Pricing & Cost Reference

Pricing is a Frontier Model Comparison snapshot. This section is the durable part: how to reason about cost so you're not surprised by a bill. As of July 2026.

How LLM pricing works. You pay per token, and input and output are priced separately (output is typically 3 to 8× input, because it's generated sequentially). "Tokens" ≈ ¾ of a word in English. A back-of-envelope estimate: cost ≈ (input_tokens × in_price + output_tokens × out_price) / 1,000,000, multiplied by your request volume.

The four levers that cut cost the most:

1. Prompt caching. Reuse of a repeated prefix (system prompt, big context) at a large discount (often ~90% off input, sometimes ~99% for cache hits, e.g. DeepSeek). The single biggest win for agents and RAG that resend the same context. Structure prompts so the stable part comes first.
2. Batch APIs. Non-urgent jobs (evals, bulk generation) at ~50% off, with a delayed SLA. Most providers offer it.
3. Right-sizing the model. Use a mini/flash/haiku tier for routing, extraction, and classification. Reserve the flagship for genuinely hard steps. A router automates this.
4. Reasoning budget. Hybrid models bill their hidden "thinking" tokens as output. Cap the reasoning effort where you don't need it. It's often the largest line item on agentic workloads.

Watch-outs. Long-context tiers often cost ~2× above a threshold (e.g. >200K tokens). Newer models may use a different tokenizer. Anthropic's newest generation reportedly emits ~30% more tokens for the same text, so compare cost per task, not just the headline per-token price. Vision/audio inputs are often billed as a large fixed token count per image/second.

Ballpark tiers (per 1M output tokens, July 2026): budget/open ≈ $0.30 to $3 (DeepSeek, Flash/Haiku/mini, self-host), mid ≈ $5 to $15 (Sonnet, Gemini Pro, GPT balanced), premium ≈ $25 to $90 (Opus, GPT Pro tiers). For live cost-per-task comparisons across providers, use Artificial Analysis and OpenRouter.