[July 2026 Edition] What is Claude Sonnet 5? How to use its near-Opus 4.8 performance at a lower cost

Hello, kazu here.

On June 30, 2026, Anthropic released Claude Sonnet 5. Since I regularly use Claude Code to handle my company's AI tasks, the moment I saw this news, I felt, "This is a huge deal for small businesses and individuals!"

And there is more good news. Access to Fable 5 begins tomorrow!

I'll start with the conclusion.

• Sonnet 5 is the "most agentic Sonnet yet," capable of making plans, using tools like browsers and terminals, and working autonomously.

• Its performance is very close to the top-tier Opus 4.8 model, but you can use it at a lower price.

• Furthermore, you can choose the "effort level" yourself, allowing you to adjust the balance between cost and intelligence for each situation.

Tasks that required expensive, large-scale models just a few months ago can now be handled at a Sonnet-level price. This is the biggest takeaway this time. I will look at the official announcement from the perspective of how it will actually work in the field.

Starting today, Sonnet 5 is available on the web, mobile, and desktop apps.

For those in a hurry

Here is a summary of the key points regarding Claude Sonnet 5.

Overview

• Sonnet 5 was announced as the "most agentic Sonnet model." Its ability to autonomously handle planning and tool usage (browsers, terminals, etc.) has been improved.

• Performance is approaching that of Opus 4.8, while the price is kept lower.

• Compared to the previous model, Sonnet 4.6, it has significantly improved in key aspects of agent performance, such as reasoning, tool usage, coding, and knowledge work.

Availability and Access

• Available on all plans starting today (June 30, 2026).

• It becomes the default model for Free and Pro plans and is also available for Max, Team, and Enterprise.

• It is also available in Claude Code and Claude Platform, and developers can call it via API as "claude-sonnet-5".

Price

• The introductory price is $2 per 1 million input tokens and $10 per 1 million output tokens (until August 31, 2026).

• After that, it will transition to the standard price of $3 for input and $15 for output.

• It uses a new tokenizer, which may result in a token count approximately 1.0 to 1.35 times higher than before for the same input, but the introductory price is set so that the cost remains largely unchanged.

Notes on Performance

• By adjusting the effort level, you can choose the balance between cost and performance. Cost efficiency is particularly improved for medium-level agent tasks, and at high effort levels, results comparable to Opus 4.8 have been achieved in some tasks.

Safety Evaluation

• Compared to Sonnet 4.6, the overall incidence of undesirable behavior has decreased, and safety during agent use has improved.

• The incidence of hallucinations and sycophancy has also decreased.

• On the other hand, automated behavioral audits showed slightly higher inconsistent behavior compared to Opus 4.8 and Mythos Preview (higher-tier models).

• The ability to perform cybersecurity-related tasks is clearly lower than that of Opus-series models, and no intentional cyber-related training has been conducted.

• Because its cyber capabilities are slightly higher than Sonnet 4.6, cyber safeguards at the same level as Opus 4.7/4.8 are enabled by default (though not as strict as Fable 5).

Other

• After the article was published, it was found that there was an error in the BrowseComp evaluation method; it has been re-aggregated using the correct method (10 million token budget, compaction, programmatic tool calling), and the graph has been corrected.

• With the revision of the evaluation criteria for Humanity's Last Exam and OSWorld-Verified, the scores for Sonnet 4.6 have also been retroactively updated.

Reference: https://www.anthropic.com/news/claude-sonnet-5

What has changed in Sonnet 5

Summarizing Anthropic's announcement in my own words, there are three changes.

• Autonomy has increased (it does not stop midway and completes complex tasks to the end)

• Compared to the previous model, Sonnet 4.6, it has significantly improved in reasoning, tool use, coding, and knowledge work.

• Closing the gap with the superior Opus 4.8 while keeping the price at the Sonnet level

Originally, it was the Sonnet series models (3.5, 3.6, 3.7) that created the gateway to the era of agentic AI.
However, recent growth has been concentrated in the superior Opus series, leaving Sonnet a step behind. This new Sonnet 5 has significantly closed that gap.

What is the benefit of being 'agentic' in practical work?

Since 'agentic' might not immediately click, let's translate it into a practical, on-the-ground perspective.

