GPT-5 vs Claude: the entire cost gap is on the input side

Start here: the cheaper rate card is the more expensive model

Claude Sonnet 5.5 undercuts GPT-5.4 on both sticker lines — $2 against $2.5 on input (20% less), and $10 against $15 on output (33% less). Read the rate card and you would conclude Claude is the budget option at the flagship tier. The workload table above says otherwise, and the reason is not on either pricing page.

Anthropic’s current tokenizer emits roughly 43% more tokens than OpenAI’s o200k_base for the same English text. Price per token is therefore the wrong unit. Per unit of actual content, Sonnet costs 1.14× GPT-5.4 on input and 0.95× on output — so the 33% output discount lands as a near-wash, and the 20% input discount inverts into a premium. That is the whole story, and it’s why the gap moves around between workloads.

The gap tracks your input:output ratio

Because output lands at roughly parity once the tokenizer is applied and input does not, the flagship premium scales with how input-heavy your traffic is. Same three workloads as the table above, sorted by input:output ratio:

workload                in:out    Sonnet vs GPT-5.4
Simple chatbot          2.0:1     +3%
Multi-step agent        2.5:1     +4%
Chatbot with history    5.0:1     +10%

Monotonic, and not a coincidence. Push further in either direction and the pattern holds: a long-form generation workload (say 500 in / 2,000 out — an inverted ratio) narrows the premium well below the 3% floor above, while anything that stuffs context — RAG, long transcripts, document analysis — widens it. If you are choosing between these two, the first question isn’t “which is cheaper,” it’s how much input do I send per unit of output.

The bigger lever: tier, not vendor

Here is the uncomfortable framing for a page like this one. On the history-carrying chatbot above, GPT-5.4 runs $490/mo and Sonnet $538/mo — a 10% spread. Claude Haiku 5.5, Anthropic’s budget tier, runs the same workload for $27/mo. That is ~94% below GPT-5.4, the other vendor’s flagship.

In other words: moving down a tier saves you multiples of what switching vendor at the same tier does. If cost is the binding constraint, “GPT-5 or Claude?” is the second question. The first is whether your task actually needs a frontier model at all. The flagship-vs-flagship decision matters when quality is the constraint — which is a real and common situation, just not a cost question.

Where each one earns its price

• GPT-5.4 — cheaper per unit of content, and no cache-write premium. The tokenizer gap is what buys it the input advantage despite the higher sticker price. Anthropic also bills first-time cache writes at 1.25× input, which matters on bursty traffic where you keep paying to warm a cache you don’t reuse.
• Sonnet 5.5 — far better cache reads, and now the larger context. Cached input at $0.1/1M is 0.05× base input — half the usual Anthropic discount and the steepest of any OpenAI, Anthropic or Google model here — so a high, steady hit rate erodes the input premium quickly. The context comparison has also flipped: 1000K at standard pricing against GPT-5.4’s 272K.

Note the tension in those two bullets: Anthropic charges a $2.5/1M cache write where OpenAI charges nothing, but pays it back with a deeper read discount. Which side of that trade you land on depends entirely on your hit rate — so model it rather than guessing. The caching deep-dive works through the break-even.