AI Model Pricing: Every Model, Per-Token & Per-Seat

Frequently Asked Questions

• How much do AI models cost per token?
  Models are priced per million tokens, split into input (what you send) and output (what the model writes back). As of the verified date on this page, that spans a very wide range: from about $0.035 / $0.14 per million on the cheapest tier (Amazon Nova Micro) up to $10 / $50 on the most capable public models (Claude Fable 5). Output tokens almost always cost several times more than input tokens.
• Which AI model is cheapest per token?
  Amazon Nova Micro is the cheapest on the table at $0.035 per million input and $0.14 per million output tokens, with Nova Lite ($0.06 / $0.24) and DeepSeek V4 Flash ($0.14 / $0.28) close behind. Among the big-lab flagships, GPT-5.6 Luna ($1 / $6) and Claude Haiku 4.5 ($1 / $5) are the low-cost options; Claude Opus 5 and Opus 4.8 (both $5 / $25) are the value picks when you need a top-tier model.
• GPT-5.6 vs Claude Fable 5 — which is cheaper?
  GPT-5.6 Sol is cheaper per token ($5 input / $30 output) than Claude Fable 5 ($10 / $50). Both are generally available and deployable today, so the choice comes down to capability, safety design, and compliance rather than access. See our full GPT-5.6 vs Claude Fable 5 comparison for the detail.
• What does "per million tokens" actually mean for my bill?
  A token is roughly ¾ of a word, so one million tokens is about 750,000 words. If a task sends 2,000 tokens of prompt and gets back 500 tokens on Claude Fable 5, it costs 2,000/1,000,000 × $10 + 500/1,000,000 × $50 = $0.045 — about 4.5 cents. Multiply by your monthly request volume to estimate spend, and remember output is the expensive side.
• Can I actually use GPT-5.6 today?
  Yes. GPT-5.6 (Sol, Terra, Luna) is generally available in the OpenAI API and Codex as of July 2026 — no waitlist. The prices above are GA list prices. Claude Fable 5 and Opus 4.8 are likewise generally available and deployable today.
• Are these prices current?
  Every row carries a verified date; the table above was last verified Jul 31, 2026 against vendor pricing pages. AI prices change often, so treat this as a dated snapshot and confirm the current number on the vendor site before you commit budget.
• What is the difference between API token pricing and a subscription plan?
  API (per-token) pricing is usage-based — you pay for exactly the tokens you send and receive, which suits automations and products. A subscription plan (ChatGPT Plus, Claude Team, etc.) is a flat per-seat monthly fee for people using the chat app, and is where admin controls and data terms live. Most businesses use both: seats for staff, API for automation.
• Does prompt caching lower these prices, and by how much?
  Prompt caching can cut input costs 80-99% for repeated prompt content, and the discount varies by vendor. Claude Sonnet 5's cached input price drops to $0.20 per million tokens, versus $2 standard (Anthropic). DeepSeek V4 Flash drops from $0.14 to about $0.003 per million on a cache hit (DeepSeek). Some vendors also offer a separate discount for batch or asynchronous processing. Confirm current batch terms on the vendor's own pricing page. The table above shows standard, real-time, non-cached rates only.
• Can I try any of these models for free before I commit budget?
  Most AI vendors offer a free tier in their chat app, separate from the paid API pricing shown here. ChatGPT, Claude, and Gemini each have a free tier with usage limits (OpenAI, Anthropic, Google). The per-token API pricing above applies once you connect programmatically, usually with a paid account. Confirm current free-tier limits on each vendor's own pricing page before you commit budget.
• Does my chat history get billed again on every message?
  Yes. Most chat and agent APIs resend the full conversation as input tokens on every turn, so a long back-and-forth costs far more than a single message suggests. A 20-turn conversation with a 500-token history re-sends roughly 10,000 extra input tokens by the final turn, on top of that turn's new prompt. Prompt caching (see above) is the main way to cut that repeated-context cost.
• Are there costs beyond the token price shown here?
  Yes. This table prices standard chat completions only; it does not include embeddings calls, fine-tuning or training jobs, vector storage, or the per-request fees some vendors add for tool use like web search. Budget those separately and confirm current rates on the vendor's own pricing page before you commit spend.
• How does reasoning effort affect cost on OpenAI vs Anthropic models?
  It does not change the per-token price on either vendor. OpenAI and Anthropic both bill reasoning tokens as regular output tokens at the model's standard output rate. A higher reasoning-effort setting makes the model generate more output/reasoning tokens before it answers, and that extra volume is what raises the bill, not a separate reasoning surcharge. See the chart above for how the fixed output-token rate compares across GPT-5.6 and Claude tiers.