Haiku, Sonnet or Opus? Choose the Reasoning Effort Too

Key takeaways

• Claude model selection has two variables: the model family and the effort level.

• Effort alone moves Haiku 5.5 from 2 cents to 21 cents per benchmark task.

• Higher effort can add minutes of latency. Opus 5.5 at max took about 72 times longer to produce its first token than at low on these benchmark tasks.

• A smaller Claude model often already passes a bounded task. Better context and examples can matter more than another jump in model size.

• A cheaper model earns the step by reproducing outputs people have already accepted from the real workflow.

Which Claude model should power an enterprise AI agent? The answer is usually more than one. The model that plans a workflow does not have to be the model that classifies a document, drafts a reply or checks the final output. At Beam, we choose models one agent step at a time.

A new step usually starts on a strong model. Early on, the prompt is still changing, the context may be incomplete, and we may still be learning what a good output looks like. A capable model gives us room to solve those problems before we start optimising the bill.

The expensive model does not get a permanent job, though. Once people have accepted enough outputs, those cases become the test set. We run the step again at lower effort and on smaller models, and keep moving down until the output changes in a way that matters.

This is where the neat Haiku, Sonnet, Opus ladder starts to break. Claude also lets you control how much effort a model spends on the task. In the benchmark data used for this article, Haiku 5.5 moved from 2 cents per task at low effort to 21 cents at max. Sonnet 5.5 went from about one second before its first token to six and a half minutes.

So we compared the model and effort level together. The result is a grid of useful combinations, five patterns behind the trade-offs, and a practical test for deciding when a working step is ready to move down.

What does the Claude model and effort grid show?

The grid shows that model size and effort do not form a clean ladder. Haiku at a higher effort can overlap with Sonnet at a lower one. Sonnet at max can cost more than Opus at a similar score. The useful unit is the combination of model and effort, tested on one specific agent step.

The Claude model and effort combinations worth testing first. Scores and costs: [Artificial Analysis](https://artificialanalysis.ai/models/releases/comparisons/claude-opus-5-5-vs-claude-sonnet-5-5), 19 October 2026.

Two comparisons show why the second axis matters. Haiku 5.5 at max reaches roughly the same score as Sonnet 5.5 at medium while costing less than half as much. At the top of the grid, Sonnet 5.5 at max costs more per task than Opus 5.5 at xhigh, even though their scores are similar.

The grid gives us a shortlist rather than an answer. A classification step may begin with Haiku, a well-specified writing task with Sonnet, and open-ended reasoning with Opus. The final choice comes from running those combinations against the same accepted cases.

How Claude model and effort affect cost, speed and quality

1. How much can Claude effort levels change cost?

Effort can change the cost of one model by roughly ten times. Haiku 5.5 costs 2 cents per benchmark task at low effort and 21 cents at max. Across model families at high effort, the same task costs 8 cents on Haiku, 88 cents on Sonnet and $1.82 on Opus.

Cost per benchmark task by Claude model and effort level. [Artificial Analysis](https://artificialanalysis.ai/models/releases/comparisons/claude-opus-5-5-vs-claude-sonnet-5-5), 19 October 2026.

The model sets the capability range. Effort controls how much reasoning that model is allowed to spend on the task. Both affect the bill, but changing effort is usually the cleaner experiment. The model, prompt and surrounding system stay in place, which leaves fewer variables to debug.

If the lower effort level still passes the accepted cases, the saving is immediate. If it fails, the failure itself is useful. It tells you whether the step needs more reasoning or whether the stronger setting had been covering for a weak prompt or missing context.

2. How much latency does higher effort add?

Higher effort can add minutes before the model starts answering. Sonnet 5.5 produced its first token in about one second at low effort on these benchmark tasks and took six and a half minutes at max. Opus 5.5 became about 11 times more expensive and 72 times slower to start between low and max.

Time to first token on hard benchmark tasks. [Artificial Analysis](https://artificialanalysis.ai/models/releases/comparisons/claude-opus-5-5-vs-claude-sonnet-5-5), 19 October 2026.

That makes effort a product decision as well as an infrastructure decision. A six-minute pause can disappear inside a background research job. In a voice conversation, it ends the conversation.

