Choosing a 'Claude 5th Generation' Model: Anthropic's Official Recommendations for When You're Unsure
Hello, this is Horiuchi from Renewer Inc.
In the previous article, we decoded Anthropic's official 'New Rules for Context Engineering.' It was about what to write in a prompt and what not to write.
This time, as a follow-up, we will address another question: Which model should I choose?.
We will be decoding this official document:
Claude models explained: choosing the best model for your use case
Model selection is not just a matter of accuracy; it requires thinking that includes costs, making it similar to business decision-making.
For example, this article introduces the framework that 'starting with the cheapest model' can actually end up costing more.
This is a counter-intuitive way of thinking and is something that should be updated to master next-generation models.
With the intention of clearly summarizing and widely distributing such updates, we have released the 'Claude 5 Generation Master Guide'.
It is over 50 pages long and covers how to master the Claude 5 generation, focusing on official Anthropic articles, including 'New Context Engineering Techniques' and the 'Claude 5 Model Selection Guide.'
Continuing from last time, we have summarized it as a Japanese translation and commentary while replacing terms with easy-to-understand language.
The examples appearing in the guide have also been translated into a usable format in Japanese and re-edited for non-engineers.
You can download it for free from the link below.
1. Claude 5 Generation Model Overview: 4 Tiers + 1 Limited Model
First, let's organize the Claude 5 models.
The Claude model system consists of four tiers—Claude Haiku 4.5, Claude Sonnet 5, Claude Opus 5, and Claude Fable 5—plus the limited-release Claude Mythos 5. As the tier increases, intelligence and unit price rise, while response speed decreases.
Claude Haiku 4.5 — Fastest and lowest cost. Handles simple, high-frequency tasks like classification and routing.
Claude Sonnet 5 — Balanced for speed and intelligence. The default for daily tasks.
Claude Opus 5 — Advanced reasoning and analysis. The official recommendation is 'start here if you are unsure.'
Claude Fable 5 — Top tier. For long-running agents and unprecedented problems.
Claude Mythos 5 — Limited release. Not available to the general public.
2. Starting Point for Selection: 4 Decision Axes
When you actually want to execute tasks with Claude Code or Cowork, or when you are incorporating model usage into an AI service as an API, which model should you choose?
The official decision axes are as follows.
1. Task Difficulty — How much deep reasoning, planning, and creativity are required? Is the cost of failure high?
2. Latency — How much response speed is required? Is real-time performance essential?
3. Access Constraints — Are there restrictions on which models can be used due to data location or compliance?
4. Unit Economics — Is the cost per task worth it? What happens when the volume increases?
1, 2, and 4 represent the trade-off between performance, speed, and cost.
The top-tier Fable 5 is the model you want to choose when you have high-difficulty tasks to assign. However, the response is slower and the unit price is higher. Conversely, the lightest model, Haiku 4.5, is fast and cheap, but it may not be able to fully solve the user's problem.
The third point, access constraints, has a slightly different nature. In terms of Claude, Mythos 5 is not publicly available, so it cannot be chosen. In corporate adoption, the perspective of whether that model can be used due to compliance and security weighs heavily.
To be honest, this is not a particularly novel approach. On the other hand, models will continue to be replaced one after another. Opportunities to compare with other companies' models will also increase. To avoid having to rethink every time, it is better to have a fixed form of questioning. I perceive these four axes as having that nature.
3. Task Unit Price and Token Unit Price: The Most Important Topic in This Article
The main argument of this official guide is: Shift your accounting unit from 'tokens' to 'tasks'.
・Price-per-token ── The unit price per 1 million tokens listed on the price sheet
・Cost-per-task ── The total cost incurred to complete a single business task. This includes all multiple interactions, thinking, tool execution, and retries in case of failure
If you compare them on a price sheet, naturally, cheaper models look cheaper. However, the official guide states the following.
"Cost-per-task is often lower for more intelligent models, especially at lower effort levels, even if the price-per-token is higher."
(The cost-per-task is often lower for more intelligent models, even if the price-per-token is higher.)
Why does this happen? Three reasons are given.
①Smarter models finish tasks in fewer turns
Because the number of trial-and-error attempts is reduced. In actual measurements of financial modeling tasks, it is reported that Claude Opus 5 reduced the number of turns and tool calls by about one-third, shortened the required time by 60%, and improved accuracy by 9 points.
② Less wasted thinking
It requires fewer tokens to reach the same quality. Harvey, a legal AI, reported that Claude Opus 5 achieved performance equivalent to the previous generation's maximum inference settings with 26% fewer tokens on average.
③ Failures and retries are less likely to occur
If a task fails, you have to run the same process again. Even if the unit price is one-third, if you fail twice, the total cost will be reversed. And the rate of retries is not listed on the price sheet.
