Claude Opus 5.5 Pricing: The Cost-per-Task Test for AI Founders

Anthropic's September 22 launch puts Opus 5.5 at $4 per million fresh input tokens and $20 per million output tokens. Both are 20% below Opus 5. Reusing context through a cache is the sharper cut: reads fell from $0.50 to $0.20 per million tokens. Five-minute cache writes fell from $6.25 to $5.00. These are API list prices, not a promise about a fixed-price subscription.

Opus 5.5 is available through the Claude API as claude-opus-5-5, Claude Code, and major cloud platforms. Anthropic also says it produces output more than 30% faster. It reports a 66.4% score on Terminal-Bench 4.0 at xhigh effort, versus 52.3% for Opus 5 in its setup. Those are vendor-reported evaluations with different settings, so treat them as a reason to run a trial, not proof your agent will improve by the same amount.

The Claude team's cost explainer uses an illustrative session with 2 million cache-read tokens, 200,000 fresh input tokens, and 60,000 output tokens. Hold usage fixed and the arithmetic is straightforward: Opus 5 costs $3.50; Opus 5.5 costs $2.40. That is about 31% less, before any change in task length or success rate.

Anthropic's 40% typical-workload saving includes its claim that the new model needs fewer tokens to finish work. That second piece cannot be inferred from the rate card. A task that requires another attempt may erase the discount; one that finishes in fewer turns may beat it. Output includes billed thinking tokens, so a longer answer can also move the bill.

In the Hacker News launch discussion, builders are comparing real sessions and debating whether lower token prices translate into lower bills. Their anecdotes are useful prompts for testing, not a representative measurement of every workload.

First, separate your agent's work into repeatable task types: a support resolution, a code fix, a lead qualification, or a research report. Pick a sample with easy and difficult cases. Run the same inputs on Opus 5 and Opus 5.5 with the same tools, instructions, and success criteria. If the result affects customers, use a safe test environment and review the outputs.

Record fresh input, cache reads and writes, output, elapsed time, tool calls, retries, and whether the task finished correctly. Calculate total spend divided by successful tasks. Add the time a person spent fixing or checking results. That is the number you can put in a margin model. For high-volume flows, compare the cost at the median and at the expensive tail; a few runaway tasks can dominate your bill.

Also measure quality. Anthropic's release cites Zapier's AutomationBench, which checks whether agents actually complete business workflows across apps. That is the right kind of outcome to copy in your own test. A cheaper call that updates the wrong record costs far more than its tokens.

Keep model selection configurable while you test. If Opus 5.5 wins on supervised feature work but struggles on unsupervised, high-stakes tasks, route those task types separately. Our earlier Sonnet 5 agent economics analysis explains why a lower rate card can still mislead. The broader founder task benchmark guide offers a way to compare models on work that resembles your product.

Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks. Their prices and real task performance will matter to founders deciding which work needs Opus. For now, the Opus 5.5 launch makes a familiar pricing mistake more expensive: treating a model announcement as a new unit-economics forecast.

Run a small evaluation before changing your default, then revisit your routing and customer margins with measured results. The opportunity is real if a better model finishes more work for less money. Your own completion data is how you know whether that opportunity belongs to your business.