uygnoey Choosing a Claude model class for a workload
Claude model classes do not specialize by domain. There is no finance model and no science model — every Claude model is trained to excel at coding, agentic tasks, and knowledge work. The difference between classes is how hard a problem the class can reliably carry, and what that capability costs in price and speed.
So model selection is not "which model is good at my field." It is "how hard is my task, and what am I willing to pay in latency and dollars to have it reliably done."
Instructions
1. Start with the most intelligent generally available model
The default recommendation is to start smart: begin on the most intelligent model class you can access, then use effort level to dial in performance and cost.
Two reasons this beats starting small:
- Cost-per-task is often lower on more capable models, especially at lower effort levels, even though price-per-token is higher. More capable models tend to take fewer turns and less thinking time to get a task right.
- Starting on a smaller model blurs your diagnostics. When something fails, you cannot easily tell a model failure from a setup failure.
Then, as latency- or cost-sensitive use cases arise, test lower classes until you find the fit.
The opposite direction is also legitimate: some organizations start with the most cost-effective model and move up classes until the quality bar is met. Both directional approaches are supported. Pick one and be consistent, so the results of your tests are comparable.
2. Know what each class is for
See references/model-classes.md for the full descriptions, including the Mythos vs. Fable packaging distinction and the rule of thumb for Opus vs. Fable.
Short form:
| Class | What it is for |
|---|---|
| Mythos / Fable | Most capable class, frontier across domains. Coding, long-running agent tasks, and problems AI has not reliably handled before. |
| Opus | Powerful class for reasoning-intensive enterprise tasks. Leading on GDPval-AA (knowledge work) and Terminal-Bench 2.1 (agentic coding). |
| Sonnet | Versatile class for everyday tasks. Best balance of performance, cost, and speed for the widest set of general-purpose use cases, including high-volume sub-agents in multi-agent orchestration. |
| Haiku | Lowest cost, fastest class. Built for high-frequency workloads where latency and cost matter. |
3. Ask the four selection questions
Work through these in order for the specific workload. Full treatment with follow-up prompts in references/selection-questions.md.
- How hard is this task? If it typically takes a lot of time, involves multiple steps, or is previously unsolved, a more capable model class is appropriate.
- What are the latency needs? If the model sits in a high-frequency, customer-facing workload, Sonnet is often the best choice.
- What are the access constraints? Mythos is only available to organizations under Project Glasswing, and not all organizations make all model classes available to all roles.
- What are the unit economics? Higher production volumes may suit lower model classes, particularly where evaluations show those tasks are completed satisfactorily. Models are priced differently per token and have different price-per-task costs depending on capability and effort.
4. Use effort level as the second dial
Effort level also shifts the balance of quality, speed, and cost. Two facts to hold together:
- Higher-class models at higher efforts give the best possible performance.
- Higher-class models at lower efforts can sometimes be more efficient than smaller models.
That second point is why "start smart and dial down effort" is often cheaper than "start small." Do not treat class and effort as one axis — search the grid.
5. Consider the advisor strategy before settling for a single model
You do not have to run the whole task on one class. In the advisor strategy, a faster, lower-cost worker model calls a more intelligent model to check its plan and evaluate its work. The executor is coached only when needed.
The reported result: on SWE-bench Pro, Sonnet 5 with a Fable 5 advisor is within 10% of Fable 5's score at 63% of the price of using Fable 5 for the whole task.
Details and the two roles in references/advisor-strategy.md; the
role definitions are also shipped as subagents in this post's agents/ folder.
6. Decide with evals, not with vibes
Standard benchmarks are useful directional guides across classes and providers, but they saturate: powerful models such as Opus and Fable solve nearly everything on the test, so the benchmark stops discriminating.
When that happens, run the models on real workloads or on your own evaluations — a curated set of problems drawn from production, including difficult tasks where your current tools fall short, with success criteria your team defines. That is where the capability and creativity of frontier models separate from the pack and from one another.
See references/evals-and-benchmarks.md.
7. Re-run the choice when conditions change
There is no one-size-fits-all model choice, which is why multiple classes exist. Selecting well means understanding the basics of each class and understanding your use case in depth — which in practice means building, maintaining, and deploying strong evaluations, then revisiting the choice when the workload, the volume, or the model lineup moves.
Examples
Worked scenarios in examples/selection-scenarios.md. Summary:
A previously unsolved research-style task, low volume. Question 1 dominates: multi-step, long-running, not reliably solved before. Start on the most capable class available. Volume is low, so unit economics barely bind. Tune effort down only if evals hold.
A customer-facing support assistant answering thousands of sessions an hour. Question 2 dominates: high-frequency and customer-facing, so Sonnet is often the best choice. Validate with a custom eval built from real production tickets before committing.
A high-volume classification or extraction step inside a larger pipeline. Questions 2 and 4 dominate: latency and cost matter, the task is narrow. Haiku is the class built for this shape. Confirm with an eval that the task is completed satisfactorily at that class.
Sub-agents inside a multi-agent orchestration. Sonnet is explicitly called out for high-volume sub-agents. Keep the orchestrator on a higher class if the planning is the hard part.
Opus clears the bar but you are tempted by Fable. Rule of thumb: if evals or internal testing show Opus struggling on some tasks, Fable is the answer. If Opus already clears the quality bar, its speed and price profile may make it the better choice.
Cost pressure on a workload that currently runs entirely on the top class. Before dropping a class outright, test the advisor strategy: worker on the cheaper class, advisor on the top class, invoked only to check plans and evaluate work.
Source
"Claude models explained: choosing the best model for your use case" by Michael Segner — published July 24, 2026.