Model vs model, real cost
LLM API cost comparisons
The per-token rate is the easy part. What actually decides your bill is the input/output mix of your workload. Each comparison below runs both models across the same four real jobs — and stays current, because every page reads live from our pricing data instead of a blog post that went stale six months ago.
- Comparisons
- 10
- Workloads each
- 4
- Pricing
- Live
Every head-to-head
Claude Opus 4.8 vs GPT-5.5
Anthropic vs OpenAI — cost per 1,000 calls across chat, RAG, long-doc summary, and bulk classification.
Compare cost →Claude Opus 4.8 vs Gemini 3.1 Pro
Anthropic vs Google — cost per 1,000 calls across chat, RAG, long-doc summary, and bulk classification.
Compare cost →GPT-5.5 vs Gemini 3.1 Pro
OpenAI vs Google — cost per 1,000 calls across chat, RAG, long-doc summary, and bulk classification.
Compare cost →Claude Sonnet 4.6 vs GPT-5.4
Anthropic vs OpenAI — cost per 1,000 calls across chat, RAG, long-doc summary, and bulk classification.
Compare cost →Claude Sonnet 4.6 vs Gemini 3.1 Pro
Anthropic vs Google — cost per 1,000 calls across chat, RAG, long-doc summary, and bulk classification.
Compare cost →GPT-5.4 vs Gemini 3.1 Pro
OpenAI vs Google — cost per 1,000 calls across chat, RAG, long-doc summary, and bulk classification.
Compare cost →GPT-5.4 Mini vs Gemini 3.5 Flash
OpenAI vs Google — cost per 1,000 calls across chat, RAG, long-doc summary, and bulk classification.
Compare cost →Claude Haiku 4.5 vs GPT-5.4 Mini
Anthropic vs OpenAI — cost per 1,000 calls across chat, RAG, long-doc summary, and bulk classification.
Compare cost →DeepSeek V4 Pro vs Claude Sonnet 4.6
DeepSeek vs Anthropic — cost per 1,000 calls across chat, RAG, long-doc summary, and bulk classification.
Compare cost →DeepSeek V4 Pro vs GPT-5.5
DeepSeek vs OpenAI — cost per 1,000 calls across chat, RAG, long-doc summary, and bulk classification.
Compare cost →
How we compare
Every comparison prices the same four workloads, so the input-heavy vs output-heavy trade-off between two models is visible at a glance. A model that's cheapest on short chats can be the expensive one on long-document summaries — the mix is the whole game.
Short chat turn
1K in / 500 out — a typical assistant reply
RAG answer
8K in / 800 out — retrieved context + grounded answer
Long-doc summary
50K in / 2K out — summarize a long document
Bulk classification
2K in / 50 out — label/route at high volume
Frequently asked questions
- Why does the cheapest LLM depend on the workload?
- Because providers price input and output tokens differently — often output costs 3–5x more than input. A model that's cheapest for short, input-heavy chats can be the most expensive for long-document summaries that generate lots of output. Each comparison here prices the same four real workloads (chat, RAG, long-doc summary, bulk classification) so the trade-off is visible instead of hidden behind a single per-token headline.
- Are these prices current?
- Yes — every comparison reads live from our pricing data module rather than being typed into a blog post, so the numbers update when provider pricing changes. We re-verify the underlying rates against each provider's published pricing page on a regular cadence.
- Does prompt caching change the comparison?
- Significantly, for input-heavy workloads. Cached input tokens are billed at roughly 10% of the normal input rate by most providers, so a model with a large repeated system prompt can become far cheaper once caching is applied. Use the LLM API Cost Calculator to model your own cache-hit ratio.
- Which model should I default to in production?
- Pick the one that's cheapest on the workload that dominates your actual traffic, not the one with the lowest headline rate. Find your dominant workload in the four profiles below, then read the verdict on the relevant comparison page — the cheapest model often flips between chat-heavy and summary-heavy products.