Low effort still includes thinking. Haiku 5.5 at low took eight seconds to produce its first token on these tasks. For interactive work, Anthropic exposes settings that remove the up-front reasoning phase: `disabled` on Haiku 5.5 and `between_tools` on Sonnet 5.5. Opus 5.5 does not offer an off setting (Anthropic). Max effort belongs in work where nobody is waiting and the evaluations show that the extra reasoning changes the result.

3. When does Claude Haiku perform well enough?

Haiku can be enough when the task is bounded, the output is easy to check and the context already contains what the model needs. In several small tests, larger models or higher effort added cost without improving the result. These comparisons use different methods, so they are useful clues rather than a shared leaderboard.

Four tests in which a cheaper Claude model or effort level remained competitive. Each result should be read within its own task and method.

• Coding with hidden tests. Sonnet 5.5 passed 40 of 40 tests at high effort and the same 40 of 40 at max, while max used 11 times the output cost (@stas_sorokin_ on X).

• Customer-service tool use. On τ²-Bench Airline, Haiku 5.5 scored 76.0% at about a cent per task, Sonnet 5.5 scored 78.7% and Opus 5.5 scored 75.5% (OpenRouter). A short, fixed policy gave the larger models little extra room to help.

• Everyday prompts. At max effort, Haiku 5.5 matched or beat Opus 5.5 on 22 of 24 prompts at roughly a fiftieth of the cost (ModelLineup).

• Work to a clear brief. Across seven real workflows at high effort, Sonnet 5.5 won four and Opus 5.5 won three, while Sonnet cost about half as much (Nate Herk).

Once a task is sufficiently bounded, more reasoning may have very little left to discover. A one-cent task that fails still creates review work, so the next step is to inspect the failures. A missing example, an ambiguous label or context the model never received may explain more than the model size.

Two studies show how far better inputs can move the result.

Feedback and labelled examples moved the result more than a stronger model in two separate 2024 studies. These illustrate the effect of context and training data rather than current model rankings.

On code, giving GPT-4o feedback from running its own tests lifted the score from 90.4% to 96.3%, ahead of the reasoning model o1-mini at 93.9% (Wang et al., 2024). On classification, small models trained on labelled examples beat zero-shot Claude Opus on all four tasks, by as much as 0.71 F1 (Bucher and Martini, 2024).

Before paying for more intelligence, make sure the model can see the examples, labels and feedback that define the job.

4. Can Claude Opus cost less than Sonnet?

Yes. Sonnet costs less per token, but it can use enough tokens at high effort to make the completed task more expensive. At a benchmark score of 56, Opus costs 37% less than Sonnet. At about 51, it costs roughly a third less. Token price and completed-task cost can point in different directions.

Claude model combinations with similar benchmark scores and different costs per task. [Artificial Analysis](https://artificialanalysis.ai/models/releases/comparisons/claude-opus-5-5-vs-claude-sonnet-5-5), 19 October 2026.

Lower down the grid, the order changes again. Around a score of 42, Haiku 5.5 at max costs less than half as much as either Sonnet at medium or Opus at low.

This is why we compare model-and-effort combinations rather than API prices. If a step needs Sonnet above high effort, Opus belongs in the test. If the quality bar is lower, Haiku at a higher effort may clear it for much less. The number that matters is the cost of an accepted result.

5. Writing: Sonnet when there is a spec, Opus when there isn't

Sonnet is a strong candidate when the brief carries most of the judgement: a fixed structure, clear examples and an explicit definition of done. Opus becomes more useful as the brief opens up and the model has to supply more taste. The evaluation should match the real writing task rather than a generic score.

Open writing compared with work to a brief. [ToneBench](https://benchmark.towardsai.com/compare/family-anthropic.html), 17 October 2026; [Nate Herk](https://www.youtube.com/watch?v=7eo-11K2e3c), 28 September 2026.

On ToneBench, Opus 5.5 at low effort scores above Sonnet 5.5 at xhigh while costing less. In the separate seven-workflow comparison, Sonnet won four tasks at about half the cost, while Opus performed better on the more open briefs.

A strong brief carries part of the reasoning. Templated documents, summaries and product copy with a clear definition of done are good candidates for Sonnet, and for Haiku at volume. Open creative briefs stay on Opus. A generic writing benchmark cannot tell you whether the output sounds right for your product or your audience.

How should you choose a Claude model for each agent step?

