Same price per token, smaller bill: what a job costs

Same fare, different bill

Watch this part · 1:23Claude Sonnet 5.5: same token price, so why is it cheaper?

Picture 2 taxis. They are the same car, with the same fare on the door, making the same trip across town. One meter stops at 12 cents. The other stops at 15.

The first taxi is Claude Sonnet 5.5, which Anthropic released on 28 September 2026. The second is Sonnet 5, the model it replaces. Per tokenTokenThe small piece of text a language model reads and writes, often part of a word. Prices, limits and the context window are all counted in tokens. they cost exactly the same. Yet the new one is sold as up to 30% cheaper per task, on its maker's own tests.

So we put it on the meter ourselves: 2 rounds, 3 models, the same jobs. One round matched the promise. The other did not. The reason applies to any model you will ever pay for.

The price puzzle

Start with the price, because this is where most people go wrong.

You may have read that Sonnet 5 went up to $3 and $15 on 1 September. It did not. That rise was cancelled and the launch price stayed.

So the saving is not a price cut. At the same price there is only one way a job can cost less: fewer tokens to finish it.

Back to the taxi. The fare per kilometre is the price per token. The distance is how many tokens the job takes. With the same fare and a shorter drive, the bill is smaller. That is the whole promise.

What the meter counts

The meter counts tokens, the small pieces of text a model reads and writes. Your question goes in as tokens. Every word of the answer comes out as tokens. Every big model works this way, whether it is Claude, ChatGPT or Gemini. The bill is tokens times price.

There is one line on the receipt most people never see. When a model stops to think before it answers, those thinking tokens are billed too, as output, at the higher price.

So a model that thinks more can cost more, even when its answer is exactly the same length. This is the twist in round 1.

How we tested

We ran 3 models: Sonnet 5.5, Sonnet 5, and Opus 5.5 at twice the fare.

• Every model got the same prompt, the same short instructions and no extra files.
• Every model ran at the same effort setting, Medium, which is the default in the Claude apps.
• Every job ran twice and we took the middle value.
• Costs are worked out at the public list prices, so you can check the sums.

We ran 2 rounds because they answer 2 different questions. Round 1 is quick questions, the kind you type into a chat app every day. Round 2 is small jobs with several steps, where the model uses tools and checks its own work.

It is a small test. Read it as a direction, not a verdict.

Round 1: quick questions

We asked 6 everyday questions: plan a 5-day trip in New Zealand, fix a small bug, write a budget, an email and a slide outline from a page of notes, and count the r's in strawberry.

Speed. Sonnet 5.5 wrote about 105 tokens a second against 73 for Sonnet 5. That is 45% faster. The speed claim holds.

Cost. Across all 6 questions Sonnet 5.5 used 43% more tokens and cost 28% more.

Faster but more expensive? Look at what happened before the first word. On the trip and the budget, Sonnet 5.5 chose to think first, with about 900 tokens of thinking on the trip alone. Sonnet 5 barely thought at all. The answers you actually read were only about 4% longer. The extra cost was almost all thinking, a detour you never see that still runs on the meter.

Was the thinking worth anything? We read every answer.

• The easy ones everybody got right: 3 r's in strawberry, every email, every slide outline.
• On the bug, all 3 models found it. But Sonnet 5 slipped twice. Once it said the broken code returned 3.5 when it returns 2.33, and once it said a tricky case was handled when it was not. Sonnet 5.5 and Opus 5.5 got it right both times.
• On the trip, Sonnet 5.5 and Opus 5.5 planned both islands. Sonnet 5 stayed on one island each time.

On a sample this small that is a hint and not proof. On a short hop across town, the new model spends a little more to think, and its answers came out more careful.

Round 2: jobs with steps

Round 2 is where the promise should show. We set 3 small jobs, each needing several steps: fix 5 bugs and keep going until every test passes, clean up a messy spreadsheet of expenses and total it, and turn meeting notes into an email and a 5-slide outline.

