A task in Claude Code is a loop: the model reads the conversation, calls a tool, reads the result, and repeats. Each trip around the loop is one request, and every turn resends the conversation so far. Four things set the cost: turns, cache reads, output token type, and model. At Opus 5.5 API list prices of $4 per million input tokens, $20 per million output, and $0.20 per million cache reads, a task that grows from 20K to 120K tokens costs about $1.62 in input over 40 turns at 90% cache reads, and about $1.02 in 25 turns. With no cache, that input costs $11.20. Output is expensive because thinking is billed as output; 60K output tokens cost $1.20, equal to reading 6M cache tokens.

Opus 5.5 changes prices and turns. Input and output tokens are 20% cheaper than Opus 5; cache reads are 60% cheaper, going from a tenth of the input price to a twentieth. On subscription plans, the lower price passes through to limits, so cached context goes about 25% further, and five-hour limits went up. The estimate that Opus 5.5 costs 40% less to run applies to typical token-billed workloads at default settings, combining lower prices with fewer tokens per task at medium default; it is not a token-price cut. On identical token counts, Opus 5.5 is about 31% cheaper from price alone. Long runs also end with a report that reduces reruns.

Effort is the main in-session lever. Opus 5.5 defaults to medium, one level below Opus 5’s default high. Low suits mechanical work, medium well-scoped daily work, and high when medium stalls; use xhigh or max only after measuring a gain. Higher effort adds thinking tokens to every turn but can save a retry loop: 20K extra thinking tokens cost about $0.40, similar to a ten-turn retry loop at 100K cached context. A cheaper first move is giving the model a test or build script to check its work. Changing effort mid-session keeps the cache on API keys and Claude subscriptions, but on Amazon Bedrock, Google Cloud’s Agent Platform, or a gateway it clears the cache and triggers a full write.

Model choice moves the bill more than effort. Opus 5.5 is the daily driver for supervised work. Fable 5.1 lists at $10 per million input and $50 per million output tokens, 2.5 times Opus 5.5, but its cache reads are only $0.25 per million, 1.25 times, so the gap is smallest on cache-heavy runs. Move up to Fable 5.1 for long unsupervised runs, novel problems, and large multi-subagent changes; switch back after. Use Sonnet or Haiku for lookups and log reading. Agent teams are experimental and use about seven times the tokens of a standard session in plan mode.

Caching and compaction shape long sessions. A cached read costs 5% of a fresh input token; a cache write costs 1.25 times the input price for a five-minute cache and twice for an hour at today’s prices. Cache lifetime is one hour on a subscription and five minutes on an API key or cloud provider; a subscription drops to five minutes once usage credits are drawn. A steady session keeps a high hit rate; changing tool definitions, system prompt, MCP servers, models, or effort on cloud/gateways breaks it. Every turn resends context, so 30 turns at 150K tokens spend about $0.90 on reads versus $0.12 at 20K. /clear is free, /compact costs about $0.25 at 150K and pays for itself within about ten turns; compact before a break, not after. Keep CLAUDE.md under 200 lines. Fast mode runs Opus 5.5 up to 2.5 times faster at twice the standard price, and it should be enabled at session start because the first request pays fast-mode input on the whole conversation.

Measure instead of extrapolating. Opus 5.5 requires Claude Code v2.1.280 or later, so run claude update first. Then run /usage or /cost at the end of a task for estimated list-price dollars, tokens, and cache share. Run the same real task on Opus 5 and Opus 5.5 across three or four tasks. Check that cache share is high, output does not dominate small changes, and total input is not many times the conversation size. The post reports an average enterprise spend of about $13 per developer per active day, with 90% of users under $30. On an internal 44-ticket customer-support benchmark, moving from Opus 4.8 to Opus 5.5 at low effort cut cost about 18%, and prompt-audit added a further 9%, but it is one benchmark, not a guaranteed number.