If you run a small business, build products, or write code for a living, you have probably asked some version of this question: which AI provider is actually cheaper for what we do? The answer is less obvious than it looks, and it usually turns on something that never appears on a pricing page: how the text you send gets counted.
Social media is full of claims that OpenAI tokens and Anthropic tokens are not the same thing, and that one consumes noticeably more than the other. That part is true. The implications run further than a few extra tokens, though, and for a smaller organization they can be the difference between a product that pays for itself and one that quietly does not. So we went through the published analyses, the rate cards, and the tokenizer differences from both companies. Here is what we found.
The cheaper-looking model is the more expensive one
The tokenizer is the part of the system that chops your text into billable units. It is also the part nobody reads about before signing up.
According to Solvimon’s pricing comparison, Anthropic’s newer models — Fable 5, Sonnet 5, and Opus 4.7 and later — use a tokenizer that produces roughly 30% more tokens for the same text than OpenAI’s does. Not more words, not longer answers. The same sentence, counted differently.
Now watch what that does to a comparison that looks settled on paper.
The two current flagships are priced within a whisker of each other. Both OpenAI’s gpt-5.6-sol and Anthropic’s Opus 4.8 charge $5 per million input tokens. They differ on output, where Opus is the cheaper of the two at $25 per million against Sol’s $30 — which Layer3 Labs calls the tightest price gap in the flagship lineup.
So take a modest, realistic workload: 1,000 requests a day, roughly 1,000 tokens in and 1,000 out on each. That is about 30 million tokens each way per month. On the published rates, Solvimon puts it at $1,050 a month on Sol and $900 on Opus 4.8 — Anthropic comes out about 14% cheaper, and the decision looks made.
Then apply the tokenizer. Those same requests become roughly 30% more billable tokens on the Anthropic side, taking that $900 to about $1,170 — now around 11% more expensive than the model it appeared to undercut.
Nothing about the workload changed. Nobody wrote longer prompts. The entire swing came from a counting rule that neither rate card mentions.
This is the part worth carrying away: a rate card is a unit price, not a cost estimate. Comparing providers on the published number alone tells you very little, because you do not control how many units your text becomes.
Prices have been anything but stable
Both companies have moved their pricing repeatedly, and in opposite directions.
OpenAI’s flagship input rate has quadrupled in under a year: $1.25 per million when GPT-5 launched in August 2025, $2.50 by March 2026, and $5.00 from late April. Anthropic went the other way first — in November 2025, Opus 4.5 arrived at $5 / $25, down from Opus 4.1’s $15 / $75, a threefold cut — before drifting back upward through new tiers, fast-mode pricing, and the tokenizer change.
Between them, that is six pricing events in twelve months. If you are building anything with a margin attached, the implication is uncomfortable but simple: the cost of your product is a moving target, and someone else is holding it. Usage-aware billing, and knowing which model is doing which job, stops being a nice-to-have.
The compounding cost of a long conversation
There is a third effect that catches people out, and it has nothing to do with which provider you picked.
Language models have no memory between turns. To answer your next message, the model is re-sent the conversation so far — every previous question and answer — because that history is the only thing giving it context. Which means the fiftieth message in a thread is not billed like the first. It is billed for everything that came before it, again. The growth is quadratic, not linear.
IntuitionLabs documents what that looks like in practice. In one analyzed Claude session, a 14-token question cost $0.0018 on the first exchange. The same 14-token question, asked at the 260th exchange, cost roughly $2.41 — a 1,339-fold increase for identical input, driven entirely by the weight of the history dragged along behind it.
Caching changes this arithmetic, and most people never turn it on
That figure assumes something important: that the history is re-processed from scratch every single time. In practice it does not have to be.
Both providers offer prompt caching, which stores the unchanging front portion of a request — the system prompt, the reference documents, the earlier turns — and bills re-reads of it at a fraction of the fresh-input rate. OpenAI’s cached input tier charges one-tenth of the standard price after the first call. IntuitionLabs cites one team that cut roughly 90% of its token spend on Sonnet by turning caching on.
For anything built on long context or repeated reference material, this is the single largest cost lever available, and it is usually a configuration change rather than a rewrite. It is also the one most often left switched off, because nothing on the pricing page tells you it is there.
What actually reduces the bill
Once you understand where the tokens go, the optimizations are unglamorous and effective. These come from IntuitionLabs’ analysis and match what we see in practice:
- Say it in fewer words. Prompts are written once and then run thousands of times, so tightening them compounds. One documented case cut a prompt from 25 tokens to 7 — $0.025 down to $0.007, a 72% saving on every call from then on, and the response came back in half the time.
- Send only the context that earns its place. Attaching the whole document because it was easier than finding the relevant section is a decision you pay for on every single request.
- Break long inputs into coherent chunks rather than one undifferentiated block, so you can send the part that matters.
- Cache the stable parts. See above. This is the big one.
- Use the tooling. Claude Code, for instance, provides
/clearto reset a conversation,/compactto condense it,/costto see what a session has actually spent, and a.claudeignorefile to keep whole directories out of the context in the first place.
None of this is exotic. It is housekeeping — and like most housekeeping, it goes undone until somebody owns it.
What this means if you are running a smaller business
Three things are worth taking from all of this.
Do not choose a provider from the rate card. As the worked example above shows, the published price can point you at the more expensive option. Tokenizer behavior, caching, and your own conversation patterns move the real cost far more than the headline number does.
Treat token usage as an operating cost, not a technical detail. It behaves like one. It grows with your success, it varies with decisions engineers make quietly, and nobody notices until the invoice does something surprising.
Get the design right before the volume arrives. Almost everything above is cheap to fix early and expensive to retrofit once a product is live and the traffic is real.
The encouraging part is that none of this requires you to become an expert in token economics. It requires that someone who understands it looks at your setup before the bill teaches you the same lesson more slowly. If you are weighing providers, or you have a bill that has grown faster than your usage, that is a conversation worth having — and it is the sort of question we spend most of our time on.
For a related look at what happens when cheap AI turns out not to be cheap, see what DeepSeek’s price rise revealed about pricing for smaller businesses.