Choose a Claude model for coding: Sonnet, Opus, Haiku or Fable?

Short answer: use Claude Sonnet 5.5 for most coding. Switch to Opus 5.5 for the hard parts (architecture, large refactors, multi-layer bugs). Use Haiku 5.5 for small, repetitive work. Fable 5.1 costs the most; use it only when a cheaper model cannot do the job. With RocketKey's daily USD allowance, matching the model to the task decides how much you get done in a day.

Prices of the main models

USD per 1M tokens, from the Models page on 25 October 2026:

Model	Input	Output	Good for
Claude Haiku 5.5	$0.10	$0.50	Naming, commit messages, summaries, parallel subagents
Claude Sonnet 5.5	$2	$10	Everyday coding, multi-step agents
Claude Sonnet 5	$2	$10	Same price as Sonnet 5.5
Claude Opus 5.5	$4	$20	Design, large refactors, hard debugging
Claude Opus 5	$5	$25	Costs more than Opus 5.5
Claude Fable 5.1	$10	$50	Work the cheaper models cannot finish

All of them have a 1M-token context window and up to 128K output tokens. The full list, older models included, is on the Models page.

How many tokens per day does each plan buy?

The allowance is in USD, so the token count depends on the model. Estimates from the output price, rounded down:

Model	Mini ($3/day)	Starter ($8/day)	Pro ($20/day)	Max ($50/day)
Haiku 5.5	≈ 6M	≈ 16M	≈ 40M	≈ 100M
Sonnet 5.5	≈ 300K	≈ 800K	≈ 2M	≈ 5M
Opus 5.5	≈ 150K	≈ 400K	≈ 1M	≈ 2.5M
Fable 5.1	≈ 60K	≈ 160K	≈ 400K	≈ 1M

A real Claude Code session spends input tokens too (code, plus the conversation history sent again with every request), so the usable amount is lower than the table. Try your own numbers in the daily cost estimate on Pricing.

Which model when

Sonnet 5.5: the default. Writing and fixing code, reading a codebase, writing tests, multi-step agents in Claude Code, Codex, Cline or Cursor. When unsure, start with Sonnet.

Opus 5.5: when Sonnet falls short. Architecture, refactors across many modules, bugs whose cause spans several layers. Opus 5.5 costs less than Opus 5 and twice Sonnet 5.5, which suits "Opus plans, Sonnet builds": plan with Opus, then switch back to Sonnet to implement.

Haiku 5.5: small jobs in volume. 20 times cheaper than Sonnet. Good for subagents that search files, summarize logs, classify, generate sample data, and for batch API scripts. Cheap as it is, the plan's per-minute request limit still applies.

Fable 5.1: on purpose only. 2.5 times the price of Opus 5.5 and 5 times Sonnet 5.5. On Starter that is about 160K output tokens a day. Turn it on for the task that needs it, then switch back.

Switching models in Claude Code

Claude Code picks models through three slots, opus, sonnet and haiku. Two ways to choose:

1. During a session: type /model to switch mid-session, for example to Opus for one hard question and back to Sonnet.
2. Map a model to a slot: set ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_OPUS_MODEL or ANTHROPIC_DEFAULT_HAIKU_MODEL to a model ID from the Models page, for example ANTHROPIC_DEFAULT_OPUS_MODEL=claude-opus-5-5. How to set the variables is in the Claude Code guide.

Claude Code's cost documentation gives the same advice: Sonnet for most coding, Opus for complex architectural decisions, and model: haiku for subagents doing simple tasks.

In Codex, Cline and Cursor, the model is the model ID in each tool's settings. See Codex, Cline, Cursor.

Models that always reason use extra output tokens

Per Claude Code's documentation, Opus 5.5, Sonnet 5.5, Haiku 5.5 and the Fable models always use extended thinking, and thinking tokens are billed as output. For simple work, lower the effort with /effort in Claude Code to spend fewer thinking tokens. More ways are in Save allowance in Claude Code.

FAQ

Does switching models cost allowance? No. The allowance is USD per day; each request is charged at the price of the model it used.

How do I write the model ID? Use the exact ID from the Models page, for example claude-sonnet-5-5. A wrong ID returns error 404.

Can I use non-Claude models? The Models page lists what RocketKey offers. Claude Code supports Claude models only.