Claude Haiku 5.5 vs Sonnet 5.5 vs Opus 5.5: 40× cheaper isn't 40× worse

Anthropic released Claude Haiku 5.5 on 7 October 2026. Two days later, everyone I coach was asking the same thing: do I still need the expensive model?

The best evidence so far comes from Nick Saraev, who spent more than $700 putting Haiku 5.5, Sonnet 5.5 and Opus 5.5 through 11 head-to-head tests: 3D games, a product website, a two-minute animated video, a personal finance app and a forensic fraud audit. Same prompt for every model, one shot each, real API bills.

Below is what he found, and how I'd turn it into decisions for your own work.

How much do Haiku 5.5, Sonnet 5.5 and Opus 5.5 cost?

Prices are per million tokens (roughly 750,000 words). Input is what you send the model; output is what it writes back.

Haiku 5.5 also has a 1 million token context window, up to 128,000 tokens of output and adjustable effort. It comes with one catch, the 100k cliff, which I explain further down.

What did the real bills look like?

Across seven of the builds, the totals on screen were:

Two things stand out. Haiku was also the fastest, at about twice Opus's speed. And Sonnet ended up almost as expensive as Opus, because it used more tokens to get to the same place. A cheaper price per token doesn't guarantee a cheaper job.

Which model won each test?

Opus won most of the visual builds. Sonnet beat Opus outright twice. Haiku won one and tied the one pure-text business task. Several tests were games, but games are a good stress test: graphics, logic, controls and story in one prompt.

What do the results actually mean?

1. Text work is saturated. On the fraud audit (massive spreadsheets, one clear answer) all three models found the same eight schemes. For reading, summarising, sorting and drafting, the cheapest model is now good enough most of the time.
2. Opus still owns visual and complex builds. 3D worlds, detailed websites, responsive game controls: the gap shows up the moment something has to look good and work across many moving parts.
3. Sonnet and Opus are closer than you think. Sonnet made the better finance app and the better animated video. If you default to the biggest model "just in case", you're often paying for nothing.
4. The quality gap is nowhere near 40×. Opus's builds were better, but not 40 times better. That changes what's affordable: work you'd never run on Opus because of cost becomes cheap enough to try.

The question isn't "which model is best?" It's "which model is best for this step?"

Plan with Opus, execute with Haiku: the pattern to copy

The most useful idea from the test is about combining models, not choosing one. A frontier model (Opus or Sonnet) does the thinking: it breaks the job down and decides what to do. Haiku is the hands. It runs the repetitive steps, many times, for almost nothing.

Jobs to hand to Haiku 5.5:

• Summarising reports, call transcripts, long email threads.
• Classifying support tickets, leads, invoices, CVs.
• Querying a database or spreadsheet and returning the answer.
• Extracting one number or field from thousands of rows, like the fraud audit.
• Research sweeps: reading 50 sources and pulling out what matters, so the bigger model only sees the summary.

Here's how it looks in real businesses. An events company gets hundreds of supplier emails and quotes a month: Haiku reads and tags each one, and the expensive model only steps in to draft the client proposal. A recruiter screens 300 CVs against a brief: Haiku scores each one, Opus writes the shortlist rationale for the top ten. Same quality where it matters, a fraction of the bill everywhere else.

If this sounds like a map of jobs and handoffs, it is. I call it graph engineering: you design the work first, then give each step to the cheapest model that can do it well.

The 100k token cliff: how to keep Haiku cheap

Once a Haiku 5.5 conversation passes 100,000 tokens, the price jumps 5×. But when you keep asking questions about the same material in one session, the repeated part is read from cache at a tenth of the price.

So the rule is simple:

1. Keep each Haiku thread short. One task, one batch of material.
2. Reuse the same context for follow-up questions so you benefit from cache reads.
3. Cut the thread at about 100k tokens. Have the bigger model summarise what Haiku produced, then start a fresh Haiku thread with that summary.

You don't have to manage this by hand. Tell your orchestrating agent the rule once and it will do it for you (there's a prompt for that below).

Which Claude model should you use for your task?

Pick the job you want to hand to AI:

Copy this prompt: sort your own tasks by model

Paste this into Claude or ChatGPT with a list of the things you use AI for. It gives you a routing plan you can act on today.

You are my AI cost advisor. Below is a list of tasks I use AI for each week.

For each task, recommend one model tier and explain why in one line:
- HAIKU 5.5 ($0.10/$0.50 per M tokens): summarising, classifying, extracting data, short drafts, high-volume repetitive steps.
- SONNET 5.5 ($2/$10): building apps and tools, important writing, most everyday work.
- OPUS 5.5 ($4/$20): planning, complex multi-step work, anything visual or design-heavy.

Rules:
1. Default to the cheapest tier that can do the task well.
2. If a task mixes planning and repetitive work, split it: Opus or Sonnet plans, Haiku executes.
3. For Haiku threads, keep each thread under 100k tokens: summarise and restart when it gets long.
4. Flag any task where a mistake would be costly, so a human checks it.

Return a table: Task | Model | Why | Split? | Human check?
Then tell me my 3 biggest savings.

My tasks:
[paste your list here]

How do you use Haiku 5.5 without being a developer?

• In a chat app: switch model per conversation. Use the small one for summaries and sorting, and switch up when the work gets complex.
• In Claude Code or another agent tool: set your sub-agents to Haiku and keep the main agent on Opus or Sonnet. That's the plan-and-execute pattern in one setting.
• In automations (Make, n8n, Zapier or an API call): pick Haiku for any step that runs hundreds of times, like tagging every new email.

If you're not sure which of your tasks fall where, that's exactly what we map in the first session of Personal AI Setup.

FAQ

Sources: Nick Saraev, "$700 test: Claude Haiku 5.5 vs Sonnet 5.5 vs Opus 5.5" (YouTube, October 2026): test set, results, costs and the plan-and-execute pattern. Anthropic's Claude Haiku 5.5 launch (7 October 2026) and Claude API pricing: prices, context window and the 100k pricing tier. Business examples and recommendations from my own coaching work. Prices as of 9 October 2026.