OpenAI vs Claude: I've Shipped 126 Systems on Both. Here Is How I Actually Choose.

A client called me in June with a spreadsheet. They had been running their support agent on OpenAI for eight months, and their bill had climbed from $180 a month to just under $900. Nothing about the product had changed. Volume had roughly doubled. The rest was output tokens, and nobody had ever looked at the output rate.

That conversation is why I keep coming back to the openai vs claude question with business owners rather than engineers. The two platforms are close enough on capability now that the deciding factors are boring ones: what you pay per million output tokens, whether a promotional price is going to expire on you, and which failure mode you can live with. I have shipped 126 production systems, and I have built on both.

Here is the short version. As of August 2026, Claude Opus 5 and GPT-5.6 Sol cost exactly the same on input, $5 per million tokens. On output, Opus 5 is $25 and Sol is $30. For a typical support agent that is a difference of about $12.50 a month, which is nothing. The real decision is which tier you drop down to, and that is where the two vendors diverge sharply.

If you would rather skip the analysis and have someone size this against your actual call volume, book a discovery call and bring your last three months of invoices.

What OpenAI and Claude actually charge in August 2026

Every price below is the published list rate, checked on 3 August 2026 against Anthropic's own model documentation and OpenRouter's live model pages. I am listing the current lineup, not last year's. Most comparison articles still ranking for this query are quoting GPT-5.4 and Claude Opus 4.6, both of which have been superseded.

Two things jump out. Claude is cheaper on output at the top tier. And OpenAI's bottom tier, Luna, is dramatically cheaper than anything Anthropic offers, at $0.20 in and $1.20 out.

Claude Sonnet 5 also carries an introductory rate of $2 and $10 per million tokens, but that expires on 31 August 2026. If you are budgeting past September, use the $3 and $15 list rate. I have watched three separate teams build a business case on a promotional number and then get surprised. Do not be the fourth.

What that means in real monthly cost

Rate cards are useless until you put a workload against them. So here is one, and I am showing my assumptions so you can run it again with your own numbers.

Assume a support agent handling 1,000 conversations a month. Each conversation burns roughly 40,000 input tokens once you count the system prompt, retrieved documents and the growing message history, plus about 2,500 output tokens. That works out to 40 million input tokens and 2.5 million output tokens a month. These are the numbers I see on a mid sized deployment.

The spread between the cheapest and most expensive option is 25x. That is the actual finding. The openai vs claude decision is worth about $12.50 a month at the top tier. The tier decision is worth $264.

This is what my June client had gotten wrong. They were not on the wrong vendor. They were running every single request, including "what are your hours", through a flagship model. We moved intent classification and simple lookups to a cheap tier and kept the expensive model for genuine reasoning. Their bill went to $310 and their customers noticed nothing.

Where OpenAI and Claude genuinely differ

Output pricing is the gap that compounds

Claude charges $25 per million output tokens at the top tier against OpenAI's $30. On input the two are identical. That sounds trivial, and on a chat product it is. On an agent it is not, because agents generate far more output per request than a chatbot does. Every tool call, every reasoning step and every retry is billed as output.

In the voice deployments I run, output is between 30 and 40 percent of total spend even though it is a fraction of the token count. If your workload is agentic rather than conversational, weight the output column heavily. I wrote about how that shakes out in practice in taking AI agents to production.

Context windows are effectively a tie

Claude Opus 5 and Sonnet 5 both offer a 1M token context window. The GPT-5.6 family offers 1.05M. Nobody should pick a vendor over a 5 percent difference in a window that almost no business application fills. Claude Haiku 4.5 is the outlier at 200K, which is still more than enough for most support and booking work but will bite you on long document analysis.

What matters far more is what you put in the window. Most teams stuff it and pay for it. Retrieval done properly keeps the window small, and I covered the mechanics of that in the business guide to RAG.

The discount trap

This one costs people real money. Right now OpenAI is running a 50 percent promotion on the Terra and Luna tiers. On OpenRouter, Terra shows as $1 and $6 per million tokens with the list price of $2 and $12 struck through.

