Claude Opus 5 Released: What Has Changed in the API and Why It Matters to AI Agents

On July 24, 2026, Anthropic released Claude Opus 5. At the name level, this is the next step after Opus 4.8, but the main change is not in the next row of the benchmark table. The model became noticeably better in long chains: plan, call tools, check the result and continue the task after intermediate errors.

Opus 5 is already available through the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. The base price has not changed: $5 for a million input tokens and $25 for a million weekends. Therefore, the update is interesting not only for those who need the maximum result, but also for developers who build agents with a large number of steps.

## Claude Opus 5 in numbers

Fast mode is still a research preview and is available in the Claude API. The Anthropic documentation specifically states that Amazon Bedrock, Google Cloud and Microsoft Foundry are not yet available. Opus 5 is available through these platforms.

## The main upgrade - the behavior of the agent

For a common question, the difference between the two versions may not be noticeable: both models can write text, explain code, and analyze documents. The gap appears when the request turns into a workflow of dozens of steps.

For example, an agent needs to add a function to an existing project. It should find related files, understand the architecture, change several modules, run tests, disassemble the error, fix the implementation and check the result in the browser. The weak agent stops after the first successful assembly and declares the task completed. Opus 5 is more likely to continue until behavioral testing, not just until a plausible diff appears.

Anthropic describes the following most notable improvements:

• more stable reasoning on long chains;
• agential development and multifile changes without left stubs;
• search for real errors in code with fewer false positives;
• work with tables, documents, presentations, diagrams and interfaces;
• coordination of several agents;
• stronger result with increased force level, up to max.

An important nuance: the model often checks its own work itself. Therefore, old instructions like “be sure to check everything through a separate agent at the end” can lead to unnecessary re-checking and additional costs. In a prompt, it is better to describe the readiness criterion, rather than force the model to mechanically repeat the same step.

## That changed compared to Opus 4. 8

The most important change for existing integrations is the default reasoning. On Opus 4.8, requests without a specific option were usually executed without one. On Opus 5, the model itself chooses how much to reason, and effort becomes the main regulator of the quality, delay and consumption of tokens.

There is also incompatible behavior. If you transmit thinking: {"type": "disabled"} along with the effort: "xhigh" or effort: "max" layer, the API will return error 400. With reasoning disabled, only high and lower levels are allowed. For agent scripts, Anthropic recommends leaving reasoning on and reducing flow through low or medium if quality is maintained.

## What does migration look like in Claude API

The minimum change is to replace the model ID. But for a predictable result, it is worthwhile to simultaneously explicitly choose the level of effort and increase the max_tokens budget: it limits the total scope of reasoning and the usual response.

{
"Model": "claude-opus-5",
"max tokens": 64000,
"output config": {
"effort": "high"
}
"messages":
{
"role": "user,"
"Content": "Check the architecture of the project, make changes, run tests and explain what's left to check manually."
}
]
}

In this example, the thinking field is not specified intentionally: for Opus 5, adaptive reasoning is activated automatically. For relatively simple tasks, you can start with medium, for long-term refactoring or complex analysis with high. xhigh and max levels should be included only after checking on your own tasks: they can improve the result, but will increase the latency and consumption of tokens.

Before switching the model, check three places in the code:

1. The logic of processing the blocks thinking and tool_use in response.
2. Limitation of max_tokens, especially if earlier reasoning was turned off.
3. Error handling 400 and a backup model for queries that are blocked by security classifiers.

For queries where the toolkit changes over the course of the dialogue, Anthropic added a mid-conversation tool changes beta capability. It allows you to add or remove the tool between moves without sending a fixed list again and without losing the prompt cache. The name mid-conversation-tool-changes-2026-07-01 is used for inclusion.

Another beta option is automatic alternate routes. With the server-side-fallback-2026-07-01 header, the API can select the recommended model if the Opus 5 request is stopped by the security classifier. This is useful for production agents: model failure does not necessarily translate into a total failure of the user script.

## What the benchmarks say

According to Anthropic, Opus 5 shows particularly strong growth where it is not the individual answers that matter, but the completion of the whole task

These figures are useful for understanding the direction of development, but do not replace your own checks. Results have a test version, a specific level of effort, number of attempts, tool settings, and possible deviations due to protective classifiers. The model that passes CursorBench better isn’t necessarily cheaper or more reliable in your Telegram bot, CRM, or internal agent.

## How to check the update in your project

Secure migration does not begin with the replacement of the model string, but with a set of control tasks. Take 20-30 real queries from logs and divide them into several groups: code, documents, tool calls, errors, and long chains.

For each version, measure:

• percentage of tasks completed without manual intervention;
• the number of moves and calls of tools;
• the number of tokens and the cost;
• time before the first response and time until the final result;
• the number of corrections after the tests;
• the number of false claims that the task has already been completed.

For an agent with access to the file system and production services, add manual confirmation points. A more autonomous model does not abolish restrictions on rights, sandboxing, journaling and rollback. Particularly carefully check for actions that can change data, send an email, publish a code, or perform a financial transaction.

## Should you switch to Opus 5

The transition seems justified if the agent regularly works with large repositories, complex code review, multi-step research, tables, documents or a browser. In these tasks, the cost of one generation is only a fraction of the cost: a failed attempt, manual fix and re-start can cost more than the tariff difference.

For simple questions, short texts, and tasks where minimal latency is more important, a flagship model will not always be a rational choice. Start with medium, compare the result with the current version, and leave Opus 5 only where the increase in quality outweighs the additional expense.

Claude Opus 5 is an update that should be evaluated as a new basis for long-term agent processes. His strength is not that he answers smarter to any question, but that he more often holds the goal through a few actions, notices his own mistakes and brings the work to a verifiable result. This is what makes the release important for developers who use Claude not as a chat, but as a participant in the workflow.

## Sources

• Официальный анонс Claude Opus 5 от AnthropicXX
• Что нового в Claude Opus 5: документация Claude PlatformXX
• Обзор моделей и тарифов ClaudeXX
• Рекомендации по промптам для Claude Opus 5XX