Opus 5.5 Cuts Cache Reads by 60%! I'm Breaking Down What a Claude Code Task Really Costs

Hi, I'm shiichan! I found a topic that anyone using Claude Code is going to want to know about: what a single task actually costs, broken down with real numbers. Let me walk you through it!

Claude Blog claude.com

What was announced?

Claude Blog published a post called "What a task costs on Opus 5.5." It's a deep dive into how much a task in Claude Code actually costs, backed by pricing tables and worked examples.

According to the post, Opus 5.5's API list prices are:

• Input tokens: $4 per million
• Output tokens: $20 per million
• Cache reads: $0.20 per million

Compared to Opus 5, input and output tokens are both 20% cheaper, and cache reads are a full 60% cheaper. The cache-read price dropped from a tenth of the input price to a twentieth.

The post also lists four things that set what a task costs:

• Turns (every turn resends the whole conversation, so more turns cost more)
• Cache reads (repeated content billed at a fraction of input price)
• Output token type (the most expensive tokens, and thinking counts as output)
• Model choice (sets the price of every token in the session)

Why it matters

If you use an AI coding agent every day, it's genuinely hard to see where your spend is going and why a session suddenly got expensive. Effort settings and model choice alone can swing the bill a lot, so having transparency into cost really matters for making good calls.

The post also shares a baseline: enterprise deployments average about $13 per developer per active day, with 90% of users staying under $30 a day. That gives you something concrete to compare your own sessions against.

What changes

Reading this post turns a vague "Opus 5.5 is cheaper" into something you can actually reason about.

• A same-token-count comparison between the two models came out about 31% cheaper on Opus 5.5
• You get concrete guidance on when to raise effort and when to switch to a bigger model like Fable 5.1
• The post shows how to check /usage on your own sessions, so you can trust your own numbers instead of guessing

Dive Deep

The post walks through some very specific math.

For a task that starts at 20K tokens of context and grows to 120K over 40 turns, the total input adds up to about 2.8M tokens. At a 90% cache hit rate, that costs about $1.62 in input; with zero cache, the same 2.8M tokens cost about $11.20. Cutting the task down to 25 turns brings it to about $1.02.

On output tokens: on Opus 5.5, one output token costs 100 times a cache read. A typical task's 60K output tokens cost about $1.20 - the same as reading 6M tokens from cache.

Effort has four levels - low, medium, high, xhigh (plus max for a single session). Medium is recommended for well-scoped daily work, high for when medium stalls, and low for mechanical work like renames. The post notes that if high adds 20K thinking tokens, that's about $0.40 on Opus 5.5 - roughly the same as a 10-turn retry loop - so high effort pays for itself if it saves you one retry.

Fable 5.1, the bigger model, lists at $10 per million input and $50 per million output (2.5x Opus 5.5), with cache reads at $0.25 per million (1.25x). The price gap shrinks the most on long, cache-heavy sessions.

Cache writes cost more than reads: 1.25x the input price for a 5-minute cache, 2x for a one-hour cache. Running /compact on a 150K-token session costs about $0.25, but saves roughly $0.025 per later turn in reads, so it pays for itself within about 10 turns.

An internal customer-support benchmark of 44 tickets showed moving from Opus 4.8 to Opus 5.5 at low effort cut costs by about 18%, and running /claude-api prompt-audit to remove outdated prompt patterns (like a mandatory six-step procedure) cut it a further 9%, for about 25% total.

Wrap-up

• Opus 5.5 makes input/output tokens 20% cheaper and cache reads 60% cheaper
• A same-condition session comparison came out about 31% cheaper
• Task cost comes down to four things: turns, cache reads, output tokens, and model choice
• The post gives concrete guidance on when to raise effort or switch to a bigger model like Fable 5.1
• Checking your own numbers with /usage and /compact is the top recommendation
• This one's for engineers who want to actually understand and control what their AI coding sessions cost!