Cheaper per Token, Yet Pricier per Task: How Claude Opus 5.5 Pairs a Price Cut with Deeper Thinking
Anthropic released Claude Opus 5.5 on October 20, 2026. Per-token prices dropped about 20%, yet at max effort the cost per task was reported to rise by roughly 20% instead. Here's what's behind the price-cut headline, and how to tame it with reasoning-effort settings in your operations.
On October 20, 2026, Anthropic released Claude Opus 5.5, the new version of its flagship model. The pitch is clear: it surpasses the previous generation at agentic coding and knowledge work, while costing "40% less to run than Opus 5" on typical workloads. For developers who run coding agents day in and day out, it's an update worth paying attention to—but if you take away only the words "it got cheaper," your actual bill may catch you off guard. The key lies in the gap between price per token and cost per task.
The Price Tag Dropped, the Benchmarks Climbed
Let's start with the facts that aren't in dispute. Opus 5.5 lowered its per-token pricing from the previous generation. Here are the main rates confirmed at launch (per million tokens).
| Item | Opus 5 | Opus 5.5 |
|---|---|---|
| Input | $5 | $4 |
| Output | $25 | $20 |
| Cache read | ~$0.50 | $0.20 |
| Cache write | $6.25 | $5 |
Input and output are each down about 20%, and cache reads are down roughly 60%. The cheaper cache reads pay off most for unattended workflows that repeatedly load long contexts, and for long-running sessions that carry a lot of history.
On performance, Anthropic says the model beat Opus 5 across benchmarks for agentic coding, computer use, and knowledge work. It ranks near the top on SWE-bench Pro-style metrics, and some reports have it beating the GPT-5 family on Terminal-Bench 4.0 (the exact figures vary by evaluator). Output speed is also said to be about 30% faster than before, so perceived wait times should shrink as well.
Look at Cost per Task and the Story Flips
This is the point that deserves the most caution this time. Even though the per-token price dropped, there's a reported reversal in which the total cost to finish a single piece of work can actually go up. In Artificial Analysis's estimate, the cost per task when run at max effort came to about $13.04—roughly 21% higher than Opus 5's figure of about $10.79. The reason the total rises even as the unit price falls is simple: the new model thinks more deeply and emits more tokens. In the same estimate, output tokens per task ballooned from 137,000 to 333,000—more than double.
In other words, "40% cheaper" is a figure averaged out on the assumption of typical workloads, including the gains in token efficiency. Throw max effort at a hard problem and consumption jumps in exchange for deeper reasoning, so the bill per job can actually get heavier. We don't mean to sound alarmist, but it's worth keeping in mind that reading "lower unit price" as simply "cheaper" is risky.
The "Throttle Knob" That Matters at Adoption
What's provided to control this reversal is a per-task reasoning effort setting. In practice, the following distinctions double as cost management.
- High only for the hard parts, low for the routine: Assign high effort to design decisions and intricate debugging, and low effort to formatting, small fixes, and summaries, keeping token waste down.
- Lean on cheap cache reads: For subagents that reuse large contexts and for unattended workflows, the drop in read pricing translates directly into lower operating costs.
- Switch to measuring by total cost: Track spending "per task" and "per day" rather than per unit price, and learn the relationship between effort and consumption for your own workloads.
On top of that, the system prompts are said to have been rewritten in a more concise, direct style, with adjustments aimed at reducing verbosity in responses.
What Changes on the Ground in Development
Opus 5.5 is available through Claude Code, the Claude desktop app (after updating), and the API. The baseline for coding agents as "standard equipment" has stepped up a notch—but whether you actually reap the efficiency gains comes down, now more clearly than ever, to how the user designs their operations. The smarter the model gets, the more it thinks and the more it writes. Teams that can rein in that extra cost through effort settings and context design, versus teams that run on the defaults and get surprised by the bill, will see the value they get from the same model diverge sharply. Rather than stopping at the price-cut headline, we recommend measuring early what a single task actually costs you.
References: Anthropic Newsroom (Claude Opus 5.5 announcement) / AlphaSignal: Opus 5.5 tops coding benchmarks but costs $13 per task / MindStudio: Claude Opus 5.5 Benchmarks, Pricing, and Real-World Performance / LLM-Stats: Claude Opus 5.5 Launch