Each step is a round trip. The model thinks, uses a tool, looks at the result and decides what to do next.

That is 24% faster and 20% cheaper.

Why fewer steps? Not because it did less work. Across every run, both models used their tools the same number of times. Sonnet 5.5 bundles several of those calls into one step and does not drive back and forth.

The biggest gap was the spreadsheet. Sonnet 5 stopped to thinkChain of ThoughtHaving a model write out the steps of its reasoning before it gives an answer. Once a prompting trick, now built into "thinking" models. for about 340 tokens on each run. Sonnet 5.5 took about 65. This time the older model took the detour.

And the work itself? Every model finished every job in both runs and passed every check. The tests were green, the totals were right, and not one number was made up.

The promise was up to 30%. We measured 20% on 3 small jobs. It is not the headline number, but it is a real saving that points the same way.

The luggage nobody mentions

Watch this part · 1:14Hidden tokens that made Sonnet 5.5 look cheaper

There is a catch in testing AI models that almost nobody talks about, and it nearly fooled us.

We ran the test inside Claude Code, Anthropic's coding tool. Claude Code sends every model a set of instructionsSystem PromptInstructions the builder of an AI app gives the model before you type anything. It sets the role, the tone and the rules, and it outranks what users say. for its tools. For Sonnet 5 that set is about 4,000 tokens heavier than for Sonnet 5.5, on every single step. With the default setup it is about 7,000.

That has nothing to do with the model. It is luggage the tool puts in the boot, and you pay to carry it on every step.

So before comparing, we gave every model the same short instructions and took the extra weight out.

• With the luggage left in, Sonnet 5.5 looks 25% cheaper.
• With it taken out, 20%.

5 points of that saving were the luggage and not the model. If you work in Claude Code you really do get the bigger number, because the luggage is on your bill too. It just is not the model's doing.

So if you ever compare models yourself, weigh the bags first. Check what your tool sends, not only what you type.

Do you need the bigger model?

Opus 5.5 costs twice as much per token as either Sonnet.

On our 3 jobs it passed exactly the same checks for 27.6 cents. That is 83% more than Sonnet 5 and more than double Sonnet 5.5. On the quick questions its trip plan was the most complete of all, but the Sonnets got the same things right for a fraction of the cost.

Anthropic says Opus still pulls ahead on long, open-ended work that needs sustained judgement. That is exactly the kind of job our small test did not cover.

Our rule of thumb: start with the smaller model. Move up when you can see it struggling, not before.

Why bigger jobs save more

If the saving is real, why did we see only 20%? Because the size of the job decides it.

• A quick question: there is hardly any distance to save. In our test it cost more.
• Small jobs with a few steps: a fifth less.
• Big jobs: the gap grows fast. On a test of agentsAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done. working at the command line, Anthropic reports that Sonnet 5.5 at Medium effort beats Sonnet 5's best score for less than a tenth of the cost. On 2,441 finance tasks, the firm Balyasny saw about 121,000 tokens per answer against 497,000 for Sonnet 5.

The longer the drive, the more a shorter route saves. That is our reading and not a measurement, and it is why your own saving depends on your own work.

What to do with this

If you use a chat app

• For everyday work and anything with several steps, the newer model is the better choice.
• The effort setting controls how long the model thinks before answering. For quick questions turn it down to Low. In our test that used 21% fewer tokens with the same answers.

If you run a team or a product on the API

• Measure cost per task, not price per token.
• Run your own jobs through both models, twice, and weigh the bags.
• On Sonnet 5.5 the API starts on High effort, not Medium, and thinking can no longer be switched off.
• Coming from Sonnet 4.6 or older? Count your tokens again. The same text makes about 30% more.

The same sum works for any model from any maker: tokens in times the price in, plus tokens out times the price out, with thinking counted as output. The price calculator does that sum for the current models.