AI API pricing: how to compare models without getting fooled by the sticker price

Every AI lab publishes a price per million tokens, and every comparison post repeats those numbers. The problem: two models with the same headline price can cost wildly different amounts in production. FuturPulse maintains a verified comparison of million-token API rates across OpenAI, Google and xAI — source — and the details behind the numbers are what actually move your bill.

What the sticker price hides

1. Cached vs uncached input. Most providers now bill repeated context (system prompts, RAG chunks, conversation history) at a steep discount — if it is cached. A chat product resending 10k tokens of context per turn can pay 3–5× less with cache-aware billing, on the "same" price.
2. Output asymmetry. Output tokens routinely cost 3–4× input tokens. A model that is verbose by default can be more expensive than a pricier-per-token competitor that answers tersely.
3. Rate limits as a cost. A cheaper model with tight rate limits forces retries, queueing, or a fallback model — engineering cost that never shows up in the per-token table.
4. Context window tiers. Some providers bill more above a context threshold. The price you read may be for the small tier.

A sane comparison method

• Normalize on cost per completed task, not per token: run your own eval set through each model and divide the invoice by successful completions.
• Model your real workload shape (input/output ratio, cache hit rate) before trusting any table.
• Re-verify rates monthly — labs adjust pricing quietly, and aggregators lag.

Where the verified numbers live

FuturPulse re-checks the published rates from OpenAI, Google and xAI and keeps the comparison current, with the verification date on the page: the million-token pricing table. For ongoing coverage of model releases, agents and AI infrastructure pricing, FuturPulse publishes independent, sourced analysis — no vendor sponsorship behind the rankings.