Two lessons sit in that screenshot. First, the promotional rate is explicitly described as limited, so budget on the $2 and $12 list price and treat the discount as a bonus. Second, look at the provider column. The identical model runs $2.20 through Amazon Bedrock and $2.50 through Azure against $2.00 list from OpenAI directly. If your agency or platform resells you model access, you are very likely paying that markup without seeing it itemised.

I ask every new client one question early: who is billing you for tokens, and can you see the raw usage? If the answer is no, that is usually the cheapest thing to fix.

What I actually choose, and when

Across the builds I have shipped, my selection rarely comes down to brand preference. It comes down to three questions.

Is the task reasoning or retrieval? If the agent needs to weigh options, chain several tools together or handle genuinely novel requests, I use a top tier model and I lean Claude, because the output rate is 17 percent lower for work that is output heavy. If the task is classification, extraction or looking something up, I use the cheapest tier that passes evaluation, and OpenAI's Luna tier at $0.20 and $1.20 is currently unbeatable on price.

How much does a wrong answer cost? A booking agent for a dental practice that books the wrong appointment slot costs a little time and some goodwill. A quoting agent that misprices a job costs margin on every quote until someone notices. On the second kind I spend more per token without hesitating and I add a verification pass. On the first I optimise for cost.

Are you locked in? Both vendors ship breaking changes. Anthropic deprecates on a published schedule with a migration guide. OpenAI moves faster and communicates less. Neither is disqualifying, but if you build so that swapping the model means changing one configuration value rather than rewriting your prompts, this whole debate stops being expensive. Every system I ship is built that way, and it is the single highest return decision in the stack.

For most of the small and mid sized businesses I work with, the answer ends up mixed. A cheap tier handles 70 percent of traffic, a top tier handles the rest, and which vendor supplies each one changes about once a year. That is a normal, healthy setup, and it is roughly what the operations autopilot and knowledge agent packages are built around.

Is this decision even right for you yet?

Be honest about where you are, because picking a model is the wrong problem for a lot of businesses.

• You have not shipped anything yet. Pick either one and build. At prototype volume the entire monthly bill is under $20 on any tier. Deciding this now is procrastination wearing a spreadsheet.
• Your bill is under $200 a month. Your time is worth more than the difference. Revisit when you cross $500.
• Your bill is over $500 a month and rising. This is the right conversation, and tier routing will save you more than vendor choice will.
• You are quoted a flat monthly fee with no token visibility. Ask for the underlying usage before anything else. You may be paying a markup of 20 to 25 percent on top of a flagship model you did not need.
• You are not sure whether you need an agent at all. Quite often the answer is that you need a workflow, not an agent. I laid out the difference in when to use AI agents versus automation, and the AI readiness assessment will tell you in about five minutes.

Frequently asked questions

Citation Capsule: As of 3 August 2026, Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens, against GPT-5.6 Sol at $5 and $30. Claude Sonnet 5 lists at $3 and $15 with an introductory rate of $2 and $10 through 31 August 2026. GPT-5.6 Terra lists at $2 and $12 and Luna at $0.20 and $1.20, both currently discounted 50 percent on a limited promotion. Claude models offer a 1M token context window against 1.05M for the GPT-5.6 family. Sources: Anthropic Models Overview 2026, OpenRouter GPT-5.6 Terra 2026, OpenRouter GPT-5.5 2026, OpenAI Platform 2026.

The thing worth doing this week

Pull your last three invoices and find the split between input and output spend. If output is more than a third of the bill, you are running an agentic workload and the output rate matters, which nudges you toward Claude. If your traffic is mostly simple and repetitive, the cheapest OpenAI tier will save you more than any vendor argument.

Then check what percentage of requests genuinely need a flagship model. In every audit I have run the honest answer was under 30 percent, and in one case it was 6 percent.

If you want that done properly against your own numbers rather than my modelled ones, book a discovery call and I will go through your usage with you. If you are earlier than that and still working out what to automate first, start with the AI readiness assessment or look at how the agent packages are scoped.