Claude Opus 5 vs Fable 5 vs Sonnet 5: Which to Use

Claude Opus 5 is the model most people should use for serious work in 2026: it comes close to Fable 5's intelligence at half the price. Sonnet 5 is the fast, inexpensive option for volume. Fable, now at version 5.1, is the most capable family for long, difficult, asynchronous tasks where getting it right matters more than cost.

I run all three every week, across Claude Code, content pipelines and client brains. Here's how they compare, what they cost and the rule I use to choose.

Claude Opus 5, Fable 5 and Sonnet 5 at a glance#

Prices and dates come from Anthropic's announcements for Opus 5 and Sonnet 5 and its Fable page.

Claude Opus 5: the default for real work#

Anthropic describes Opus 5 as "a thoughtful and proactive model" designed to be used every day. It's the default on Claude Max and was the strongest model on Pro when it launched.

The numbers behind the recommendation, all from Anthropic's launch post:

• On CursorBench at max effort, Opus 5 performed within 0.5% of Fable 5's peak score at half the cost per task.
• On OSWorld 2.0, a computer-use benchmark, it surpassed Fable 5's best result at just over a third of the cost.
• It's priced the same as Opus 4.8 while more than doubling its score on Frontier-Bench.

In practice, what stands out is verification. Opus 5 checks its own work and iterates until it succeeds, which is exactly what you want in long Claude Code sessions. Fast mode runs about 2.5 times faster at twice the price when latency matters.

Claude Fable 5 and 5.1: when you need the most capable model#

Fable is Anthropic's frontier family for "the hardest knowledge work and coding problems." Simon Willison's first impressions of Fable 5 sum it up well: slow, expensive and a beast. He also notes a 1 million token context window, 128,000 maximum output tokens and a January 2026 knowledge cutoff.

Fable 5.1 kept the $10 and $50 prices but cut cache reads to $0.25 per million tokens, 75% less than Fable 5. Anthropic estimates that reduces typical workload costs by about 25% and highly agentic workloads by up to about 45%.

Three things to know before you default to it:

• Safeguards reroute some requests. Many cybersecurity and biology queries go to Opus models instead, and you aren't charged Fable prices for them.
• Data retention. Using Fable requires 30-day retention for safety monitoring by default.
• Limits. It burns through plan usage faster, so save it for work that earns it.

Claude Sonnet 5: the fast workhorse#

Sonnet 5 is the most agentic Sonnet yet. Anthropic says its performance is close to Opus 4.8's at a lower price, and its introductory $2 and $10 pricing is now permanent. It's the default on Free and Pro plans.

Use it for high-volume and well-defined work: classifying, extracting, summarizing, simple edits, first drafts and subagents that do narrow jobs inside a larger workflow.

Which Claude model should you use?#

The rule I teach:

1. Start with Opus 5 for anything you'd give a capable colleague.
2. Drop to Sonnet 5 when the task is repetitive, well specified or runs thousands of times.
3. Escalate to Fable when a task is long, ambiguous or high-stakes, or when Opus has already failed at it.

Two refinements matter more than the model choice itself. First, adjust effort before switching models: higher effort often closes the gap on hard tasks. Second, put structure around the model. A good CLAUDE.md, skills and a verification step improve results on every tier. I cover that setup in how to use Claude Code.

How to switch between Claude models#

• Claude Code: run /model to choose the model for the current session, and adjust effort when a task is harder than usual.
• Claude apps: pick the model from the selector in the message box before you send.
• Claude API: set the model parameter, for example claude-sonnet-5, claude-opus-5 or claude-fable-5-1.

One habit worth building: write the model choice into the skills and scripts behind recurring jobs. That way a simple classification task never quietly runs on your most expensive model, and a hard review never runs on your cheapest.

What each model costs on a real task#

Take a session that reads 200,000 input tokens and writes 20,000 output tokens, without caching:

On a subscription you don't pay per token, but the same ratio shows up as how fast you reach your limits. For plan details, read Claude Code pricing. If you're weighing Claude against OpenAI's agent, see Claude Code vs Codex.

Where to go next#

Pick one recurring task this week and run it on Sonnet 5 and Opus 5 side by side. Keep the cheaper model wherever the output is equally good. Model choice, effort and verification are part of what we practice in my Claude Code course.