LimitDigest
Home / Guides / 1M context LLM API pricing: what a 500K-token prompt costs

1M context LLM API pricing: what a 500K-token prompt costs

Updated 2026-11-17. From each provider's official pricing page.

Quick answer

For a single 500K-token prompt, GPT-6 Luna and Qwen3.7 Flash tie as the cheapest 1M context APIs we track at about $0.10 a request (our estimates at list prices, 4,000 output tokens). The catch is long-prompt pricing: OpenAI doubles input rates past 272K tokens and Claude Haiku 5.5 charges 5x past 100K, while Claude Sonnet, Opus and Fable, Gemini 4, DeepSeek, Kimi K3.1 Pro and Qwen4-Max keep one price across the full 1M window.

A 1M token context window doesn't mean a 1M token prompt costs what the headline price suggests. Several providers switch to a higher rate once a prompt crosses a size line, and it usually applies to the whole request, not just the tokens past the line. Below is what one long request costs on 23 models with a 1M window, using list prices from our API price tables.

The long-prompt rules are confirmed on each provider's pricing page, read Nov 16, 2026. Every dollar figure per request is our estimate and is labeled that way.

What a 100K, 500K and 900K token prompt costs

ModelPer 1M (in / out)Long-prompt rule100K prompt500K prompt900K prompt
GPT-6 Luna$0.10 / $0.50Over 272K: $0.20 in, $0.75 out$0.012$0.10$0.18
Qwen3.7 Flash$0.03 / $0.1332K to 256K: $0.10 / $0.40; over 256K: $0.20 / $0.80$0.012$0.10$0.18
DeepSeek V4.1 Flash$0.30 / $1.20Flat to 1M (half price off-peak)$0.035$0.15$0.27
Muse Ember 1.0$0.10 / $0.50Over 200K: $0.30 in, $1.50 out$0.012$0.16$0.28
Claude Haiku 5.5$0.10 / $0.50Over 100K: $0.50 in, $2.50 out$0.012$0.26$0.46
Qwen3.5 Plus$0.40 / $2.40Over 256K: $0.50 in, $3 out$0.05$0.26$0.46
Gemini 3.8 Flash$0.75 / $3.75Flat to 1M (doubles Jan 1)$0.09$0.39$0.69
Qwen3.6 Flash$0.25 / $1.50Over 256K: $1 in, $4 out$0.031$0.52$0.92
Muse Spark 1.3$1.25 / $4.25Flat to 1M$0.14$0.64$1.14
DeepSeek V4 Pro$1.32 / $3.96Flat to 1M (half price off-peak)$0.15$0.68$1.20
Qwen4-Max$1.50 / $4.50Flat to 1M$0.17$0.77$1.37
Gemini 4 Hydrogen$1.50 / $7.50Flat to 1M$0.18$0.78$1.38
Qwen3.6 Plus$0.50 / $3Over 256K: $2 in, $6 out$0.062$1.02$1.82
GPT-6.1 Terra$1 / $5Over 272K: $2 in, $7.50 out$0.12$1.03$1.83
Claude Sonnet 5.5$2 / $10Flat to 1M$0.24$1.04$1.84
Gemini 4 Argon$2 / $10Flat to 1M (doubles Dec 1)$0.24$1.04$1.84
Kimi K3.1 Pro$2 / $10Flat to 1M$0.24$1.04$1.84
Grok 4.3$1.25 / $2.50Over 200K: double ($2.50 in, $5 out)$0.14$1.27$2.27
GPT-6.1 Sol$2 / $10Over 272K: $4 in, $15 out$0.24$2.06$3.66
Claude Opus 5.5$4 / $20Flat to 1M$0.48$2.08$3.68
Muse Wonder 1.0$4 / $20Flat to 1M$0.48$2.08$3.68
Claude Fable 5.2$8 / $40Flat to 1M$0.96$4.16$7.36
GPT-6 Astra$10 / $50Over 272K: $20 in, $75 out$1.20$10.30$18.30

Cost of one request with 4,000 output tokens, our estimate at list prices from each provider's pricing page (checked 2026-11-17). No caching, batch or off-peak discounts. Whole request billed at the long-prompt rate once the prompt crosses the line.

GPT-6 Luna and Qwen3.7 Flash lead at 500K. The cheapest model with one flat rate for the whole window is DeepSeek V4.1 Flash at about $0.15. At the top end, a 500K prompt on GPT-6 Astra costs about $10.30, about 100 times the cheapest.

Which LLM APIs charge more for long prompts

  • OpenAI: GPT-6.1 Sol, Terra, Astra and GPT-6 Luna bill everything at a higher "long context" rate once input passes 272K tokens, per OpenAI's pricing page. On GPT-6.1 Sol that's $4 in and $15 out instead of $2 and $10. Going from a 270K to a 300K prompt raises the bill about 2.2x, by our math.
  • Anthropic: Claude Haiku 5.5 is the exception in Claude's lineup. Past 100,000 prompt tokens the whole request costs $0.50 in and $2.50 out, even when part of the prompt is a cache hit, per Anthropic's pricing docs.
  • Meta: Muse Ember 1.0 bills requests over 200K input tokens at its long rates in full. Muse Wonder and Muse Spark have no long-context premium, per Meta's pricing page.
  • xAI: Grok 4.3 costs double for prompts of 200K tokens and up, per xAI's models page.
  • Alibaba: most Qwen models price by prompt size. Qwen3.7 Flash has three tiers (up to 32K, 32K to 256K, over 256K), and Qwen3.5 Plus, 3.6 Plus and 3.6 Flash step up past 256K, per Alibaba Cloud Model Studio pricing.

