OpenAI vs Anthropic vs Google: the cost of 100,000 identical requests
Abstract “per million tokens” numbers are hard to compare with a real budget. Take a concrete job — 100,000 requests, 2,000 input and 500 output tokens each — and compute straight money.
Flagships
GPT-5.5 ($5/$30): 100,000 requests = $2,500. Claude Opus 5.5 ($4/$20): $1,800. Claude Sonnet 5.5 ($2/$10): $900. Gemini 4 Argon on promo ($2/$10): $900, $1,800 after December 1 ($4/$20). Opus is exactly 2× Sonnet — that is the price of “maximum quality” over “very good”.
Middle class
Gemini 3.8 Flash ($0.75/$3.75 until end of 2026): $338 — the cheapest near-Sonnet option. GPT-5.6 Terra ($2/$12): $1,000. DeepSeek V3.1 ($0.27/$1.10): $109 — an order of magnitude cheaper with quality enough for RAG and summarisation.
Budget tier
Claude Haiku 5.5 ($0.10/$0.50): $45. GPT-5.6 Luna ($0.20/$1.20): $100. On these models the whole batch costs one day of GPT-5.5. If your task fits a simple prompt, the 20–50× gap is the argument for a “budget first, flagship if needed” cascade.
Caching reshuffles the table
With a shared system prompt and 70% cache hits, GPT-5.5 drops from $2,500 to about $1,870 and Claude Opus 5.5 from $1,800 to $1,270. On long-context agents caching saves more than switching vendors: turn caching on before comparing lists.
Bottom line
The spread for one workload is $45 to $2,500 — 55×. Prices move: Gemini 3.8 Flash rises in January 2027, GPT-5.6 Sol runs on promo only until November 30. Live third-party quotes and uptime monitoring are in the model catalogue; the per-model calculator re-prices any token mix.
Published 2026-11-11. Prices are vendor list prices on the publish date; live quotes are in the catalogue.