#201 — Choosing the best Claude models for your use case

Opus and Fable both handle coding, agents, and knowledge work well, but Anthropic's own rule of thumb is simple: if Opus fails your evals, move up to Fable; if Opus already clears your quality bar, stick with it for the better speed and price. In practice, Fable tends to show more wisdom, creativity, and writing skill than raw benchmark scores suggest, so it's worth spot-checking outputs manually before assuming Opus is "good enough."

Haiku is built specifically for high-frequency, latency-sensitive workloads, which makes it a strong fit for routine, repetitive support tickets rather than complex escalations. The tradeoff is capability, not speed or cost — for anything requiring multi-step reasoning, Anthropic's framework points you toward Sonnet or higher instead.

Claude Mythos is restricted to organizations vetted under Anthropic's Project Glasswing, which covers dual-use cybersecurity and biology research work — it isn't available through a standard API signup. If your startup doesn't fall into that category, Fable gives you the same underlying model with additional public-safety guardrails instead.

Once frontier models like Opus and Fable start solving nearly every question on a standard benchmark — a pattern Anthropic calls saturation — the benchmark stops showing meaningful differences between models. Anthropic's recommended fix is to pull a curated set of real problems straight from your own production logs, including tasks your current tools already fail on, and grade them against success criteria your team defines.

The advisor strategy runs a cheaper, faster model as the primary worker and only calls in a smarter model to check its plan or grade its output, rather than running every task on the expensive model. Anthropic's own benchmark shows the payoff: on SWE-bench Pro, Sonnet 5 paired with a Fable 5 advisor lands within 10% of Fable 5's solo score at 63% of the cost of running Fable 5 on the whole task.

Anthropic's guidance leans toward mixing model classes by task rather than picking one model for the entire product — Sonnet for high-volume sub-agents in a multi-agent setup, Haiku for latency-sensitive front-end calls, and Opus or Fable reserved for the hardest reasoning steps. This is effectively the same logic behind the advisor strategy, just applied architecturally instead of just at runtime.

Effort level is a separate lever from model class entirely — a high-class model set to high effort gives you the strongest possible output, while the same model set to low effort can sometimes be more cost-efficient than a smaller model doing the same task. This means the real cost comparison isn't just Opus vs. Sonnet, it's Opus-at-low-effort vs. Sonnet-at-high-effort, which most teams never actually test.

This usually means you've hit saturation — a point where models like Opus and Fable solve almost every question on the test, so the benchmark no longer has room to show a gap between them. Anthropic's advice at that point is to stop trusting the leaderboard and switch to real workload testing or custom evaluations instead, since that's where frontier models actually separate from each other.

GDPval-AA is one of the industry benchmarks Anthropic uses to rank Opus-class models on real knowledge-work tasks, not just coding puzzles. If your use case looks like research, analysis, or document-heavy work rather than software, this benchmark is more predictive of Opus performance than generic leaderboards.

Terminal-Bench 2.1 is the benchmark Anthropic cites when ranking Opus-class models on agentic coding tasks — meaning tasks where the model operates a terminal or dev environment autonomously rather than just answering a prompt. Founders building coding agents or dev-tool products should treat this benchmark as more relevant than general chat benchmarks when comparing Opus and Fable.

Anthropic runs a dedicated Startups program separate from its standard API pricing, aimed at early-stage companies building on Claude. Founders evaluating model costs from this article should check that program before assuming they're stuck paying full API rates during the evaluation phase.

Claude models are priced per token, and price-per-task varies further based on capability and effort level, not just the base rate for that class. This is why Anthropic's own guidance warns against comparing models on token price alone — a cheaper-per-token model can still cost more per completed task if it needs more turns to get there.

Sonnet is Anthropic's recommended default for high-volume, high-frequency customer-facing workloads because it balances performance, cost, and speed. Haiku goes a step further on cost and latency but is positioned for the highest-frequency, most routine workloads specifically — the tradeoff shows up on tasks requiring more reasoning depth.

Anthropic explicitly recommends Sonnet for high-volume sub-agents inside multi-agent orchestration setups, reserving Opus or Fable for the harder reasoning steps in the same pipeline. This mirrors the advisor strategy's logic — cheap models do the volume work, expensive models get called in selectively — just applied at the architecture level rather than at runtime.

Saturation happens when a model like Opus or Fable solves almost every question on a given benchmark, which erases the ability of that test to distinguish between models. If your top two model candidates are scoring within a point or two of each other on a standard benchmark, that's a signal to switch to custom evaluations built from your own production data instead.

Both Mythos and Fable require limited data retention settings to be used safely, which is a compliance detail founders in regulated industries need to configure before deployment, not an optional setting. This matters specifically for teams in finance, health, or legal use cases where data retention policy is already under scrutiny.