There is one more often-overlooked point. When a failure occurs with a smaller model, it takes human time to determine whether the cause is 'insufficient model capability' or a 'prompt issue.' The official guide also points this out.
"Starting with a smaller model can also make it harder to distinguish between model failures and setup failures." (Starting with a smaller model can also make it harder to distinguish between model failures and setup failures.)
If read somewhat critically, there are parts that feel like positioning talk from the model provider wanting you to use expensive models.
Nevertheless, verifying costs from this perspective should not be a waste. Trying to recalculate your company's billing amount not by 'token volume' but by 'cost per task.' Even just doing that will likely change your perspective significantly.
4. Start smart (Top-down approach)
So, how should you start when actually selecting a model?
The approach the official guide recommends first is start smart.
"start with the most intelligent generally available model and use effort level to dial in performance and cost" (Start with the most intelligent generally available model and use effort level to dial in performance and cost)
1. Confirm the upper limit of quality for that task using the highest-performance model × default effort (thinking depth setting).
2. Lower the effort first, within the range where quality is maintained.
3. If further reduction is needed, lower the model tier by one level.
Instead of climbing up from a cheap model, you descend from a smart model.
If you know the upper limit first, you can effectively isolate issues, such as 'This task should be able to do this much, so why did it fail?'
The point is the order: lower the effort before lowering the model.
Effort is a setting for 'how deeply to make it think,' allowing you to adjust speed and cost while keeping the same model. The official guidance for Claude Opus 5 also states, 'As long as quality is maintained, feel free to use low or medium as the primary control for cost and response time.'
This is a mindset that can be used as-is when facing the model selection menus of Claude Code or ChatGPT. First, get satisfactory results with the smartest model, then try dropping down to a lighter model. I feel that trying it in reverse order and concluding that 'this is all AI can do' is the most wasteful way to use it.
5. Advisor strategy ── 90%+ performance at 63% of the cost
However, it is not realistic to run everything on the top-tier model. That is why the official guide introduces the 'Advisor Strategy', which combines models.
The mechanism is as follows: A low-cost model leads the task as a worker (executor), and only in situations where it is unsure of a decision, it calls a high-performance model as an advisor. The official guide describes this as a 'coach-and-execute' type.
"Sonnet 5 with a Fable 5 advisor is within 10% of Fable 5's score at 63% of the price" (Sonnet 5 with a Fable 5 advisor is within 10% of Fable 5's score at 63% of the price)
Figures are showing over 90% performance at 63% of the cost.
In a human organization, this is like a staff member handling the practical work and only consulting a supervisor at key points. I find it an interesting conclusion that this division of labor is ultimately the most rational approach in the world of AI as well.
6. Ultimately, which model should you start with? ── The official answer is 'When in doubt, Opus 5'
You might be thinking, 'I understand the mindset and principles, but specifically, which one should I start with?'
The official documentation (Models overview) answers this question by name.
"If you're unsure which model to use, start with Claude Opus 5 for complex agentic coding and enterprise work. For workloads that need the highest available capability, use Claude Fable 5."
(If you are unsure which model to use, please start with Claude Opus 5 for complex agentic coding and enterprise tasks. For workloads that require the highest available capability, please use Claude Fable 5.)
When in doubt, start with Opus 5. This is the official answer.
Combined with the 'start smart' approach introduced earlier, the practical procedure is as follows.
1. First, confirm the upper limit of task quality using the default settings of Opus 5
2. Reduce effort within the range where quality is maintained
3. If you want to further reduce costs or increase speed, move to Sonnet 5; conversely, if it cannot handle the task, move to Fable 5
And the official documentation writes the conclusion on model selection as follows.
"the best way to select a model is to understand the basics of each model class and understand your use case in-depth. That means building, maintaining, and deploying strong evaluations."
(The best way to select a model is to understand the basics of each model class and understand your use case in-depth. This means building, maintaining, and deploying strong evaluations.)
Rather than a benchmark leaderboard, the way of thinking presented throughout this article is more useful for corporate decision-making regarding model selection.
Models will be replaced in half a year, but these decision-making criteria will remain as an asset.
Thank you for reading this far!
Ultimately, this discussion boiled down to the problem of how to measure AI costs.
Choosing based on low unit price is similar to hiring people solely based on low hourly wages. What actually matters is how much it cost to complete a single job. Once you can switch to this perspective, the feeling that 'expensive models are a luxury' should be reconsidered.
The guidebook 'Claude 5 Generation Master Guide,' which summarizes the content of this article in diagrams, can be downloaded for free below. In addition to the 'model selection' covered in this article, it also includes the previous theme, '6 New Rules for Context Design,' all in one volume.
Information on Claude that can be used for work and business is summarized in the magazine below.
Please read it as well. If you like, please follow.