What counts as context vs output?

Context is the input budget — everything the model reads before generating: system prompt, prior conversation turns, tool/function definitions, retrieved documents, and the latest user message. Output is what the model writes back. They're billed and limited separately, even though both consume the same underlying attention budget. A 200K-token context window with an 8K output cap means you can feed ~200K tokens in but only get up to ~8K tokens back per call.

Do system prompts count toward the window?

Yes. Every token the model sees consumes the input budget — system prompt, developer instructions, few-shot examples, tool schemas, prior assistant turns, retrieved RAG chunks. The only tokens that don't count are the ones the model emits as output (those count toward the separate output cap). Long system prompts on multi-turn conversations are the #1 cause of mysterious context-overflow errors.

What is RAG and how does it interact with context size?

Retrieval-Augmented Generation (RAG) means fetching relevant passages from a vector store at query time and pasting them into the prompt so the model can ground its answer in fresh or domain-specific facts. Each retrieved chunk eats context budget. A typical RAG setup retrieves 5–20 chunks of 500–1000 tokens each, so 5K–20K tokens of context vanish before the user message is even appended. Plan for that — large windows let you retrieve more aggressively, but precision matters more than volume.

Are reasoning tokens included in the output budget?

For OpenAI's o-series and similar reasoning models, yes — internal chain-of-thought tokens count against the output cap, which is why those caps are dramatically larger (e.g. 100K for o1) than non-reasoning models. The user only sees the final answer, but the model may have generated 10–50× that many hidden tokens to get there.

Why does the same model show different context limits at different vendors?

Open-weight models (Llama, DeepSeek, Mistral) are hosted by many providers (Together, Groq, Fireworks, Bedrock, Vertex), and each picks its own deployment configuration — KV-cache sizing, max sequence length, batching strategy. The headline 128K Llama 3.3 70B context might be 32K on one host and 128K on another. Check your specific provider's docs.

How accurate is the token estimate?

Within roughly ±10% for English prose and code. Real tokenization needs each model's exact BPE or SentencePiece table — Convertitive doesn't ship those in the browser because they'd add tens of megabytes of JavaScript. For final cost or limit sign-off, run your prompt through the vendor's own tokenizer (OpenAI tiktoken, Anthropic's count-tokens endpoint, Google's count_tokens).

Should I always pick the largest window?

No. Bigger windows let you cram more in, but model quality typically degrades past 30–50% utilization — attention becomes diffuse and the model misses details in the middle of long contexts (the 'lost in the middle' effect). For retrieval-heavy workloads, fewer high-quality chunks beat dumping the whole corpus. For analysis tasks, summarize-then-process often beats one giant prompt.

LLM Context Window Visualizer

See exactly how much of each model's context budget your prompt eats.

Buğra SözeriAI

Published May 30, 2026

Every hosted LLM has a hard cap on input tokens — exceed it and the API rejects the request or silently truncates the front of your prompt. The widget below counts the tokens in whatever you paste, then renders a horizontal bar per model showing the share of its context window your input would consume. Bars turn amber above 50% utilization and red above 80%, where prompt-engineering choices start to dominate model quality.

Tokens are estimated with the same heuristic used by the Token Counter: ~4 characters per token for prose, ~3.5 for dense code. For per-call cost on top of utilization, see the LLM Cost Calculator. Background reading: context window, LLM, and GPT token.

