OpenAI vs Anthropic: what LLM API costs actually differ

The rate cards are nearly a wash — deliberately

At the flagship tier the sticker prices are close enough to be a marketing decision: GPT-5.4 at $2.5/15 per 1M tokens, Claude Sonnet 4.6 at $3/15. Identical output price, 20% apart on input. If you compare providers by rate card, you’ll conclude it barely matters. Three structural differences — none of them on the pricing page — decide the real bill.

1 · The tokenizer: Claude bills more tokens for the same text

Providers meter their own tokenizers, and they don’t tokenize alike. For typical English prose, Claude’s tokenizer produces roughly 10% more tokens than OpenAI’s o200k_base for the same input. Stack that on the sticker gap and Sonnet’s effective input price lands at ~1.32× GPT-5.4’s — not the 1.2× the rate card implies. This is the single most-missed line item in provider comparisons.

2 · The cache-write premium: Anthropic charges to warm the cache

Both providers discount cached input ~90%. The difference is on the write side: Anthropic bills first-time cache writes at 1.25× input ($3.75/1M on Sonnet); OpenAI charges no write premium at all. At low cache hit rates you pay Anthropic’s premium without earning it back — the two providers’ caching economics only converge once your hit rate is high. If your traffic is spiky, that asymmetry quietly favors OpenAI.

3 · The budget tier is where the % gap is widest

GPT-5.4 mini ($0.75/4.5) undercuts Claude Haiku 4.5 ($1/5) by ~25% on the rate card — and the tokenizer gap applies here too, so the effective spread is wider. High-volume, budget-tier workloads (simple chat, classification, summarization) are where the provider choice moves the most dollars in relative terms.

Reading the workload table honestly

On the realistic-column defaults above (8% retries, 40% cache hits, 15% infra overhead), the OpenAI models come out ahead at both tiers — e.g. ~$490/mo vs ~$628/mo (+28%) for a history-carrying chatbot at 1,000 requests/day on the flagships. That’s a real, recurring gap. It is not, however, a verdict — three things legitimately flip it:

• Quality fit on your task. If Sonnet resolves your tickets in one turn where GPT-5.4 takes two, the per-token gap is irrelevant — $/completed-task is the metric that matters.
• High, steady cache hit rates. Past the write premium’s break-even, Anthropic’s caching economics are as strong as anyone’s — a stable-prefix, high-traffic workload narrows the gap materially.
• Long context. Sonnet’s 1M-token context at standard pricing changes the architecture question for document-heavy workloads — sometimes eliminating a RAG pipeline’s worth of calls entirely.