Claude Opus 5.5 pricing fell 20% on the list, and the same model can cost 4.5 times more per task depending on a drop-down nobody on your team has opened. The cut is a rounding error beside the setting. Set medium as the default, and make anyone who wants max name a reason.
Start with what Anthropic published. Its launch page lists Opus 5.5 at $4 per million input tokens and $20 per million output, against $5 and $25 for Opus 5. Cache reads drop to $0.20 from $0.50. Anthropic’s pricing docs show the same rates.
Key points
- How much did Claude Opus 5.5 pricing really fall? The price per token fell 20%. Anthropic also says typical workloads will cost 40% less because the model uses fewer tokens to finish a job.
- What does the effort setting do to a monthly bill? It multiplies it. More thinking means more output tokens, and output tokens are the expensive ones.
- Does GPT-6 Sol change the arithmetic? It changes the comparison, though less than OpenAI’s page suggests.
- Who should be allowed to pick max? Almost nobody, by default. We’d set medium for the whole team, allow high where a wrong answer costs real money, and allow max only with a named reason.
How much did Claude Opus 5.5 pricing really fall?
The price per token fell 20%. Anthropic also says typical workloads will cost 40% less because the model uses fewer tokens to finish a job. The first number is on a price list. The second is the company’s claim with no workload named, so treat it as a hypothesis until your invoice agrees.
The side options are worth a look. Fast mode, up to 2.5x speed, costs $8 and $40, double the standard rates. Batch work is half price at $2 and $10 per Anthropic’s docs. Paid Claude plans get higher five-hour limits and a saveable rate-limit reset, according to the launch page, though it gives no figures. On an API plan the list price is your number. On a seat plan the limits are.
What does the effort setting do to a monthly bill?
It multiplies it. More thinking means more output tokens, and output tokens are the expensive ones. On Artificial Analysis data reported by Digital Applied, per-task cost climbs steeply with each step up the dial. Here is that cost, and what it would mean for a team running 2,000 tasks a month (the monthly column is our arithmetic).
| Effort | Cost per task | Monthly at 2,000 tasks |
|---|---|---|
| Low | $0.55 | $1,100 |
| Medium (default) | $1.34 | $2,680 |
| High | $1.82 | $3,640 |
| Xhigh | $3.46 | $6,920 |
| Max | $5.98 | $11,960 |
Max used 119,200 output tokens per task against 25,700 at medium. That’s about 4.6 times the tokens and 4.5 times the cost, while the intelligence index score moved from 51.2 to 57.6. Paying 4.5x for roughly 12% more score isn’t automatically wrong. Some tasks justify it. Most email drafts and spreadsheet clean-ups don’t.
Look at the shape of the table, too. Going from low to medium costs about $0.79 more per task. Going from xhigh to max costs about $2.52 more for a step most work won’t notice. Each rung up the dial buys less than the one before and costs more. That’s an ordinary pattern in pricing, and it’s why a default matters more than a discount.
The gap between medium and max on that team is roughly $9,300 a month. It comes from a menu choice, not a contract, and no finance team would approve it if it showed up as a line item. A 20% list cut is worth having. The dial is worth more than four times as much.
Does GPT-6 Sol change the arithmetic?
It changes the comparison, though less than OpenAI’s page suggests. OpenAI’s announcement puts GPT-6 Sol at $2 and $10 and Luna at $0.10 and $0.50, each 50% below earlier promotional prices. The page carries no date, and Sol and Luna were available in ChatGPT Work and Codex, not regular Chat.
OpenAI’s own cost-per-task claims compare Sol with Opus 5, not Opus 5.5. The company says Sol scores 33.2% on AutomationBench at $0.27 per task, against 26.9% for Opus 5 at about 11.1 times the cost. That’s a vendor claim on one benchmark. And cheap per token can still be expensive per job, a point we made about Claude Opus 5’s cost per task and about the previous price cut in our Opus 4.8 pricing piece.
Second-order effect worth watching. When vendors keep cutting list prices, the next competition isn’t over the rate card. It’s over defaults, limits and what’s bundled. A model maker that cuts the list price while letting the typical task burn more tokens has raised prices in practice, and a model maker that does the opposite has cut them. The rate card can’t tell you which one you’re dealing with. Only the invoice can. Buyers who only compare per-million-token prices will miss where the margin moved.
Who should be allowed to pick max?
Almost nobody, by default. We’d set medium for the whole team, allow high where a wrong answer costs real money, and allow max only with a named reason. Then track cost per accepted output, not cost per request, because a rejected draft still costs you the tokens.
We cover that habit in our piece on cost per successful task and in our guide to AI costs for Canadian businesses. The same logic applied when token prices fell and a budget still blew out.
The obvious objection is that nobody in a 40-person firm will tune an effort dial, so the vendor’s default wins. Agreed, and it’s the argument for looking once. If your tooling defaults to medium, the dial matters only when someone pushes it up, and you should find out who. If it defaults to high or max, you’re paying the premium without noticing. One check beats a quarterly argument about the bill.
Is cheaper per task the same as better value?
No, and our own case has a soft spot. It leans on cost-per-task figures from Digital Applied’s report of Artificial Analysis data published on September 22, and Artificial Analysis may revise them. They also show Opus 5.5 running with a fallback enabled, which means some requests that tripped a safeguard were answered by another Claude model.
Artificial Analysis’s own test setup scored Opus 5.5 below Anthropic’s figure on Terminal-Bench 4.0, so vendor benchmarks and independent ones differ. The earlier release made the same promise, and the Claude Opus 5 cost per task told a different story once measured.
A bigger caution cuts against cheap defaults. The same report says Opus 5.5 gave a wrong answer 68.4% of the time when it didn’t know, where two other models more often declined. Cost per task isn’t cost per correct task. If your documents are the kind where max effort catches errors medium misses, our advice flips for that work. The per-task costs come from Artificial Analysis’s test set, not yours, so the monthly totals above are an illustration. Only your own sample will settle it.
Run 50 of your real tasks at medium and at high this week, count the ones your team accepts, and divide the spend by that count.
The mid-tier model followed: what Claude Sonnet 5.5 pricing means in Canadian dollars. Meta now matches this rate card; see how Muse Wonder pricing compares.
Frequently asked questions
How much does Claude Opus 5.5 cost?
Anthropic lists it at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Fast mode costs $8 and $40, and batch jobs cost $2 and $10.
Is Opus 5.5 really 40% cheaper than Opus 5?
That is Anthropic’s claim for typical workloads, based on the model using fewer tokens per task. The list price cut is 20%. Check the 40% against your own invoice.
What effort setting should a small business use?
Medium is a sensible default, since Artificial Analysis data shows it at $1.34 per task against $5.98 at max. Use higher settings only where a wrong answer is costly.
Archive entry dated September 23, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.
