Published: 2026-10-08
Summary

Haiku 5.5 vs Sonnet 5.5 and Opus 5.5 on a dozen builds: what each cost, and where the gap was small

Chapters / key moments (click to jump — plays here on the page)

Nick Saraev spent over $700 giving Claude Haiku 5.5, Sonnet 5.5 and Opus 5.5 the same prompts: games, sites, an accounting-fraud audit and a finance app. Haiku cost a fraction (fifths to hundredths) and often looked close; Opus won the games. His routing advice: let a frontier model plan and use Haiku as the hands for summaries, classification and database queries.

Source video

"I Spent 24 Hours Testing Haiku 5.5: Here Are The Results" by Nick Saraev — Watch on YouTube →

The prices he starts from

Per million tokens, as he lists them: Fable 5.1 $10 in and $50 out; Opus 5.5 $4 and $20; Sonnet 5.5 $2 and $10; Haiku 4.5 $1 and $5; Haiku 5.5 $0.10 and $0.50. Anthropic's pricing page matches for Haiku 5.5 and adds the catch he explains at the end: those rates apply to prompts up to 100,000 tokens; above that a request pays $0.50 and $2.50. Our changelog entry has the full launch details.

What each build cost (as he states it)

Task	Haiku 5.5	Others	His verdict
3D MOBA in Three.js	under $10, 3 h 46 min, about 90M tokens	Sonnet roughly 5x the cost	Sonnet and Opus had far better graphics
Forensic-accounting fraud audit (planted fraud in large CSVs)	8 of 8 found	Sonnet and Opus also 8 of 8; Haiku to Sonnet about 20 cents apart, Haiku to Opus about 4 to 5x	Text analysis is 'saturated': no meaningful gap
Voxel sandbox (Minecraft-style)	about $0.20, 13 min, 2.2M tokens	Opus clearly best, with shadows	Haiku playable but plain
Personal-finance web app ('Ledgerly')	about $0.50, 17 min, 7M tokens	Sonnet about 5x the cost with a better UI; the other two used 4M and 3.39M tokens (captions don't say which is which)	Sonnet's app best; Opus not better
First-person shooter	$5.70	Sonnet $27, Opus about $20.78 (captions garble this figure)	Opus top
Two-minute 'history of flight' video (SVG, in code)	under $1	Opus about $7; Sonnet cost more than both	He would pick Sonnet's
Attack-on-Titan-style game	about $5, 72M tokens	Sonnet about $4.50 to $5 with fewer tokens	Haiku surprisingly good

All prompts are in a free page he links from the description. These are single one-shot runs, scored by his own eye.

His routing advice

Have Opus 5.5 or another frontier model plan the work; Haiku is 'the fingers' that do it.
Give Haiku summarisation of reports, classification of tickets, database queries and pulling one number from thousands (for example fraud analysis).
Mind the 100K-token price cliff: prompts past it cost five times more. His idea: for sub-agent research, cut a sub-agent off near 100,000 tokens, compact its output and start a fresh one. Cache reads are a tenth of the input price (about 1 cent per million up to 100K), so follow-up prompts in the same thread are very cheap.
Previously you would delegate research to a Haiku-class model at $1 and $5; the same job is now a tenth of that.

Gotchas & Caveats

Captions garble several dollar figures; the table keeps only the ones that read cleanly and flags the rest. Check his linked page for exact numbers.
One run per model per prompt: it shows typical cost and rough quality, not reliability. Games are judged by playing them for a minute.
Anthropic says code written for Haiku 4.5 can break on 5.5 (manual extended thinking returns a 400, adaptive thinking is on by default, and the same text counts as more tokens). The video does not cover migration.

Key Takeaways

On text-heavy analysis the three models tied, so default to the cheapest that passes your own check.
On games, UI polish and 3D, the premium models still win, but not in proportion to price.
Count tokens as well as dollars: Haiku used more tokens on several builds, which the lower price more than covers.
Test a model on your own top three jobs before routing by reputation.