How Much Does 1 Million Tokens Cost? (2026 LLM Pricing)
The price of a million input and output tokens on every major model — plus what a million tokens actually holds.
Updated Aug 5, 2026 · Reviewed by PromptCostLab
Token and pricing information changes frequently. Use this guide for planning and verify model-specific limits and prices with the provider before production billing.
What a million tokens is
Scale first: a million tokens is roughly 750,000 English words — ten novels, a large documentation set, or a solid codebase. It is also the size of the largest current context windows, so '1M tokens' is simultaneously a unit of library scale and a single request's ceiling.
Providers price per million tokens, separately for input and output. Every number below was verified against the provider's own pricing page in August 2026.
Input: what 1M tokens cost to send
The spread across the market is enormous — nearly 100x from floor to ceiling:
- DeepSeek V4 Flash: $0.14 (cache hits: $0.0028)
- DeepSeek V4 Pro: $0.435 (cache hits: $0.003625)
- Gemini 3.6 / 3.7 Flash: $0.75 intro through Dec 2026, then $1.50
- Kimi K2.7 Code: $0.95 (cache hits: $0.19)
- Mistral Large 3: $0.50 · Llama 4 Maverick (hosted): $0.20
- Claude Sonnet 5: $2.00 · GPT-5.6 Terra: $2.00 · Grok 4.6: $2.00
- Gemini 3.1 Pro: $2.00 up to 200K, $4.00 above
- Claude Opus 5: $5.00 · GPT-5.6 Sol: $5.00 (both $10 above their long-context thresholds)
- Claude Fable 5 / Mythos 5: $10.00
Output: the expensive side
Generated tokens cost 3–6x input everywhere. The range for one million output tokens: DeepSeek Flash $0.28, Gemini 3.6/3.7 Flash $3.75 (intro), Kimi K2.7 Code $4.00, Mistral Large 3 $1.50, Grok 4.6 $6.00, Claude Sonnet 5 $10.00, GPT-5.6 Terra $12.00, Gemini 3.1 Pro $12.00 ($18 above 200K), GPT-5.6 Sol $30.00, Claude Opus 5 $25.00, and Claude Fable 5 $50.00 — where the always-on thinking tokens also land.
Practical implication: controlling output length (structured answers, sensible max-output limits) often saves more than switching input models.
A million tokens in a real workload
A documentation assistant ingesting 1M tokens of input per day on Sonnet 5 with a 70% cache-hit rate pays about $0.86/day in input (~$26/month). The same traffic uncached is $2/day. If it also generates 100K output tokens a day, add $1/day. On Fable 5 instead, the identical workload costs roughly $13/day uncached — the model choice moved the bill 6x more than caching did.
The point of per-million pricing is exactly this arithmetic. The PromptCostLab comparison calculator applies it across every registry model with your volumes, cache share, and long-context thresholds included.
Watch the fine print
Three adjustments move real bills away from headline rates: long-context tiers (OpenAI above 272K, xAI at 200K, Gemini 3.1 Pro above 200K), batch discounts (typically 50% off at Anthropic, Google, Mistral), and request-level fees outside token pricing (Perplexity's search-context charges; Gemini's hourly explicit-cache storage). Region and data-residency surcharges add single-digit percentages. Re-verify quarterly — this market reprices constantly.
How much does 1 million tokens cost?
It depends entirely on the model and direction. In August 2026, 1M input tokens cost $0.14 on DeepSeek V4 Flash, $2 on Claude Sonnet 5 or GPT-5.6 Terra, and $10 on Claude Fable 5. Output is pricier: $0.28 (DeepSeek Flash) to $50 (Fable 5) per million. Cached input is dramatically cheaper — often 10% of the input price or less.
What can 1 million tokens hold?
About 750,000 English words — roughly ten novels, 2,000 typical blog posts, or a mid-sized codebase. On current 1M-context models it is also the entire window: everything you send plus everything the model replies must fit inside it.
Why do token prices differ so much between models?
Flagship models charge for capability and compute (Fable 5 at $10/$50), fast models charge for efficiency (Gemini Flash intro at $0.75/$3.75, DeepSeek Flash at $0.14/$0.28). The spread between the cheapest and most expensive output token is over 100x, which is why model routing matters more than micro-optimizing prompts.
Prices, token behavior, and model limits change. These primary provider docs are the right place to verify the details before shipping.