Benchmark · 2026
LLM Token Cost by Provider
What does a million tokens actually cost across the major providers? Claude Opus $15/M input, GPT-4 $10/M, Gemini Pro $3.50/M, Claude Sonnet $3/M, GPT-4o-mini $0.15/M. Output is 3–5× input across all providers.
By the sipi.bot engineering team · Published 2026-07-18 · Methodology
Token pricing by model
Pricing as of Q2 2026, per million tokens, USD. Prices are list pricing from provider websites and exclude volume discounts and enterprise deals.
| Provider | Model | Input ($/M) | Output ($/M) | Output/Input ratio | Tier |
|---|---|---|---|---|---|
| OpenAI | GPT-4o-mini | $0.15 | $0.60 | 4.0× | Budget |
| Anthropic | Claude Haiku | $0.25 | $1.25 | 5.0× | Budget |
| Gemini Flash | $0.075 | $0.30 | 4.0× | Budget | |
| OpenAI | GPT-4o | $2.50 | $10.00 | 4.0× | Mid |
| Anthropic | Claude Sonnet | $3.00 | $15.00 | 5.0× | Mid |
| Gemini Pro | $3.50 | $10.50 | 3.0× | Mid | |
| OpenAI | GPT-4 | $10.00 | $30.00 | 3.0× | Premium |
| Anthropic | Claude Opus | $15.00 | $75.00 | 5.0× | Premium |
The three tiers
Budget tier ($0.075–$0.25/M input)
GPT-4o-mini, Claude Haiku, Gemini Flash. Use these for high-volume, low-stakes work: classification, routing, simple extraction, formatting. An agent making 10,000 classification calls per day at $0.15/M and 500 tokens per call spends $0.75/day — negligible.
Mid tier ($2.50–$3.50/M input)
GPT-4o, Claude Sonnet, Gemini Pro. The workhorses for most production agents — capable enough for tool use, multi-step reasoning, and coding. This is where most production agent spend lands. An agent making 1,000 tool-augmented Q&A calls per day at $3/M and 2,500 tokens per call (input + output) spends ~$7.50/day.
Premium tier ($10–$15/M input)
GPT-4, Claude Opus. Reserve for the hardest tasks — complex research, novel reasoning, high-stakes decisions. An agent making 100 research calls per day at $15/M and 8,000 tokens per call spends ~$12/day. Easy to justify for quality; hard to justify for routine work.
Provider pricing trends
Provider token pricing has fallen roughly 40–60% over the last 12 months across all tiers. The budget tier has seen the steepest relative cuts (GPT-4o-mini launched at $0.15/M, roughly 1/10th of GPT-4-mini's launch price). The premium tier has been relatively stable — Claude Opus and GPT-4 remain expensive because the underlying capability is still scarce.
The practical implication: agent cost-per-task is falling, but the runaway tail (the top 1% of tasks that cost 50× the median) is not improving — see cost per task. Cheaper models make routine work cheaper; they do not make retry loops less damaging.
Methodology
- Provider list pricing as published on OpenAI, Anthropic, and Google AI websites as of 2026-07-18. Volume discounts, enterprise deals, and cached-prompt pricing are excluded.
- Output/Input ratio is calculated from list pricing; providers structure their pricing consistently so the ratio is stable.
Pricing changes frequently — verify current pricing on the provider's website before making decisions based on these numbers. Provider pricing changes often; re-check each provider's pricing page before relying on these figures. Released under CC BY 4.0 — cite sipi.bot.
How to use this benchmark
- Right-size your model. Most tasks don't need the premium tier. Move routine work to the budget or mid tier.
- Watch output pricing. For long-output agents, output pricing can dominate — pick a model with a low output/input ratio.
- Compose with a firewall. Token pricing is per-call. A retry loop multiplies that per-call cost 40×. A velocity rule prevents the multiplication.