model incapable of security tasks” and more a “model that opens up defensive tasks while more cautiously blocking risky execution steps.”
Migration Points Developers Should Immediately Check
The simplest change when moving from Opus 4.8 to Opus 5 is the model ID.
model = "claude-opus-4-8" # before
model = "claude-opus-5" # afterHowever, in practice, the following should also be checked:
| Check Item | Why it’s Important |
| — | — |
| max_tokens | With thinking enabled by default, the thinking process and response share the same output budget. |
| output_config.effort | This is a key parameter to directly adjust the balance of cost, latency, and quality. |
| Thinking disable condition | Setting thinking disabled with xhigh or max effort will result in a 400 error. |
| Prompt cache minimum length | Opus 5 has a lower minimum cacheable prompt length of 512 tokens. |
| Fallback design | Decide whether to pass to another model or notify the user of interruption upon safety classifier rejection. |
| Evaluation criteria | Do not rely solely on provider benchmarks; perform regression tests with internal task samples. |
Especially if you are using it connected to an in-house agent or automation system, simply changing the model is not enough. The same prompt might yield longer responses, perform more self-verification steps, or change its tool usage patterns. While these could be positive changes, they might also lead to unexpected cost increases or redundant steps in finely tuned workflows.
Which Teams Should Try Opus 5 First?
Opus 5 seems particularly well-suited for the following teams:
For development teams, it’s worth trying for complex bug analysis, code reviews, multi-file refactoring, test enhancement, and understanding legacy codebases. While simpler code snippet generation might be sufficient with cheaper models, tasks that require following the entire repository context and self-verification are where Opus 5’s strengths become apparent.
For security teams, it’s suitable for defensive tasks such as policy document analysis, security checklist generation, code vulnerability reviews, cloud configuration audits, and incident report drafting. However, tasks where safeguards might intervene, like automated penetration testing or exploit generation, require separate verification and policy review.
For data and research teams, it can be used for summarizing long reports, comparing multiple documents, analyzing complex tables, and interpreting experimental results. The improvements in Opus 5 for document and spreadsheet tasks and long contexts align well with this area.
For operations and business automation teams, it’s worth testing automation flows that complete real-world tasks through multiple steps, as highlighted in Zapier AutomationBench. Typical tasks include checking account status, classifying risk signals, sending notifications, and generating summary reports.
Still, Benchmarks Are Just Benchmarks
While the Opus 5 announcement is impressive, it’s risky to directly translate benchmark numbers into purchasing decisions. Model evaluation is becoming increasingly complex. Even for the same model, results can vary depending on effort settings, tool provision methods, number of attempts, cost budget, presence of fallbacks, and intervention of safety filters.
Therefore, it’s best to start practical adoption with small internal evaluations:
- Collect 20-50 samples of tasks that frequently fail in your actual work.
- Compare Opus 4.8, Sonnet 5, and Opus 5 under the same conditions.
- Look not only at the accuracy rate but also at cost, latency, number of retries, and human correction time.
- Define the scope of use that aligns with your security, privacy, copyright, and internal data policies.
- Incrementally switch the default model, starting with tasks that yield good results.
The choice of AI models is increasingly moving away from “picking the smartest single model” to “combining effort and models per task.” Opus 5 is a model that quite well demonstrates this trend.
Conclusion: The Real Message of Opus 5
The real message of Claude Opus 5 is not “yet another frontier model release.” More accurately, it’s a signal that the price point for high-performance agent models is coming down, and the way models are used is shifting from one-off Q&A to long-term task execution.
Fable 5 remains the top-tier choice for the most complex long-term autonomous tasks. However, for daily recurring development, documentation, analysis, automation, and security review tasks, Opus 5 can become a more realistic default. This is because it offers a longer context, default thinking, effort adjustment, and enhanced agent performance at the same price as Opus 4.8.
However, the better the model, the better evaluation habits are needed. Benchmark tables are a starting point, and the final judgment should be based on whether it makes fewer mistakes, gets lost less often, and completes tasks more cost-effectively in our specific work. Opus 5 is a model worth thoroughly testing against those criteria.
Leave a Reply