Conventional AI (conversational AI) was a tool that returned one good answer for every question asked.
It was smart, but the human side was always the one giving instructions.
You had to provide instructions from the human side every single time.

Agentic AI changes this.
When you hand it a goal, it continues to work through multiple steps on its own until it reaches that goal.
Along the way, it reads files, executes commands, checks results, and decides on the next move.
Early users have reported that 'complex tasks that would have stopped halfway with the previous Sonnet are now completed to the end by Sonnet 5' and 'it self-checks its own output without being asked'.

What does this bring to small companies and individuals? It means you can entrust entire tasks to it without having a human hovering over it to give instructions step by step.
I personally run my internal AI company (AI organization) using Claude Code, and I have increasingly been handing off tasks like organizing work logs and drafting reports by just providing the goal. The significance of being able to do this with a cheaper model may seem modest, but it is actually quite large.

On the topic of price (the substance of 'cheap yet smart')

Here is the pricing.

• Sonnet 5 (Introductory price, until August 31, 2026): $2 per 1 million input tokens, $10 per 1 million output tokens

• Sonnet 5 (Standard price after introductory period): $3 per 1 million input tokens, $15 per 1 million output tokens

• Reference - Opus 4.8: $5 per 1 million input tokens, $25 per 1 million output tokens

Compared to the superior Opus 4.8, it is less than half the price on the input side and significantly cheaper on the output side as well. Yet, its performance is approaching that of Opus 4.8.
This is the benefit this time around.
The model name on the API is claude-sonnet-5, and it can be used from the Claude API documentation.

Note that Sonnet 5 uses a new tokenizer (the mechanism that breaks text into tokens), so the number of tokens for the same text may increase by about 1.0 to 1.35 times. However, the introductory price is reportedly set so that the transition remains roughly cost-neutral even after accounting for this increase. Judging whether it is cheap or expensive by looking only at the unit price will lead to a slight misunderstanding.

A new concept called 'effort'

Personally, I think the concept of effort is what will be most effective for practitioners this time.

With Sonnet 5, you can choose the level of how hard it should try to think, all while using the same model.

• Setting effort to medium provides excellent cost efficiency

• Setting effort to high makes it comparable to the superior Opus 4.8 for some tasks

In other words, you can switch between working lightly and cheaply for light tasks, and having it think thoroughly only for heavy tasks.
Even on the official cost-performance graph, Sonnet 5 outperforms Sonnet 4.6 across the board while covering a wider range of cost and performance than Opus 4.8.

For small companies and individuals, this is close to the feeling of buying intelligence according to your budget. If you run everything at peak performance, it costs money, but most routine tasks only require a medium level of effort, allowing you to focus your power only where it really counts. That is the kind of operational approach you can take.

Safety (Points that business owners should care about)

As long as you are using it for business, safety cannot be ignored. I will look at this from a practical perspective as well.

• Sonnet 5 has a lower overall rate of undesirable behavior compared to the previous model, Sonnet 4.6.

• Resistance to refusing malicious requests and to prompt injection (hijacking by unauthorized instructions) has improved.

• Hallucinations (plausible lies) and sycophancy (a tendency to be overly agreeable to the user) have decreased.
  (It is great for Claude users that both have decreased!)

On the other hand, there is also an honest note.
Compared to the higher-end Opus 4.8 and Mythos Preview, the rate of undesirable behavior is reportedly slightly higher in Sonnet 5.
Rather than overhyping it, disclosing these limitations makes it more trustworthy from the user's perspective.

Regarding cybersecurity capabilities, Sonnet 5 is intentionally kept low. In evaluations involving the creation of exploit code to target software vulnerabilities, the success rate for creating fully functional exploit code was 0%.
Furthermore, a safeguard that detects and blocks dangerous cyber usage in real-time is enabled by default. Detailed evaluation results are summarized in the Claude Sonnet 5 System Card.

From the perspective of someone actually running agents every day

This is not an official announcement, but my own field experience.

As mentioned above, I run my company's AI organization using Claude Code, and I have multiple AI agents assigned roles to handle daily practical tasks.

I explained this in detail in this article.

What I feel strongly after this past year is that for small companies and individuals, the ability to run smart models autonomously at a low cost is more effective than the models simply becoming smarter. is the point.

