Claude Opus 5 Puts a Dial on Your AI Bill. Use It Right

Anthropic released Claude Opus 5 on July 24, and the headline number is simple. It costs $5 per million input tokens and $25 per million output, the exact same price as the Opus model it replaces, while scoring at or above Claude Fable 5, Anthropic's top model, on several hard benchmarks at half the cost.

But if you run AI anywhere in your business, the pricing is not the most important part. The most important part is a new setting called Effort. It is a dial with four positions, low, medium, high, and max, that controls how much reasoning the model does before it answers. More effort means better answers and a bigger bill. Less effort means faster, cheaper output.

This is the first time a frontier lab has handed operators direct per-task control over the cost and quality tradeoff. This article covers what Opus 5 actually is, what the Effort dial does to your bill, and a simple rule for setting it.

What did Anthropic actually ship on July 24?

Claude Opus 5 is a frontier-class model priced at $5 per million input tokens and $25 per million output that gets close to Fable 5 performance at half Fable's cost.

Start with the benchmark that matters most for business automation. On Frontier-Bench v0.1, a benchmark built around agentic coding tasks, Opus 5 scored 43.3%. The model it replaced, Opus 4.8, scored 18.7%. Fable 5, the most capable model Anthropic sells, scored 33.7%.

Read that again. The mid-tier model now beats the flagship on an agent benchmark, at half the price. Independent testing from CodeRabbit and Vellum points the same direction, with Opus 5 showing particular strength in debugging and root-cause analysis.

The second spec worth knowing is the context window. Opus 5 handles 1 million tokens in a single prompt, which is roughly 750,000 words. In practical terms, that is your entire SOP library, a full quarter of sales call transcripts, or a complete contract stack analyzed in one pass instead of chunked across twenty prompts.

You can use it today through the Claude apps or the API under the model name claude-opus-5. If you pay $20 to $25 a month for a Claude seat, you already have access.

What is the Effort dial and why should you care?

Effort is a per-request setting with four levels that controls how much reasoning Opus 5 does, and the gap between levels can roughly double your token spend on the same task.

Here is how it works. At low effort, the model answers quickly with minimal reasoning. At max effort, it works the problem hard before responding. Benchmark data from FrontierCode shows a high-effort query burns roughly twice the tokens of a medium-effort query on the same task.

The part most coverage missed: medium is the sweet spot, not max. The same FrontierCode data shows Opus 5 at medium effort delivers near-top benchmark scores at about half the compute cost of the higher settings. Past medium, you pay double for single-digit gains on most tasks.

Think about how you staff your own business. You do not put your most senior person on every task. You match the person to the stakes. The Effort dial is the same decision, applied to a model, on every single request.

Until now, your only cost lever was switching models entirely, which usually meant rebuilding prompts and re-testing everything. Now the lever is a parameter you change in one line.

Does Opus 5 actually lower your AI bill?

Only if you pay for tokens directly. If your AI spend is seats and bundled software, nothing on your bill changes this week.

Your AI spend lives in three places. Seats, like the $20 to $25 per person you pay for Claude or ChatGPT. Bundled tools, like Make.com starting near $10 a month, n8n cloud around $24, or GoHighLevel at $97 to $297, where the AI cost is baked into the subscription. And direct API tokens, where you pay per unit of work.

Opus 5 and its Effort dial only touch the third layer. Your Claude seat does not get cheaper because the API has a new dial.

But if you run API workflows, the math gets interesting. Say you have an automation that processes 40 sales call transcripts a day, about 5,000 tokens in and 800 tokens out per run. On Opus 5 at medium effort, that costs about $1.80 a day, or roughly $54 a month. Run the same job at high effort and the reasoning tokens push it to about $78 a month, a 44% increase for output that is barely different on a routine summarization task.

And here is the check that keeps you honest: that same transcript job on GPT-5.6 Luna, OpenAI's fast tier at $1 input and $6 output per million tokens, costs about $12 a month. Opus 5 is not the tool for cheap bulk work. It is the tool for work where reasoning quality actually changes the outcome, and the Effort dial is how you avoid paying reasoning prices for non-reasoning tasks.

How do you decide which effort level to use?

Use a rule I call Effort Matching: set the effort level by what a wrong answer costs you, not by what a right answer is worth.

Most people will set this dial by asking how important the task is. That is the wrong question, because everything feels important. The better question is what happens if the output is wrong, and how fast you would catch it.

