Claude Opus 5.5 Cache Reads at $0.20 Match GPT-6 Sol — Fresh Tokens Run 2x the Price
Claude Opus 5.5 launched September 22 with a cache read price of $0.20 per million tokens — a 60% cut from the $0.50 rate on Claude Opus 5. That figure is the same as GPT-6 Sol’s cached read rate, which OpenAI derives from a 90% discount on its $2 per million input price. The convergence matters because cache reads now dominate the bills for long-running agentic and coding sessions.
Anthropic’s data on Claude Code usage shows input-to-output ratios moving from 189:1 to 324:1 over six months, with context per request growing 2.6x. In that environment, the cached read rate carries more weight than the headline input price.
The full Opus 5.5 price table
| Token type | Opus 5.5 | Opus 5 | Change |
|---|---|---|---|
| Cache read | $0.20 / M | $0.50 / M | −60% |
| Input (fresh) | $4.00 / M | $5.00 / M | −20% |
| Output | $20.00 / M | $25.00 / M | −20% |
| Cache write (5 min) | $5.00 / M | $6.25 / M | −20% |
Cache read at $0.20 is a 95% discount to fresh input at $4.00. Writing to the cache costs $5.00 per million — more than a fresh read — so caching only saves money if the prefix is reused. The break-even is 1.26 calls per write. In steady state, any session above a 21% cache hit rate comes out ahead; most active Claude Code sessions run well above that.
Versus GPT-6 Sol and Luna
| Opus 5.5 | GPT-6 Sol | GPT-6 Luna | |
|---|---|---|---|
| Fresh input | $4.00 / M | $2.00 / M | $0.10 / M |
| Cached read | $0.20 / M | $0.20 / M | $0.01 / M |
| Output | $20.00 / M | $10.00 / M | $0.50 / M |
At the cached read level, Opus 5.5 and GPT-6 Sol land at exactly the same rate. The gap re-opens on every other token type: GPT-6 Sol charges half the Opus 5.5 price for fresh input and output. GPT-6 Luna is 20x cheaper on cached reads. OpenAI prices cache writes at 1.25x the base input rate: $2.50/M for GPT-6 Sol and $0.125/M for GPT-6 Luna. GPT-6 Sol’s single-prefix caching break-even works out to 1.28 calls per write, nearly identical to Opus 5.5’s 1.26.
At 90% cache hit rate, effective input cost converges
A session running at 90% cache hit rate has an effective input cost of $0.58 per million on Opus 5.5, calculated as (0.10 × $4.00) + (0.90 × $0.20) = $0.40 + $0.18. GPT-6 Sol at the same hit rate works out to $0.38 per million — (0.10 × $2.00) + (0.90 × $0.20) = $0.20 + $0.18. The gap narrows to $0.20 from the nominal $2.00 difference in fresh input price. Output pricing does not compress the same way: Opus 5.5 output at $20.00 is twice GPT-6 Sol’s $10.00 regardless of caching.
Anthropic estimates 40% total cost savings for typical Claude Code workloads on Opus 5.5 versus Opus 5. That combines the price cuts, the model’s lower token usage per task at default effort, and the improved cache hit rates in Claude Code’s harness. The effective bill reduction for teams already running high cache hit rates is closer to 60% on the input line alone.
Effort changes no longer reset the cache
One operational change that compounds the pricing: on Opus 5.5, changing the effort level during a session does not clear the cached conversation on the Claude API and Claude subscriptions. This lets teams dial effort up for a hard step and down again without paying a cache-write penalty on the full context. The exception is Amazon Bedrock and Google Cloud’s Agent Platform, where effort changes still clear the cache.
Typical enterprise Claude Code spend is running at approximately $13 per developer per active day across deployments Anthropic tracks, with 90% of users under $30 per active day. Both figures apply to current models.
Fast Mode for Opus 5.5 is available in Claude Code at $8/$40 per million input/output, providing up to 2.5x faster generation at twice the standard price. Switching Fast Mode on mid-session charges the fast-mode input rate on the entire conversation uncached; Anthropic advises enabling it at session start.