Which Claude Model to Use: Anthropic's Own Guide

§ 01The short answer: start on Opus 5.5#

Anthropic’s reason is short: “There’s a reason it’s the default model in Claude Code.” Opus 5.5 costs half as much per token as Fable 5.2, and the guide recommends it as the starting point for most coding and knowledge work. From there, it gives three moves:

• Down to Sonnet 5.5 for well-scoped everyday work, such as a bug fix or a feature built against a spec.
• Down again to Haiku 5.5 for simple high-volume work: classification, routing, extraction, and subagents that handle a larger model’s side tasks.
• Up to Fable 5.2 for the hardest technical projects in coding, finance and the sciences, when Opus 5.5 at high effort still falls short, or for long unattended runs.

§ 02Which model for which task#

A clear spec does not always mean Sonnet. The guide notes that a migration or a performance optimization can be well specified and still hard enough to need Opus 5.5, or Fable 5.2 for a major one.

§ 03What each model costs#

Haiku 5.5 is the only model whose price depends on prompt length. A prompt over 100K tokens bills at five times the lower rate, which is still a quarter of Sonnet 5.5’s per-token price. Every model has the same 1M-token context window, so context size is not a reason to pick one over another.

§ 04Change effort before you change models#

The guide’s second rule: “Before you move models, try adjusting your effort level.” Effort controls how long the model thinks, how many tools it calls and how much it checks its own work. It is a finer dial than the model name, and past high effort one more level can cost as much as a bigger model.

Anthropic gives its own numbers for Opus 5.5 on SWE-bench Pro, measured against the high setting:

• Low: about 8 points lower, for about a third of the cost.
• Medium: about 2.5 points lower, for about 70% of the cost.
• xhigh: about 1.4 points higher, for 2.5 times the cost.

When a model is not good enough for a task, the guide says to take three steps in this order:

1. Check that the model can verify its own work, with a test, a build or a script.
2. Raise effort one level.
3. Move to the next model up, starting at a lower effort than you used on the smaller one.

There is also a cheaper pattern for work you can check automatically. Run everything at low effort and re-run only the failures at high. On the same SWE-bench Pro subset, Anthropic says Opus 5.5 run that way passed about 97% of tasks for about $0.17 each, against 95.3% for $0.29 with every task at high.

§ 05When the bigger model is the cheaper one#

“A higher price per token doesn’t always mean a higher cost per task, or a better result,” the guide says, and it gives three results from Anthropic’s own runs:

• Terminal-Bench 3.0: the previous Fable, 5.1, at max effort scored 58.0%. Opus 5.5 at high scored 58.9% with half the tokens.
• FrontierCode 2.0: Fable 5.2 at high scored 45.3%, about level with Opus 5.5 at max (45.5%), for roughly half the cost per task. So before you raise Opus 5.5 past high on a hard task, test Fable 5.2.
• A 20-problem search benchmark: two problems carried 43% of the spend. Price your hardest tasks, not only the typical ones.

§ 06Mixing models in one system#

The guide spends a third of its length on using several models together. The main points:

• Put subagents on smaller models. “Keep the expensive model’s context for the decisions, and let cheaper subagents do the verbose reading.” Small models suit work where a mistake is cheap to spot, such as finding files, running tests and reading logs. Code edits and judgment calls stay on the main model.
• Delegate bulk work. On an internal defect-search benchmark over about 22 million tokens of source code, an Opus 5.5 lead with up to 25 Sonnet 5.5 workers cost about a third less than Opus 5.5 alone and scored about the same.
• Be careful with the advisor tool. An Opus 5.5 executor at high with a Fable 5.1 advisor scored 90.1% at $2.92 per attempt. That is 1.7 points over Opus 5.5 alone for about 2.1 times the money, a gain Anthropic puts at the edge of run-to-run noise.
• Switch models at a natural break. The prompt cache belongs to the model that wrote it, so the first turn on a new model pays the cache-write price on the whole conversation.

The test Anthropic sets for any multi-model setup: “Keep a multi-model setup only if it beats every single-model point on that chart.” We covered Anthropic’s earlier advice on the same question in one agent or subagents.

§ 07What the guide does not settle#

Every number here comes from Anthropic’s own benchmark runs, some of them internal. None has been reproduced by an outside party in what we read. The guide says as much about its limits: published results get you a starting model, and the next step is to test the candidates on a sample of your own tasks.

Claude Mythos 5.2, which Anthropic calls its most capable model for cybersecurity and life sciences research, is not in the comparison. It is available only through Anthropic’s verification programs.

§ 08Where CellCog sits#

CellCog routes its agents to Claude Opus 5.5, the model this guide tells developers to start on. You do not pick a model in CellCog. You pick how deep an agent should go on a task (Flash, Core or Max), and the platform handles the rest.

§ 09Sources#

• Anthropic, Which Claude model should you build with? Fable, Opus, Sonnet, or Haiku, Addy Osmani, November 9, 2026
• Claude Devs on X (21:03 UTC, November 9, 2026; 584,000 views at 02:06 UTC on November 10)
• Anthropic, Optimizing for cost and intelligence, Claude Platform docs
• Anthropic, Claude Fable 5.2 and Mythos 5.2 launch post, November 9, 2026