Start with enough capability to get the step working, then lower the cost against evidence from the workflow itself. Test effort before changing the model, and begin with high-volume steps where the saving can repay the validation work. This keeps model optimisation tied to accepted outputs rather than benchmark movement.

Start with enough capability to remove model uncertainty

A new step has several moving parts: the prompt, context, output format, tools and model. Starting with a capable model removes one source of uncertainty while the others are still changing. On this grid, Opus 5.5 at medium is the starting point for an open or still-unstable step.

The accepted outputs are the valuable part of this stage. They show what good looks like on the actual workflow and later become the cases a cheaper combination has to pass. This is also why agent evaluation has to be designed around the task rather than a generic quality score.

Test a lower effort level before a smaller model

Run the accepted cases again at a lower effort level on the same model. If they still pass, the step becomes faster and cheaper without a model migration. Anthropic describes effort as “often a better lever than switching models” (Anthropic).

If the lower setting fails, inspect where it fails before changing the model. Some failures point to missing capability. Others reveal that the prompt or context had been relying on the stronger setting to fill in a gap.

Optimise the steps with enough volume to matter

The cost of validating a change is mostly fixed. The return depends on how often that step runs.

Under the illustrative classification workload below, moving from Sonnet 5.5 to Haiku 5.5 saves about $85 a month at 10,000 calls and about $8,550 at one million. The same experiment can be marginal at one volume and obvious at another. This is where the economics of cheaper LLMs for AI agents become practical rather than theoretical.

The saving from a cheaper Claude model grows with call volume. Anthropic list prices, 19 October 2026. Token counts are illustrative.

When should you switch to a cheaper Claude model?

Switch when the step has a stable definition of good, enough accepted outputs to test against and enough volume to justify the work. The cheaper combination should pass the same cases at a lower completed-task cost. Treat the change as a behaviour change, with review and rollback rather than a silent configuration update.

We want five things to be true:

1. The step has run long enough to produce outputs people accepted.

2. Its output can be checked, whether it is a label, a field, a route or a document with a clear brief.

3. It runs often enough for the saving to repay the testing and maintenance.

4. The cheaper combination matches the accepted cases at a lower cost per task.

5. A person reviews the change, or every accepted case is rerun before the new setting takes over.

A model swap can change behaviour even when the interface and model name look familiar. In one published comparison, GPT-4's accuracy on a task fell from 84% to 51% between two versions carrying the same name (Chen et al., 2023). The regression set is what protects the workflow.

How we see it at Beam

We start a workflow with enough model capability to make the step work while its prompt and context are still being developed. Once the agent has run and people have accepted its outputs, we test cheaper combinations on one step at a time.

If a stronger model gives us the quality we need but costs too much at volume, its accepted outputs become the standard a smaller model has to reach. In an early classification test, a Haiku model reached Opus-level accuracy at around a fifth of the cost. The order matters: establish the quality, keep the evidence, then see how far the step can move down.

The right Claude model is the cheapest model-and-effort combination that still passes the cases your users have already told you are good.

FAQ: Claude Haiku, Sonnet, Opus and effort levels

What is the difference between Claude Haiku, Sonnet and Opus?

Of the three models compared here, Haiku is the smallest and cheapest, Sonnet sits in the middle, and Opus offers the highest capability. The useful difference depends on the task. Haiku suits bounded, high-volume work; Sonnet handles many well-specified professional tasks; Opus has more room for open reasoning and judgement. Effort levels can change the comparison.

Which Claude model should I use?

Use the smallest model-and-effort combination that passes examples from your real task. Haiku is a sensible candidate for classification and extraction, Sonnet for work with a clear brief, and Opus for open-ended reasoning or writing. Start stronger while the step is unstable, then test cheaper combinations against accepted outputs.

What do Claude effort levels change?

Effort controls how much work the model spends on a response. Higher levels can improve difficult reasoning, but they also use more tokens and can add substantial latency. The same Haiku model moved from 2 cents per benchmark task at low effort to 21 cents at max in the data used here.

Should I lower Claude's effort level or switch models first?

Test a lower effort level first. It keeps the model, prompt and surrounding system unchanged, which makes the result easier to interpret. If the accepted cases still pass, you get the saving without a migration. If they fail, inspect the failure and then compare the next model-and-effort combinations.