Which 1M context APIs keep one flat price

Anthropic says it plainly: Claude 4.6 and later (except Haiku 5.5) include the full 1M window at standard pricing, and "A 900k-token request is billed at the same per-token rate as a 9k-token request." That covers Sonnet 5.5, Opus 5.5 and Fable 5.2.

Google's Gemini API pricing page lists one rate with no prompt-size tier for Gemini 4 Hydrogen, Gemini 4 Argon and Gemini 3.8 Flash. DeepSeek has one rate for its full 1M window, and Moonshot prices Kimi K3.1 Pro flat across its window too. Alibaba lists Qwen4-Max, launched Nov 16, 2026, at one price for prompts up to 1M.

Two of those flat prices are launch deals. Gemini 4 Argon goes from $2 to $4 per 1M input tokens on Dec 1, 2026, and Gemini 3.8 Flash doubles on Jan 1, 2027, per Google. DeepSeek charges half its peak rate off-peak, which makes it cheaper still for batch jobs; see DeepSeek off-peak hours.

How caching changes long-context API cost

If you ask several questions about the same long document, caching matters more than the headline price. A cache hit bills the repeated part of the prompt at a fraction of the input rate:

ModelCached input per 1M500K prompt, uncached500K prompt, cache hitNote
Claude Sonnet 5.5$0.10$1.00$0.05Anthropic: cache hits at 0.05x base input, full 1M window
Claude Haiku 5.5$0.05$0.25$0.025Anthropic: long-prompt cache hit $0.05
GPT-6.1 Sol$0.20$2.00$0.10OpenAI: long-context cached input $0.20 (short prompts $0.10)
Gemini 4 Hydrogen$0.075$0.75$0.037Google: implicit caching only
DeepSeek V4.1 Flash$0.006$0.15$0.003DeepSeek: cache hit, peak rate

Input cost only, our estimate. Cache writes and storage fees, where a provider charges them, are extra on the first request.

On Claude Sonnet 5.5, re-reading a cached 500K-token document costs about $0.05 in input instead of $1.00. That turns a flat-priced frontier model into a reasonable choice for long-document Q&A, as long as you hit the cache within its window.

Why the same text can cost more tokens on Claude

Token counts aren't comparable across vendors. Anthropic says Claude 4.7 and later use a newer tokenizer that produces about 30% more tokens for the same text than the tokenizer older Claude models used, per its pricing docs. Every vendor counts differently, so a file that is 500K tokens on one model can be noticeably more or less on another. Run your real file through each provider's token counter before you trust a per-request estimate.

Which long-context API should you pick?

  • Cheapest long requests: GPT-6 Luna or Qwen3.7 Flash. Fine for summaries, extraction and search over big files when top-tier reasoning isn't the point.
  • Cheapest flat price: DeepSeek V4.1 Flash. No size line to watch, so your bill scales in a straight line with prompt length.
  • Frontier model over a whole codebase: Claude Sonnet 5.5 or Gemini 4 Hydrogen. Both keep one rate to 1M; a 500K prompt costs about $1.04 and $0.78 respectively.
  • On OpenAI: keep prompts under 272K where you can. Trimming a 300K prompt to 270K roughly halves the input bill on GPT-6.1 Sol.

Our take: if your prompts regularly run past 272K tokens, the flat-priced models win on predictability, and DeepSeek V4.1 Flash wins on price. For shorter, everyday prompts the ranking changes, so check the cheapest LLM API ranking and cheapest AI API for coding.

Sources

FAQ

What is the cheapest LLM API with a 1M token context window?

For a 500K-token prompt, GPT-6 Luna and Qwen3.7 Flash, tied at about $0.10 a request by our estimate. The cheapest with one flat rate across the whole window is DeepSeek V4.1 Flash.

Does OpenAI charge more for long prompts?

Yes. GPT-6.1 Sol, Terra, Astra and GPT-6 Luna bill the whole request at higher long-context rates once input passes 272K tokens, per OpenAI's pricing page.

Does Claude charge more for 1M context?

Not on Sonnet 5.5, Opus 5.5 or Fable 5.2, which keep standard pricing across the 1M window. Claude Haiku 5.5 charges $0.50 in and $2.50 out per 1M tokens once a prompt passes 100K tokens.

Does Gemini have long-context pricing?

Not on Gemini 4 Hydrogen, Gemini 4 Argon or Gemini 3.8 Flash, which list one rate on Google's pricing page. Argon's launch rate doubles on Dec 1, 2026 and 3.8 Flash's on Jan 1, 2027.

How much does a 1M token prompt cost?

About $2.04 on Claude Sonnet 5.5 and $0.30 on DeepSeek V4.1 Flash at peak rates, by our estimate with 4,000 output tokens. Caching cuts repeat reads of the same text sharply.

Is long-context pricing applied to the whole request?

On OpenAI, Claude Haiku 5.5 and Muse Ember, yes: once the prompt crosses the line, every token in that request bills at the higher rate, per their pricing pages.

Related

More guides