The reason is simple: for our scale, the biggest constraints are money and time.
Even if you have a top-performance model, if the unit price is high, you cannot run it in large quantities. Conversely, if you can run a reasonably smart model cheaply without needing a human to monitor it constantly, you can handle the actual workload. Since Sonnet 5 brings low cost, intelligence, and autonomy all at once, I expect the trend of delegating daily tasks to agents to advance even further.

How to use Sonnet 5 and Opus 4.8

I will list my personal opinions on how to distinguish between them, as it can be confusing.

• For daily routine tasks (drafting reports, organizing information, most coding, research), use Sonnet 5 with a medium level of effort. It is cost-efficient and allows you to handle volume.
  

• For high-difficulty tasks where the cost of failure is high (important design decisions, complex long-term tasks), increase Sonnet 5 to a high level of effort or use Opus 4.8.

• For areas where a higher-end model is clearly necessary, such as security, choose Opus 4.8 directly.

The trick is not to try to do everything with one model. With the release of Sonnet 5, you can now adjust the cost and performance dials more finely. Take inventory of your work and decide which tasks to assign to which level. Just doing that will significantly change your monthly costs.

Summary

Three key points about Claude Sonnet 5.

• Achieves performance and autonomy approaching the superior Opus 4.8 while maintaining a low price

• Allows you to choose the level of effort, letting you adjust cost and intelligence for each situation

• For SMEs and individuals, the significance lies in being able to run smart AI cheaply without constant human supervision

You don't need to chase every new model that comes out. What matters is deciding which parts of your work to entrust to which model and at what level of effort. If you can design that, Sonnet 5 will be a cost-effective partner.

FAQ

Q. Where can I use Claude Sonnet 5?
A. It is available on all plans starting today. It is the default model for Free and Pro plans, and is also available to Max, Team, and Enterprise users, as well as in Claude Code and the Claude Platform. In the API, it can be called by the name claude-sonnet-5.

Q. Should I use Sonnet 5 or Opus 4.8?
A. For the majority of daily tasks, Sonnet 5 is sufficient. We recommend using Opus 4.8 for high-difficulty tasks where the cost of failure is high, or in areas like security where a superior model is clearly required. By increasing the effort level, Sonnet 5 can rival Opus 4.8 in some tasks.

Q. How much does it cost in the end?
A. The introductory price until August 31, 2026, is $2 per 1 million input tokens and $10 per 1 million output tokens. After that, it will be $3 for input and $15 for output. For reference, Opus 4.8 is $5 for input and $25 for output.

Q. Will the price increase with the new tokenizer?
A. Even for the same text, the token count may increase by approximately 1.0 to 1.35 times. However, the introductory price is set so that the transition is roughly cost-neutral even when accounting for this increase.

Q. Is it safe to use for business?
A. Sonnet 5 has a lower rate of undesirable behavior than the previous model and improved resistance to prompt injection. Safeguards that block dangerous cyber use are also enabled by default. However, it has been disclosed that the rate of certain behaviors is slightly higher than in the superior Opus 4.8, so it is safer to use them selectively depending on the application.

Click here if you want to learn about Generative AI and Dify through videos

In the Generative AI course I teach, I also explain how to build applications using Generative AI and Dify.

Generative AI Course (Basic)
Generative AI Course (Practical)
AI Talent Course (From the basics of machine learning and deep learning to AI agent development with Python)

Click here for project consultations

I support the construction of business efficiency agents utilizing AI, including Dify and n8n. I offer a free initial consultation, so if you are a corporation looking to build AI agents in-house, please feel free to contact me via the link below.

For inquiries, clickhere

X (formerly Twitter) is also available for inquiries.

Linktree

YouTube Channel

Additionally, I offer an AI advisory service (for individuals and corporations, 3-month one-on-one support) to provide ongoing, hands-on support for AI implementation. You can start with a 30-minute free consultation.Click here for details on the AI advisory service

Threads

If you found this article helpful, please give it a 'Like'! 🙇

Introduction to the One-Coin Membership

I deliver technical verification logs for tools like ChatGPT and Dify, as well as ready-to-use materials such as configuration files (YAML) and prompts, at least once a month. You can join for just one coin, so please consider signing up.

Generative AI Lab (Membership)

#GenerativeAI #AI #Claude #ClaudeCode #AIAgent #BusinessEfficiency #LLM #Automation #Prompt #TriedWithAI #sonnet5 #AINews