Models & Cost

The model you run is the single biggest cost lever in Claude Code. On list price Opus costs 40x what Haiku does per token, and Fable 100x, so the default posture is simple: pick the cheapest model that clears the task, override it per project, and reach up to the top tier only for the parts that actually need it.

The lineup

Model	Tier	Reach for
Haiku 5.5 (claude-haiku-5-5)	fast, cheap	routine filtering, docs, reading transcripts, classification
Sonnet 5.5 (claude-sonnet-5-5)	mid	most coding and infra work
Opus 5.5 (claude-opus-5-5)	high, the Claude Code default	hard reasoning, long agentic runs, nasty refactors
Fable 5.1 (claude-fable-5-1)	frontier, above Opus	the hardest work, when Opus at higher effort still falls short
Fast mode	Opus, faster output	Opus quality when you want speed (costs more per token)

All four models have a 1M-token context window. Since 2.1.280 Opus 5.5 is the default model on every Claude plan, including Pro and Team Standard, which used to default to Sonnet. Haiku 5.5 (2026-10-07) and Sonnet 5.5 (2026-09-28) followed Opus 5.5 (2026-09-22). Before them the lineup was Opus 5, Sonnet 5 and Fable 5 (Fable 5.1 took over on September 1), and earlier Opus 4.8, Sonnet 4.6 and Haiku 4.5.

Fast mode is not a smaller model. It is Opus with faster output, toggled with /fast, and it carries a higher per-token price for that speed.

Pricing

Per million tokens, from Anthropic's pricing page (checked 2026-10-11):

Model	Input ($/1M)	Output ($/1M)	Cache read ($/1M)
Haiku 5.5, prompts up to 100K tokens	0.10	0.50	0.01
Haiku 5.5, prompts over 100K tokens	0.50	2.50	0.05
Sonnet 5.5	2	10	0.10
Opus 5.5	4	20	0.20
Fable 5.1	10	50	0.25
Opus 5.5, fast mode	8	40	·

Three things to know. Opus got cheaper: Opus 5.5 is $4/$20 against $5/$25 for Opus 5 and 4.8, and Sonnet 5.5 kept Sonnet 5's $2/$10 (Sonnet 5's planned rise to $3/$15 was dropped). Cache reads are cheaper than the usual 10%: 5% of input on Opus 5.5 and Sonnet 5.5, 2.5% on Fable 5.1, so a well-cached session costs far less than the input price suggests (Claude Code 2.1.296 corrected its own cost figures for Sonnet 5.5 cache reads from $0.20 to $0.10). Haiku 5.5 is priced by prompt length: a request whose prompt (cache reads included) is over 100K tokens pays five times the rate on every token in it. The other models charge the same rate across the full 1M window.

Fast mode on Opus 5.5 is $8/$40, twice the standard rate, and on Pro, Max, Team and Enterprise it bills usage credits rather than your plan's included usage. The "$15 / $75" figures that still circulate are Opus 4 and 4.1 prices, now retired on most platforms. Standing rule: re-check the official pricing page before you size a real budget, since prices move.

The levers

Per-project model override. The highest-value setting: run Haiku in a docs repo, Sonnet in infra, Opus only where it matters. Set it once per project and stop overpaying by default.

/effort. Effort is now set per model, and the 5.5 models start at medium. Raise it for the hard problem and drop it again; higher levels think longer and spend more tokens. The Agent tool takes an effort parameter too (2.1.292), so cheap subagents can run low.

Ultracode. A separate toggle in /effort since 2.1.284. With it on, Claude plans a multi-agent workflow for each substantive task at whatever effort you are on, so every request uses more tokens and takes longer. Turn it on for a stretch of hard work, then off.

Multi-model within a session. You can mix tiers inside one job: a cheap model handles routine sub-steps (reading a long transcript, filtering) while Opus or Sonnet does the actual work. This keeps the expensive model off the boring tokens.

The cost reality

One user ran the same 2.3-hour session (372 turns, 57.9M tokens) on the mid-2026 lineup: roughly $8.57 on Haiku, $25.71 on Sonnet, and $290.41 on Opus. Same work, a 30x spread. Prices have shifted since, but the gap between tiers is as wide as ever. That gap is the entire argument for defaulting down and reaching up only on purpose.

Decide: which model for which work?

The work	Model
Routine filtering, doc edits, reading long inputs	Haiku
Everyday coding, infra, refactors that are not subtle	Sonnet
Hard reasoning, subtle or large refactors, tricky bugs	Opus
The hardest problems, when Opus at high effort still falls short	Fable 5.1
Opus quality but you need it fast	Opus + fast mode

Recipes

1. Set the per-project default. In a low-stakes repo (docs, content), default to Haiku or Sonnet. Reserve Opus for the repos where reasoning quality pays for itself.

2. Default down, reach up. Start a task on Sonnet. When you hit the genuinely hard part (a nasty refactor, a subtle bug), switch up to Opus or Fable for that stretch, then drop back. Note that Claude Code now defaults to Opus on every plan, so defaulting down is a choice you make with /model.

3. Multi-model session. Let a cheap model read the transcript or filter candidates each turn, and keep the expensive model for the work that needs it.

4. Ultracode for a real wall. When a problem warrants multi-agent effort, put ultracode in that one prompt, or turn the toggle on with /effort ultracode, get the answer, and run /effort ultracode off. It is a tool, not a setting to leave on.

Failure modes

• Opus everywhere. Opus is now the out-of-the-box default, so this is what happens if you never touch /model. For routine work it pays up to 40x Haiku's rate for nothing. Set a cheaper per-project default and escalate by exception.
• Reseller proxy "discounts." Posts promising "save 92% on Opus" are third-party proxies routing your traffic, not real pricing. Avoid them; you give up trust and control for a number that is not what it claims.
• Unbounded autonomous Opus loops. A detached loop on Opus can spend serious money while you sleep. Put a spending cap in the job itself. See Loops & Autonomy.
• Planning from secondhand prices. Model and token prices move. Size budgets from the official pricing page, not from a post or from memory.

Quick reference

You want	Reach for
Cheapest viable model	Haiku, by default in low-stakes repos
The everyday workhorse	Sonnet
Hard reasoning / subtle refactor	Opus
The frontier tier	Fable 5.1
Opus quality, faster	fast mode (/fast)
Stop overpaying by default	per-project model override

Cost cheat: default down, reach up on purpose, cap any autonomous run, and verify every price against the official page.

Reference pages already exist for model selection and permissions & settings; this guide is the opinionated operating layer over them.