og.cc.cd / guides / llm-context-window-comparison

LLM Context Window Comparison: GPT-6, Claude 5.5, Gemini 4, DeepSeek and Qwen

By og.cc.cd · Published · Updated · 4 min read

TL;DR: Most flagship models now list about a 1M-token context, but context window, maximum input and maximum output are three different limits, and some models charge more once a prompt crosses a threshold.

Updated · All sources ↓

Most flagship models now advertise a context of about a million tokens, but “context window” hides three different limits and, for some models, a price jump. Here are the current numbers from each vendor’s own model pages and what they mean in practice.

Count your prompt’s tokens →All guides

Three numbers, not one

  • Context window — the total number of tokens the model can handle in one request. On some models input and output share it (Google calls Gemini 4 Argon’s limit a “unified input and output token limit”).
  • Maximum input — how much you can send. OpenAI lists it separately: GPT-6 models have a 1,050,000-token window but accept at most 922,000 input tokens.
  • Maximum output — how much the model may generate in one response, including reasoning tokens on most thinking models. This is usually far smaller than the window.

The comparison (checked 2026-10-20)

ModelContext windowMax outputNotes
gpt-6-astra1,050,000128,000Max input 922,000; prompts >272K billed at 2× input, 1.5× output
gpt-6.1-sol1,050,000128,000Max input 922,000; same long-prompt pricing
gpt-6-luna1,050,000128,000Max input 922,000; same long-prompt pricing
gpt-5.3-codex400,000128,000Max input 272,000
Claude Fable 5.11M128KFull 1M at standard pricing
Claude Opus 5.51M128KUp to 300K output on the Batches API with a beta header
Claude Sonnet 5.51M128KUp to 300K output on the Batches API with a beta header
Claude Haiku 5.51M128KPrompts over 100K tokens pay higher prices
gemini-4-argon1,048,57665,536Unified input+output limit; up to 1,048,576 output with long decode
gemini-3.8-flash1,048,576 (input)65,536
gemini-3.5-flash-lite1,048,576 (input)65,536
deepseek-v4-pro1M384K
deepseek-flash1M384K
qwen3.8-max1M—Output limit not listed on the overview page
qwen3.7-plus1M—Prompts over 256K billed at a higher tier
qwen3.8-flash1M—

Sources: OpenAI model pages (gpt-6-astra, gpt-6.1-sol, gpt-6-luna, gpt-5.3-codex), Anthropic models overview, Google model pages (gemini-4-argon, gemini-3.8-flash, gemini-3.5-flash-lite), DeepSeek models & pricing and Alibaba Cloud Model Studio text generation models. “1M” is written exactly as each vendor writes it.

Reading the output column

Output limits differ more than context windows do. OpenAI’s GPT-6 models and Anthropic’s current models allow 128K output tokens per synchronous request; DeepSeek lists up to 384K; Google’s Gemini models default to 65,536, with Gemini 4 Argon able to go further using long decode. On thinking models the reasoning tokens are generated output too, so a deep reasoning pass can use a large share of that budget before the visible answer starts. If you need a long document generated, check this column first — a 1M-token window does not help if the answer is capped at 64K.

How much text is a million tokens?

It depends on the tokenizer and the language. Anthropic says 1M tokens is roughly 555k words or 2.5M Unicode characters on Claude’s current tokenizer, while models before Claude Opus 4.7 fit about 750k words. Alibaba Cloud gives about 1.5 million Chinese characters per 1M tokens for Qwen, as an illustration only. Paste a representative document into the token counter to measure your own ratio.

Long context changes the price

A big window does not mean every token costs the same. Several models switch to a higher rate once a single prompt crosses a threshold, and that rate applies to the whole request:

  • OpenAI GPT-6 family: prompts with more than 272K input tokens are priced at 2× input and cache rates and 1.5× output for the full request.
  • Claude Haiku 5.5: a request whose prompt is over 100,000 tokens pays the higher price set; cache reads and writes count toward the prompt length. Anthropic says its other current models bill the full 1M window at standard rates.
  • qwen3.7-plus: a separate price tier applies from 256K to 1M input tokens.

Our cost calculator picks the right tier from your average input size, so you can see the jump before it shows up on an invoice.

Check the size before you send

If a request might be close to a limit or a price threshold, count it first. OpenAI offers POST /v1/responses/input_tokens, Anthropic POST /v1/messages/count_tokens and Google models.countTokens; each counts the full request, including tools and attachments. Our tokens guide explains the differences. Every response’s usage field then tells you what was actually billed.

Matching the window to the job

WorkloadWhat usually matters
Chat and support botsPrice per token and latency; conversations rarely approach 100K tokens if you summarise history
Coding agentsLarge window plus cheap cache hits, because the same files and instructions are re-sent on every step
Whole-document analysisMaximum input, the long-prompt price tier, and enough output budget for the answer
Long generated reportsMaximum output (64K–384K across the models above), not the window

Should you fill the window?

Usually not. Every token you send is billed on every request, larger prompts take longer to process, and the model has to find the relevant part among everything else. Practical rules:

  1. Retrieve, then send. Search or filter documents first and pass only the relevant chunks.
  2. Keep a stable prefix (system prompt, tool definitions, reference docs) at the start so prompt caching can make repeat requests cheaper — see our prompt caching guide.
  3. Leave room for the answer. Reserve the output (and reasoning) budget you need before you count how much input fits.
  4. Summarise long conversations instead of re-sending the entire history on every turn.

Numbers in this table change with new model releases. Each row links to the vendor page it came from — check it before you design around a limit.

Sources

Vendor features and prices were checked against these official pages on 2026-10-20. Products change; if something here no longer matches, the official page is right.

Related guides