If your business pays for AI by the token, both of April's flagship upgrades can make your bill go up, just in different ways. OpenAI's GPT-5.5 API list prices are double GPT-5.4's rates, though OpenAI reports it uses roughly 40% fewer output tokens than GPT-5.4. Anthropic's Claude Opus 4.7 keeps its old price but can count the same text as more tokens. The safe move for a small business is the same either way: run a sample of your real work through the new model, compare the bill and the quality, and only then switch.

That's the short answer. Below is what changed, what it means for your costs, how to test it without a developer on staff, and which settings keep the bill under control.

$5 / $25
Opus 4.7 per million input / output tokens (unchanged)
1.0–1.35×
Tokens for the same input on Opus 4.7, per Anthropic
$5 / $30
GPT-5.5 per million input / output tokens
~+20%
GPT-5.5 xhigh: more expensive to operate on Artificial Analysis's index

Two April upgrades, two different ways your bill goes up

You shouldn't have to watch AI model launches to run your business. But if you have a chatbot on your website, an assistant that drafts client letters, a tool that summarizes intake forms, or a vendor product built on one of these models, the vendor's upgrade quietly becomes your cost change.

Anthropic: same price tag, more tokens

Anthropic released Claude Opus 4.7 across all Claude products and its API, and also on Amazon Bedrock, Google Cloud's Vertex AI and Microsoft Foundry. In its Opus 4.7 announcement, the company says pricing stays the same as Opus 4.6: $5 per million input tokens and $25 per million output tokens.

The catch is in the same announcement. Check how your provider's current tokenizer counts your own text, since the same input can produce a different number of billable tokens than it did on an older model. Anthropic says the same input can map to roughly 1.0–1.35× as many tokens, depending on the content type. Anthropic also says Opus 4.7 thinks more at higher effort levels, particularly on later turns in agentic settings, and that this means it produces more output tokens. Output tokens are the expensive ones.

Anthropic's own advice is to measure the difference on real traffic. We think that's the most useful advice in it, and we'll come back to it.

OpenAI: a higher price tag, fewer tokens (by its own account)

OpenAI announced GPT-5.5 on April 23, 2026. It is priced at $5 input and $30 output per million tokens on the API. Artificial Analysis notes that per-token pricing has doubled from GPT-5.4.

The pitch is that the new model gets the job done with fewer tokens. Implicator.ai's breakdown cites roughly 40% fewer output tokens than GPT-5.4. Artificial Analysis found GPT-5.5 (xhigh) used about 40% fewer output tokens to run its index than its predecessor, and says that reduction largely absorbs the hike, resulting in a net cost about 20% higher to run its Intelligence Index.

So the two companies took opposite routes to the same place. One raised the price and says you'll use fewer tokens. The other kept the price and counts more tokens. Neither upgrade is automatically cheaper for your work.

Claude Opus 4.7GPT-5.5
List price per million tokens (input / output)$5 / $25$5 / $30
Change vs. previous version's priceUnchanged from Opus 4.6Double GPT-5.4
Token count for the same workSame input can be ~1.0–1.35× more tokensAbout 40% fewer output tokens in Artificial Analysis testing
What independent testers saw12–27% higher costs, except short prompts, which got more cost efficient, per OpenRouterAbout 20% more expensive to run on Artificial Analysis's index

Price per token vs. tokens per task, explained without math

Think of tokens like billable minutes. Roughly speaking, you pay the rate times the minutes used.

In its example, a request that cost $0.10 on Opus 4.6 could cost anywhere from $0.10 to $0.135 on Opus 4.7, depending on content mix.

Finout also worked a monthly illustration. A coding-agent workload of 1 million input tokens and 0.2 million output tokens per day costs about $10 a day, or about $300 a month, on Opus 4.6. The same workload on Opus 4.7 at 35% token inflation comes to about $13.50 a day, or about $405 a month. That's the top of Anthropic's range, so treat it as a ceiling for that example, not a forecast for yours.

How much the tokenizer change hits you depends on what you send. OpenRouter's analysis found that for production-scale prompts of 10K+ tokens, Opus 4.7 produced 32–34% more native tokens than 4.6 for equivalent text, while short prompts actually got more cost efficient. Run your own documents through both model versions and compare the token counts, since the difference varies a lot by content type.

Translation for an owner: a quick one-line customer reply may barely change. A workflow that feeds the model long contracts, patient histories, project specs or a big policy manual on every request is where the extra tokens pile up.

Hypothetical: what "double the price, 40% fewer tokens" does to one workflow

Hypothetical: an accounting office runs a document-summary tool on GPT-5.4. Input is long (client statements); output is short (a one-page summary). On GPT-5.5, the input rate doubles and the input is the same length, so that part of the bill doubles. The output rate also doubles, but if output tokens really drop 40%, the output part ends up about 1.2 times what it was. Because most of this office's spend was input, the total lands much closer to double than to flat.

