Mythos, Opus, Sonnet, Haiku: Here's How to Pick the Right Claude Model!

Hey everyone, it's me! Today I found a really practical post from Anthropic's Claude blog. It's not a shiny new-model announcement, it's an official answer to a question I bet a lot of you have asked: which Claude model should I actually use?

Claude Blog claude.com

What was announced?

This is a post from the Claude Blog titled "Claude models explained: choosing the best model for your use case." Anthropic laid out how to think about picking between its model lineup, Mythos, Fable, Opus, Sonnet, and Haiku, instead of just guessing.

It's not a launch of anything new, it's a guidance piece. But with the model lineup growing, I think this kind of framework is exactly what a lot of teams need right now.

Why it matters

According to Anthropic, "what model should I choose for this workload" is one of the most common questions they get, and as more model classes and versions ship, the answer keeps getting more nuanced.

They also flagged something I hadn't thought about: if you start with a cheaper, smaller model, it gets harder to tell whether a failure is really the model's fault or just a setup problem on your end. That's a good reason to have a clear decision framework instead of guessing.

What changes

Anthropic's default recommendation is simple: start with the most intelligent model you have access to, then use the effort level to dial performance and cost up or down. More capable models often need fewer turns and less thinking time, so the cost per task can actually end up lower than a cheaper model, even when the price per token is higher.

When picking a model, they suggest asking four questions.

• How hard is the task? Long, multi-step, or previously unsolved problems call for a more capable model class.
• What are the latency needs? For high-frequency, customer-facing work, Sonnet is often the best fit.
• What are the access constraints? Mythos is only available to organizations approved under Project Glasswing.
• What are the unit economics? If evals show a lower-tier model handles high-volume work well enough, that's often the more efficient choice.

Cost-conscious teams can also flip the approach: start with the cheapest model and move up model classes until quality clears the bar.

Dive Deep

Here's a closer look at each model class.

• Mythos / Fable: Anthropic's most capable model class, shipped as two packages of the same underlying model. Mythos is for trusted organizations under Project Glasswing working on dual-use cybersecurity and biology tasks, while Fable adds extra safeguards for general public use. Both require limited data retention.
• Opus: built for reasoning-heavy enterprise work, and it ranks among the leading models on benchmarks like GDPval-AA (knowledge work) and Terminal-Bench 2.1 (agentic coding). Opus and Fable both excel at coding and agent tasks, but Fable tends to edge ahead on wisdom, creativity, and writing. The rule of thumb: if Opus struggles on your evals, try Fable; if Opus already clears your bar, its speed and price make it the better pick.
• Sonnet: the versatile, everyday model class, balancing performance, cost, and speed, including for high-volume sub-agents in multi-agent setups.
• Haiku: the lowest-cost, fastest model class, built for high-frequency workloads where latency and cost matter most.

One detail I really liked was the "advisor strategy": a fast, cheap model does the work, and a smarter model only steps in to check the plan or the output when needed. On SWE-bench Pro, pairing Sonnet 5 with a Fable 5 advisor landed within 10% of Fable 5's own score, at 63% of the cost of running Fable 5 for the whole task. That's a solid efficiency win.

They also touched on why standard benchmarks aren't always enough: top-tier models like Opus and Fable can saturate benchmarks by solving nearly every question, so Anthropic recommends building custom evaluations from real production workloads to make the final call.

Wrap-up

• Claude's model lineup has four classes: Mythos / Fable, Opus, Sonnet, and Haiku
• The default move is to start with the smartest model available and tune cost and performance with effort level
• Pick a model by weighing task difficulty, latency needs, access constraints, and unit economics
• Pairing a cheap model with a smarter advisor model can cut costs while keeping quality high
• Because top models saturate benchmarks, custom evals built from real workloads matter more than ever

This one's for engineers building with Claude who keep asking which model to use, and for team leads trying to balance cost against performance.