# LLM API pricing, compared honestly - published: 2026-08-31 - last verified: 2026-09-10 - read: 7 min - keywords: LLM API pricing comparison, cheapest LLM API, Claude API pricing, GPT API cost, Gemini API pricing - url: https://rotative.xyz/deep-dive/llm-api-pricing-comparison - site: Rotative (rotative.xyz) — evidence-first AI news --- LLM pricing looks simple and is not. The sticker price is per million tokens, but what you actually pay depends on input versus output mix, caching hits, batch windows, and the fact that thinking models bill you for their invisible reasoning. Two apps with identical sticker prices can differ 10x in real cost. This is the honest map. Numbers drift month to month - that is the point of the maintenance note at the end - but the structure of pricing changes much slower than the numbers. **The short version** - Output tokens cost roughly 3-5x input tokens everywhere. Apps that generate long output pay disproportionately. - Prompt caching cuts input costs 50-90 percent on repetitive workloads - the biggest lever most teams ignore. - Batch APIs cut costs another 50 percent for anything that tolerates waiting hours. - Thinking and reasoning tokens are billed as output even though you never see them. Budget accordingly. - The open-weight tier made "good enough" nearly free: for classification, extraction, and summarization, the frontier is often unnecessary spend. ## The four levers Input price. What you send. Context windows are huge now, and long context means input dominates many bills. Output price. What the model writes. Typically 3-5x input price, and the real multiplier for reasoning models, which bill hidden chain-of-thought as output. Caching. Providers discount cached input sharply - on the order of 50-90 percent off - when your requests repeat a stable prefix: system prompts, documents, tool definitions. Any app with a fixed system prompt and variable user input should be engineered for cache hits. This is free money. Batching. Async batch endpoints run large jobs at typically half price, with turnaround of minutes to hours. Classification backlogs, embeddings, document processing - anything not real-time should already be on batch. ## The tiers that actually exist Forget vendor names for a second; the market has settled into four functional tiers. Frontier reasoning. The strongest models from Anthropic, OpenAI, and Google. Priced like senior consultants: think multiple dollars per million input tokens and an order of magnitude more for output. Worth it for hard reasoning, complex code, and agentic work where the model is doing steps of thinking. Waste for classification. Frontier speed tier. The same families' smaller variants: cents per million tokens, a large fraction of flagship quality. The default for production chat, extraction, and most summarization. Open-weight value tier. DeepSeek, Llama-family, Mistral, Qwen and peers: near-frontier quality on structured tasks at prices that round to zero, often self-hostable. In 2026 this tier handles an enormous share of global token volume for exactly this reason. Edge/nano tier. Sub-cent models for labeling, routing, moderation. If your latency budget is milliseconds or your volume is billions of calls, live here. ## What people get wrong Sticker-shopping on input price while their bill is output-dominated. Benchmark-shopping while their workload is cacheable. And the reasoning trap: a thinking model that produces a better answer in 5,000 hidden tokens may cost more than a faster model producing a good-enough answer in 300 visible ones. The fix is measurement: log tokens by type per feature. Most teams find two or three call sites drive most spend, and that the frontier model is only earning its price on one of them. ## A decision rule Hard reasoning, complex code, agentic chains: frontier reasoning tier. User-facing chat and extraction at scale: frontier speed tier. High-volume structured work: open-weight. Routing and labeling: nano. Re-evaluate quarterly, because the tier boundaries move every few months - the 2026 speed tier beats the 2024 frontier at a fraction of its price. Maintenance note: exact prices change monthly. Verify against provider pricing pages before committing budgets; treat every third-party price table, including any specific figures, as indicative, not contractual.