Anthropic says Claude Sonnet 5.5 “costs up to 30% less for most work.” The Claude Sonnet 5.5 pricing card did not move. Input is still US$2 per million tokens and output US$10, so the saving shows up on a Canadian invoice only if your own tasks use fewer tokens. That is worth testing before anyone books it into next year’s budget.
- Sonnet 5.5 is priced the same as Sonnet 5: US$2 per million input tokens and US$10 per million output tokens.
- The one guaranteed price cut is cache reads, down to US$0.10 per million tokens from US$0.20.
- The “up to 30% less” figure is a per-task claim from Anthropic’s own testing, not a discount.
On September 28, 2026, Anthropic released Claude Sonnet 5.5 and led with two numbers: 30%+ faster and up to 30% less for most work. Read one paragraph further and the price per token is unchanged. The model is probably better. For a Canadian team paying in US dollars, what matters is whether the bill gets smaller, and by how much.
Key points
- What did Anthropic actually cut? One line item. Anthropic’s announcement says cache reads now cost US$0.10 per million tokens, half of Sonnet 5’s US$0.20.
- Where does the per-task saving come from? From the model finishing with fewer tokens and fewer tool calls.
- Who captures the value? If the efficiency is real, the customer does, at least at first.
What did Anthropic actually cut?
One line item. Anthropic’s announcement says cache reads now cost US$0.10 per million tokens, half of Sonnet 5’s US$0.20, and everything else on the rate card is unchanged. The rest of the promised saving rests on Anthropic’s claim that the model needs fewer tokens per task and costs up to 30% less in its own testing.
One line item. Anthropic’s announcement says cache reads now cost US$0.10 per million tokens, half of Sonnet 5’s US$0.20. Everything else on the rate card is the same as Sonnet 5. The rest of the saving comes from Anthropic’s statement that the model “typically needs far fewer tokens to do the same work” and that “in our testing, it costs up to 30% less per task.”
That is a claim about efficiency, measured by the vendor on the vendor’s tasks. It may hold for you. It may not. “Up to” sets a ceiling, and a ceiling is not an average (the marketing department knows this, which is why it chose the phrase).
Where does the per-task saving come from?
From the model finishing with fewer tokens and fewer tool calls. Anthropic’s benchmark charts plot score against cost per task at each effort setting, and they make a stronger point than the headline: on several tests, Sonnet 5.5 at low or medium effort beats Sonnet 5’s best score “for about a tenth of the cost per task.”
The same page also says that at higher effort settings it “can perform comparably” to Opus 5.5 “at a similar cost.” So the effort setting your developers pick moves the bill more than the model name does. A team that switches to Sonnet 5.5 and turns effort up to Max may pay roughly what it would have paid for Opus.
Show Me the Invoice
Here is one workload priced three ways. The workload is an illustration, not a measured customer. The rates are Anthropic’s. The exchange rate is the Bank of Canada’s daily rate for September 29, 2026: 1.4188 Canadian dollars per US dollar.
| Input | Value | Type |
|---|---|---|
| Monthly input tokens | 20 million | Illustration |
| Monthly output tokens | 4 million | Illustration |
| Monthly cache reads | 50 million | Illustration |
| Input and output rate, both models | US$2 and US$10 per million | Fact (Anthropic) |
| Cache read rate | US$0.20 (Sonnet 5), US$0.10 (Sonnet 5.5) | Fact (Anthropic) |
| Token reduction on your tasks | 0% or 30% | Vendor claim, tested both ways |
The sum, per month:
- Sonnet 5: (20 x $2) + (4 x $10) + (50 x $0.20) = US$90, or C$127.69.
- Sonnet 5.5, same token use: $80 + (50 x $0.10) = US$85, or C$120.60.
- Sonnet 5.5, 30% fewer input and output tokens: $56 + $5 = US$61, or C$86.55.
The guaranteed saving on this workload is US$5 a month, about 6%. The full 30% claim is worth US$29 a month. The gap between those two lines is the part you have to prove yourself.
Sensitivity: at ten times the volume, the same three lines read C$1,276.92, C$1,205.98 and C$865.47 a month. The percentage gap does not change with scale. It depends on token efficiency, which is why the test below matters more than the volume.
Who captures the value?
If the efficiency is real, the customer does, at least at first. Anthropic gets a faster model that is cheaper to serve per task and more likely to win work from rivals. As our Claude Opus 5.5 pricing analysis argued, the list price is only one input to the cost of a finished piece of work.
If the efficiency is real, the customer does, at least at first. Anthropic gets a faster model that is cheaper to serve per task and more likely to win workloads that would otherwise go to rivals. Our Claude Opus 5.5 pricing analysis made a similar point about the top tier: the list price is only one input to the cost of a finished piece of work.
The bull case
The strongest case for taking the claim at face value is in Anthropic’s own data. Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 against Sonnet 5’s 10.3%. A model that finishes a task on the first attempt uses fewer tokens and also skips the second and third attempts. For agent work, where retries pile up, the real saving could be larger than 30%, not smaller.
We think that is plausible for coding agents. We are less sure it carries over to short, single-turn work such as email drafts or summaries, where there are few retries to eliminate.
What would change our view
Independent cost-per-task measurements on business workloads, rather than coding benchmarks, would move us. So would a Canadian customer publishing before-and-after invoices. Until then, treat 30% as the best case and 6% (on a cache-heavy workload like ours) as the floor.
Run this on your own numbers
Take last month’s Anthropic usage export. Rerun 50 real tasks on Sonnet 5.5 at the same effort setting you use now, and record total tokens per completed task. Divide the new figure by the old. If the ratio is near 0.7, the claim holds for you. If it is near 1.0, the only saving you will see is on cache reads, and the switch is a quality decision, not a cost one. For more model cost breakdowns, see AI Tools and Reviews.
Anthropic has since told investors its pricing may change; see our read of the Anthropic S-1 risk factors.
FAQ
How much does Claude Sonnet 5.5 cost?
Anthropic lists Claude Sonnet 5.5 at US$2 per million input tokens and US$10 per million output tokens, the same as Sonnet 5. Cache reads cost US$0.10 per million tokens.
Is Claude Sonnet 5.5 cheaper than Sonnet 5?
Only cache reads are cheaper by rate. Anthropic says the model uses fewer tokens per task and costs up to 30% less in its testing, but that depends on your workload.
Does the effort setting change Claude Sonnet 5.5 pricing?
The rate per token stays the same, but higher effort uses more tokens per task. Anthropic says that at high effort, Sonnet 5.5 can cost about the same per task as Opus 5.5.