Effort Matching has three tiers.

Low effort goes to work that is internal, reversible, and high-volume. Tagging leads, sorting inbound email, summarizing meeting notes, first drafts nobody outside your team will see. A wrong answer here costs you a minute of cleanup.

Medium effort is your default for production work. Client-facing drafts, data analysis, code your developer reviews, proposal sections a human edits before sending. A wrong answer gets caught by the review step you should already have.

High or max effort is reserved for one-shot work where errors are expensive and review is hard. Long-running agent tasks, contract analysis, database migration scripts, anything that touches money or data you cannot easily restore. A wrong answer here costs you real dollars or a client.

One caveat that matters more than the dial: a medium-effort answer reviewed by a person beats a max-effort answer shipped blind. Effort buys reasoning. It does not buy accountability. Keep humans on anything that leaves the building.

Where does the Effort dial fit in your existing stack?

If you build automations in Make.com, n8n, or GoHighLevel, the Effort dial shows up as one more field in your existing AI module, and it is worth setting deliberately instead of leaving the default.

Here is what that looks like in practice. In n8n or Make.com, your Claude API step has a model field where you select claude-opus-5, and the effort setting rides along as a request parameter. That means one workflow can use different effort levels at different steps.

A post-call automation is a good example. Step one transcribes and summarizes the call. That is low effort work, high volume, internal, easy to spot-check. Step two drafts the follow-up email the client will actually read. That is medium. Step three, if the call was a closing call, extracts contract terms and flags anything unusual against your standard agreement. That is the one step worth high effort, because a missed term costs real money and nobody re-reads the raw transcript later.

Same workflow, three effort levels, and the expensive setting only fires on the 5 calls a week that are closing calls instead of all 40. That is the difference between a $54 monthly bill and a $78 one at scale, and it compounds across every workflow you run.

If your automations live inside GoHighLevel's native AI features, you do not get this control yet. Platform tools decide model and effort for you, and that convenience is part of what the $97 to $297 subscription pays for. That tradeoff is fine at low volume. It stops being fine when your automation volume grows to the point where you are effectively buying tokens at retail through a subscription markup.

Should you switch to Opus 5 this week?

Switch if you are currently paying flagship prices for work Opus 5 can handle. Do not rip out a working workflow to chase a benchmark.

The clearest case for switching is anyone running Fable 5 or another flagship-tier model on agent or coding workloads. If Opus 5 scores higher on agentic benchmarks at half the cost, and your workload looks like those benchmarks, you are overpaying today.

The clearest case against switching is a workflow that already works. Every model swap carries hidden costs: re-testing prompts, re-checking edge cases, retraining your team on new failure modes. If a workflow makes you money reliably, the burden of proof is on the new model, not the old one.

The middle path takes one afternoon. Pick your single most expensive AI workflow. Pull the last 10 real inputs it handled. Run them through Opus 5 at medium effort and compare outputs side by side against what you shipped. If Opus 5 matches or beats your current model, switch that one workflow and watch it for two weeks before touching anything else.

There is also a bigger pattern here worth naming. Last week, GPT-5.6, Grok 4.5, and Gemini 3.6 Flash shipped within days of each other and competed mostly on price. Now Anthropic ships a model at the same sticker price with a dial that lets you cut your own bill. Cost control is becoming a feature the labs compete on. Operators who know their per-task AI cost will be the ones who benefit. Operators who only know their monthly total will keep wondering where the savings went.

What should you actually do this week?

Spend 30 minutes finding out which layer of the AI bill you are actually paying, because everything else depends on that answer.

First, pull your last credit card statement and split your AI spend into seats, bundled tools, and direct API. Most service businesses find 80% or more sitting in the first two layers, which means model news like this does not touch their bill yet.

Second, if you have API spend, run the one-afternoon test above on your most expensive workflow, at medium effort. Medium, not max. The data says that is where the value is.

Third, if you are seats-only today, note the direction of travel. The gap between operators who run AI through seats and operators who run it through workflows is now a cost-control gap, not just a capability gap. That gap compounds every time a release like this ships.

The model is good. The dial is better. Use both deliberately.

---

If you want a clear picture of what AI can actually do for your specific operation, book a free AI Clarity Call. Thirty minutes, no pitch, you leave with a real answer.

If you want to learn alongside other operators and stay current on what is working, join the Abra AI community. That is where I share what I am actually building.

Subscribe to the newsletter for more breakdowns like this.