Claude Sonnet, Opus and Haiku: differences and which to choose for work

Claude Haiku is the family geared towards high volumes and fast responses. Claude Sonnet combines speed and capability for everyday work. Claude Opus is designed for complex tasks, including extended ones. The range also includes Fable, intended for the most demanding reasoning and long-running agentic work.

But these are only the manufacturer’s starting guidelines: when choosing an artificial intelligence model to work with, the choice should always be verified against the actual task.

Our guide to Claude AI introduces the assistant, how to access it and its features. Here we focus on one decision: which model to put to work, on which task, and with what criterion for accepting the result.

For a business, the answer also depends on what happens after generation. A summary delivered in a few seconds may need half an hour of corrections. A slower analysis can save time if it retains the necessary data and makes missing information visible. Choosing a model therefore means measuring the entire path to a usable result.

Claude Sonnet, Opus and Haiku: the updated comparison

Any comparison between Claude models must state the version and the date. The characteristics of an earlier generation may differ from those of the model available today. This snapshot is verified as of 14 October 2026: Sonnet 5.5, Opus 5.5, Haiku 5.5 and Fable 5.1 (Anthropic) will be the focus.

Claude Haiku: when speed and volume matter

Claude Haiku is worth testing when the work is frequent, the expected output is well defined and errors can be caught. Anthropic introduced Haiku 5.5 on 7 October 2026, with a particular focus on repetitive, cost-sensitive tasks, including classification and rapid information processing (Anthropic).

Imagine a company that receives enquiries about deliveries, invoices and support. The first task might be to assign each message to a category, keeping a separate category for unclear cases. The expected response is short; the process already specifies who will take charge of the message.

In this test, Claude Haiku should be judged above all on the boundaries. An enquiry that mentions an invoice but disputes a delivery must reach the right person. A reputational alert must be able to break out of the ordinary categories. A fast model becomes useful when the system also recognises the cases that require a check.

The figure to watch is the rate of acceptable classifications on the sample, together with turnaround times and corrections. Price per token, on its own, says little about how much it will cost to handle a request routed to the wrong department.

Claude Sonnet: how to evaluate it in day-to-day work

Claude Sonnet can be put to the test on tasks such as summarising, preparing documents, analysing data and supporting programming. The Sonnet 5.5 documentation places it in the tier that combines speed and capability (Anthropic).

Take a set of meeting minutes. The task is to extract decisions, owners, deadlines and open points. Quality depends on very concrete details: keeping track of a postponed deadline, distinguishing a proposal from a decision, flagging that no owner has been named.

Claude Sonnet should be compared with the other models on the same material and with the same instructions. If the output is accepted with few corrections and within timeframes compatible with the process, we have a useful indication. If, however, the summary is elegant but changes the status of decisions, the test has identified a problem to be resolved.

For external-facing texts, the register also needs checking: does the recipient recognise the company? Can the claims be substantiated? Do the promises match what the business offers? These criteria make the comparison more meaningful than a personal preference for a well-turned sentence.

Claude Opus: when to try a model for more complex work

Claude Opus is a family worth evaluating when the task involves multiple steps, interrelated documents or extended programming work. Anthropic describes Opus 5.5 as a model for agentic coding and complex knowledge work; in its general guidance it names it as the starting point for many workloads (Anthropic).

One possible test is to compare a sales presentation, a specification and the delivery terms. The operational question is whether the promises made to the client are borne out in the other documents. The output should cite the relevant passages, the discrepancies and the points that need checking.

When comparing Sonnet and Opus, it is precisely this ability to preserve relationships that is worth measuring. If a delivery is subject to approval, a summary that drops the condition changes the meaning of the commitment. The reviewer must be able to trace where every relevant conclusion comes from.

Claude Opus can make mistakes too. Human oversight of AI systems remains necessary where the output influences important decisions or communications. A more capable model broadens what the work can achieve, while responsibility for approval must be assigned within the process.

Where Claude Fable fits

The choice between Sonnet, Opus and Haiku must also take Fable into account. The current documentation presents Fable 5.1 as an option for demanding reasoning and long-running agentic work. In its selection procedure, Anthropic suggests considering it when tests with Opus at high effort levels continue to prove insufficient (Anthropic).

For a business, this introduces a useful step: describing the shortcoming observed. Did the model miss a condition? Did it abandon a task? Did it draw a conclusion without supporting evidence? Did it require more corrections than expected? Only once the problem has been identified can we tell whether a different model is an appropriate answer.

A disorganised document, contradictory instructions or an overly broad objective can undermine the test with any model. Before increasing the capacity deployed, it is worth also checking the quality of the material and of the task assigned.

What Claude Sonnet, Opus and Haiku cost via the API

API pricing for Claude models is separate from app subscriptions. The table shows the base rates on the Claude Platform per million tokens, verified as of 12 October 2026. It excludes caching, batch processing, additional tools and third-party platform pricing (Anthropic).

A token is the unit by which the model’s processing is counted, and it does not always correspond to a word. For a real project, you need to look at the actual consumption of requests as well as the rate. The figures should be rechecked before a budget is approved.

Consider a simple simulation: 1,000 requests, each with 2,000 input tokens and 500 output tokens. At the base rates shown, the cost would be 7 USD for Sonnet 5.5 and 14 USD for Opus 5.5. The calculation covers 2 million input tokens and half a million output tokens; it excludes caching, tools, additional reasoning tokens and repeated requests. It is an arithmetical example, not a forecast of an application’s usage.

A difference of a few dollars can be outweighed by review time. Conversely, at high volumes, a model that passes the tests at a lower cost can deliver significant savings. The economic comparison must take account of consumption, accepted results and human intervention together.

How to choose a Claude model through a business trial

To choose a Claude model, we propose a sample of ten real tasks, using authorised data or data prepared for the test. It is a starting size that allows quick learning, to be expanded before entrusting the system with more varied or more significant work.

The sample must include ordinary requests and difficult cases. For a summary, we include a text with a missing deadline. For a classification, a message that belongs to two categories. For an editorial review, a commercial promise without supporting evidence. These cases show how the system handles the boundaries of the task.

Before starting the comparison, we define the acceptable result. In the minutes, for example, no deadline may be invented and every decision must be traceable in the text. We record the model version, instructions, materials and settings. If reasoning effort is configurable, it should be noted: comparing different settings without declaring them makes the result harder to read.

Repeating some tests helps to observe variability. Ten favourable results are not enough to guarantee future behaviour, but they allow the next test to be chosen with more information. Anthropic, for its part, recommends evaluations based on one’s own prompts, data and edge cases (Anthropic).

The model Bliss would choose (and how)

When comparing Claude Sonnet and Opus, as when choosing Haiku, the cost of an error matters as much as the complexity of the text. A short message to a client during a crisis may demand more care than a very long internal document.

For Bliss, the choice must therefore be tied to where the output is going. An internal classification, a draft text and an approved communication carry different consequences. The process must establish who checks, what data the system may use and when the work stops for a review.

This link belongs to the governance of AI investment and risk. The decision on the model becomes sound when someone can explain the expected result, show the test and take responsibility for its use.

Different models can also be combined: an initial classification handled by Haiku, subsequent processing by Sonnet and a hand-off to Opus for cases defined as complex. It is an architecture that needs testing. Routing errors and the cost of the additional steps must also be measured.

The ultimate goal is to obtain approvable work with proportionate resources. If a model costs less but increases the number of corrections, the savings need to be recalculated. If a more capable model improves a marginal task without producing observable benefits, that capacity could be deployed elsewhere.