Now flip it. Hypothetical: a marketing agency's tool takes a short brief and writes long drafts. Most of the spend is output, so the 40% reduction matters more and the increase is smaller. Same model, same price change, very different result. This is why a general headline number can't tell you what your bill will do.

A one-week replay test before switching models

Finout's practical advice is to replay a representative traffic sample through 4.7 and measure token counts and quality side by side. Anthropic recommends measuring on real traffic too. We think a week of your actual work is the right sample for most small businesses, because it catches the Monday intake rush and the Friday reporting run, not just a handful of tidy test prompts.

Here's how we'd set it up, in plain terms:

  1. Start where the spend is. Check your AI provider's usage dashboard or your vendor's invoice to see where the money goes, tool by tool or key by key. Start there.
  2. Save a week of real requests. Keep the actual inputs (with client names or patient details handled under your normal privacy rules) and the outputs the current model produced.
  3. Run the same requests through the new model. Same prompts, same settings. If you use the higher effort settings, test those too, since that's where extra thinking shows up.
  4. Compare three numbers per workflow: tokens in, tokens out, and cost per completed task. Cost per task is the one that matters.
  5. Have a person grade quality. Pick a sample of outputs and have the person who normally checks the work score old vs. new without knowing which is which.
  6. Decide workflow by workflow. It's common for the new model to be worth it on hard tasks and not worth it on routine ones. You don't have to move everything at once.
Tip

If you buy AI through a software vendor rather than paying for tokens directly, ask the vendor which model your plan runs on, whether they plan to switch, and whether your per-seat or per-task price will change. The token math still lands on someone, and it's better to know who before renewal.

If you don't have the time or the in-house skills to run this, it's exactly what a token expenditure audit covers: we pull your usage, replay real traffic on the candidate models, and give you a per-workflow recommendation with the numbers behind it.

Controls that cut cost: effort settings, task budgets, caching, batch

A more expensive model doesn't have to mean a more expensive month. Both vendors give you dials, and they're worth learning.

Effort settings

Anthropic's documentation describes the effort parameter as a way to control how many tokens Claude spends when responding, trading off thoroughness against token efficiency with a single model. Opus 4.7 adds a new xhigh level between high and max. For GPT-5.5, reasoning effort supports none, low, medium, high and xhigh, and now defaults to medium.

Our advice: don't run every task at the top setting. Sorting emails or pulling a date out of a form doesn't need deep reasoning. Save high effort for the jobs where a mistake is expensive.

Task budgets

Anthropic launched task budgets in public beta to give developers a way to guide Claude's token spend. A task budget tells Claude how many tokens it has for a full agentic loop, including thinking, tool calls, tool results and output. Two caveats from the docs: it's a soft hint, not a hard cap, and the minimum accepted budget is 20,000 tokens. So it helps steer multi-step agents, but you still want spending alerts on your account.

Prompt caching

If your tool sends the same long instructions or reference documents on every request, caching can cut that cost sharply. Cache read tokens cost 0.1 times the base input price, while writing to a 5-minute cache costs 1.25 times the base input price. The default cache lifetime is 5 minutes, and it refreshes at no additional cost each time the cached content is used. For a busy front desk or support bot that reuses the same policy text all day, that adds up.

Batch processing

Not everything needs an answer in two seconds. Anthropic's Message Batches API cuts costs by 50 percent, with most batches finishing in under an hour. Overnight report drafting, bulk document tagging and end-of-day summaries are good candidates.

Heads up

Check that your software actually uses caching and batch before counting on those savings. Many off-the-shelf integrations send the full prompt fresh every time. If nobody has checked, assume yours does too.

We'd start by tightening what you send in the first place. Trimming bloated system prompts, sending only the relevant pages of a document, and not resending the whole conversation history every turn is the work we do under context engineering and token optimization.

When to stay on the older or smaller model

Newer isn't automatically better for your workload. If your workload is cost-sensitive, test whether staying on the older model or moving to a smaller one meets your needs before you upgrade. We'd put it this way:

Stay put, or step down to a smaller model, when:

Consider upgrading when:

We'd usually suggest a mix: a strong model for a few hard jobs, a cheaper one for the volume work. We treat routing tasks to the right model as a normal part of AI automation design, not an exotic extra.

What we'd do this week

If we were running a clinic, law office, trades company or agency that pays for AI today, here's the checklist we'd follow:

None of this needs a big project. We work with businesses across Southern California, and the pattern is the same whether you're in Riverside or Santa Monica: the businesses that measure first don't get surprised.

Next step

If you'd like a second set of eyes on your AI spend before you switch models, book a Free Automation Audit. We'll look at what you're running, where the tokens go, and whether the April upgrades are worth it for your work.

Sources