Which Claude Model Should I Use?

Start on Opus 5.5. Turn effort down before you switch.

This page used to say "Sonnet for doing, Opus for deciding." The Claude 5.5 models changed the math. Anthropic's models page now says to "start with Claude Opus 5.5 for most workloads," and its Opus 5.5 guide reports that at the default medium effort it matched or beat Opus 5 at high on coding tasks, in fewer steps and fewer tokens.

So the dial you reach for first is effort, not the model. If a task is simple, /effort low on Opus 5.5 thinks less and costs less, and Anthropic says low comes close to Opus 5 at high on several coding evaluations. Switch to Sonnet 5.5 when you need speed or you're running a lot of the same small job, since it costs half as much per token.

From Anthropic's models overview, checked 8 October 2026. Opus costs 2x Sonnet ($4 / $2 = 2), Fable costs 2.5x Opus ($10 / $4 = 2.5), and on prompts up to 100K tokens Haiku costs a twentieth of Sonnet ($2 / $0.10 = 20).

In the desktop app, the model and effort sit beside the send button, and both are clickable. Click the model name to get the list of models. Click the effort level to open a slider that runs from low to max.

In the terminal, /model and /effort do the same job.

When to Use the Fable Class

Fable 5.1 is Anthropic's top model. Its models page says to use it "for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short." In plain terms: try Opus 5.5 at high first, and move to Fable only when that still fails. It runs extra safety classifiers, so a small share of edge-case requests come back refused. Anthropic's own page: anthropic.com/claude/fable.

Pricing: $10 / $50 per million tokens, the most expensive of the four.

When to Use the Opus Class

Opus 5.5 is the default, and for most work it should stay that way.

• Most coding, including multi-file changes and refactors
• Debugging where your first approach already failed
• Code review: Anthropic says early testers saw more bugs caught and fewer false alarms than on Opus 5
• Reading dense charts, diagrams and screenshots
• Writing PRDs, analysis and strategy documents

It also writes output more than 30% faster than Opus 5, per Anthropic's guide, so the old "Opus is slow" complaint mostly went away. Reserve xhigh and max for work where you've seen them help; Anthropic says the same.

Pricing: $4 / $20 per million tokens.

When to Use the Sonnet Class

Use Sonnet 5.5 when speed matters or the job repeats many times.

• Quick edits where you'll read the result straight away
• Bulk work: renaming across files, formatting, generating many similar things
• Background agents doing mechanical work
• Chat-style back and forth, where a fast first word matters

Two things to know. Sonnet 5.5's guide says it tends to add tests and docs nobody asked for, and that at low and medium effort it's more likely to stop and check in before a long task is finished. If either bothers you, raise effort or add a scope line (see prompting). For the hardest long-horizon work, Anthropic itself says an Opus model is the better choice.

Pricing: $2 / $10 per million tokens, half of Opus.

When to Use the Haiku Class

Haiku 5.5 is the cheap default for scoped, mechanical work where you don't need deep reasoning. Think of it like the intern who types really fast. Anthropic's launch post names compaction, summarization and subagent work, and says Sonnet 5.5 and Opus 5.5 "remain better choices for complex agentic coding tasks." On Terminal-Bench 4.0, Haiku 5.5 scored 39.2% against 70.6% for Sonnet 5.5.

• Quick code formatting
• Simple lookups ("what does this function do?")
• Running through checklists
• Basic text transformations
• Syntax checks

The real use case is sub-agents. When you're running background code review or spinning up a test-execution agent, Haiku is fast and cheap enough that you can run several in parallel without watching your bill climb.

Pricing: $0.10 input / $0.50 output per million tokens for prompts up to 100,000 tokens, the cheapest Claude model. A prompt over 100,000 tokens bills at $0.50 / $2.50, which is 5x more per token ($0.50 / $0.10 = 5), so keep Haiku jobs small. Against Haiku 4.5 ($1 / $5) that is 10x cheaper per token under 100K ($1 / $0.10 = 10). Anthropic puts the average saving at around 75%, not 90%, because Haiku 5.5's newer tokenizer counts the same text as about 30% more tokens.

Haiku 5.5 is also the first Haiku with an effort setting (default medium), and it holds the same 1M context window as Opus and Sonnet. Haiku 4.5 is no longer in Anthropic's current lineup. Sources: Introducing Claude Haiku 5.5 and the Haiku 5.5 model page, both read 8 October 2026.

Use Claude Code v2.1.293 or later. From that version, /model haiku and --model haiku run Haiku 5.5. On v2.1.288 we tested the alias and it still ran Haiku 4.5. The full ID claude-haiku-5-5 worked there, but it printed an unrecognized_model warning and billed the cache writes at Opus's $8 per million tokens instead of Haiku's $0.20, a cost report 40x too high ($8 / $0.20 = 40).

Run an eval before you move a job that has to catch things. We moved five scheduled background jobs from Sonnet with an eval each. Three passed twice on Haiku 5.5: a log alerter, a yield sweep and a weekly action list. Two stayed on Sonnet. A meeting coach on Haiku raised 4 false alarms on a full-meeting replay, against 2 on Sonnet. A Microsoft 365 sync on Haiku dropped a [truncated] marker in 3 of 5 runs, while Sonnet kept it in 5 of 5.

How Claude Code Picks For You

On Claude Pro ($20/month), Max, Team and Enterprise, Claude Code now defaults to Opus 5.5 at medium effort. Before v2.1.280, Pro started on Sonnet. /model sonnet now picks Sonnet 5.5, the default Sonnet since Claude Code v2.1.284.

On Claude Max (from $100/month) you get 5x or 20x the usage of Pro. Get here when you're hitting Pro limits daily, not before.

On the API (pay per use) you pick the model per request. Full control, full responsibility for the bill.

Switching Models Mid-Session

In the app, use the picker beside the send button, as in the session above. In the terminal, type:

/effort low     # Same model, less thinking, lower cost
/model sonnet   # Faster and half the price, for bulk work
/model haiku    # Cheapest, for simple background jobs
/model opus     # Back to the default

You can do this mid-conversation, mid-task. The context carries over. One cost to know: Anthropic's docs say changing effort between requests invalidates the prompt cache, so set it at the start of a session where you can.

Which Plan Should I Get?

If you're just trying it out: start free, see if you like it.

If you're using it regularly: Pro at $20/month handles most people's actual usage. Start here and upgrade only when you start hitting limits and it frustrates you. You'll know.

If you're a power user building things all day: Max from $100/month gives you the headroom, with 5x the usage of Pro, or 20x on the higher tier. Pro already includes Opus, so you're paying for the extra usage, not model access.

If you're building apps that call Claude: API, pay per token, pick your model per call.

The Decision in One Line

Opus 5.5 by default, effort as the cost dial, Sonnet 5.5 for speed and volume, Fable when Opus at high still falls short.

Start there and adjust based on what you actually see, not what you imagine might be better.

Pro Tips

1. Lower effort before you lower the model. /effort low on Opus 5.5 keeps the stronger model and cuts thinking. Drop to Sonnet when you need speed.
2. Use Sonnet or Haiku for mechanical sub-agents. Test runs and lookups in the background don't need Opus-level reasoning. Keep code review on Opus.
3. Watch your context window. The 1M token limit sounds massive, but loading whole codebases repeatedly adds up. Use /compact to trim.
4. You pay for thinking. Thinking tokens bill as output, even when Claude Code collapses them on screen. Effort is the dial: /effort low on Opus 5.5 thinks less and costs less. See thinking modes.