Paste your prompt + context

Characters: 364
Words: 60
Tokens (est.): 91
Style: Prose

Context window utilization

Sort

GPT-4oOpenAI
91 / 128,000 tokens (0.07%)
Max output: 16,384 tokens
GPT-4o miniOpenAI
91 / 128,000 tokens (0.07%)
Max output: 16,384 tokens
GPT-4 TurboOpenAI
91 / 128,000 tokens (0.07%)
Max output: 4,096 tokens
o1-miniOpenAI
91 / 128,000 tokens (0.07%)
Max output: 65,536 tokens
Llama 3.3 70BMeta
91 / 128,000 tokens (0.07%)
Max output: 4,096 tokens
Limits vary by host (Together, Groq, Fireworks).
Llama 3.1 405BMeta
91 / 128,000 tokens (0.07%)
Max output: 4,096 tokens
DeepSeek V3DeepSeek
91 / 128,000 tokens (0.07%)
Max output: 8,192 tokens
Mistral Large 2Mistral
91 / 128,000 tokens (0.07%)
Max output: 8,192 tokens
o1OpenAI
91 / 200,000 tokens (0.05%)
Max output: 100,000 tokens
Includes reasoning tokens in output budget.
Claude Opus 4Anthropic
91 / 200,000 tokens (0.05%)
Max output: 32,000 tokens
Claude 3.5 SonnetAnthropic
91 / 200,000 tokens (0.05%)
Max output: 8,192 tokens
Claude 3.5 HaikuAnthropic
91 / 200,000 tokens (0.05%)
Max output: 8,192 tokens
Claude Sonnet 4Anthropic
91 / 1,000,000 tokens (0.01%)
Max output: 64,000 tokens
1M context tier — beta header required on some endpoints.
Gemini 2.0 FlashGoogle
91 / 1,000,000 tokens (0.01%)
Max output: 8,192 tokens
Gemini 1.5 FlashGoogle
91 / 1,000,000 tokens (0.01%)
Max output: 8,192 tokens
Gemini 1.5 ProGoogle
91 / 2,000,000 tokens (0.00%)
Max output: 8,192 tokens

What does this fit in?

Based on your 91-token input. Models with headroom below 100% accept the prompt; output budget still depends on each model’s output cap.

Accepts (16)

GPT-4o+127,909
GPT-4o mini+127,909
GPT-4 Turbo+127,909
o1-mini+127,909
Llama 3.3 70B+127,909
Llama 3.1 405B+127,909
DeepSeek V3+127,909
Mistral Large 2+127,909
o1+199,909
Claude Opus 4+199,909
Claude 3.5 Sonnet+199,909
Claude 3.5 Haiku+199,909
Claude Sonnet 4+999,909
Gemini 2.0 Flash+999,909
Gemini 1.5 Flash+999,909
Gemini 1.5 Pro+1,999,909

Overflows (0)

—

Token counts are heuristic (~4 chars/token for prose, ~3.5 for code). System prompts, tool definitions, and prior turns all consume the same context budget — paste the full assembled prompt for the most accurate utilization.

How to use

Paste the full prompt
Include the system prompt, every prior conversation turn, any tool definitions, and the retrieved context — all of it counts against the window, not just the latest user message.
Read the bars
Each model gets a bar showing input tokens as a share of its context window. Green under 50% is comfortable, amber 50–80% means you should start trimming, red above 80% leaves no room for the model's reply.
Sort to compare
Switch sort order to surface either the tightest fit (smallest windows first) or the most headroom (largest first). The 'what does this fit in?' panel summarizes which models accept the prompt and which overflow.
Check the output cap separately
Context window is the input budget; the output cap (shown under each bar) is a separate limit on response length. A model with a 1M-token window can still only emit ~8K tokens per reply.

Models tracked

16 production LLMs across six providers. Numbers reflect published API limits for direct (non-batch) access on the broadly available production tier.

Model	Provider	Context	Max output
GPT-4o	OpenAI	128,000	16,384
GPT-4o mini	OpenAI	128,000	16,384
GPT-4 Turbo	OpenAI	128,000	4,096
o1	OpenAI	200,000	100,000
o1-mini	OpenAI	128,000	65,536
Claude Opus 4	Anthropic	200,000	32,000
Claude Sonnet 4	Anthropic	1,000,000	64,000
Claude 3.5 Sonnet	Anthropic	200,000	8,192
Claude 3.5 Haiku	Anthropic	200,000	8,192
Gemini 2.0 Flash	Google	1,000,000	8,192
Gemini 1.5 Pro	Google	2,000,000	8,192
Gemini 1.5 Flash	Google	1,000,000	8,192
Llama 3.3 70B	Meta	128,000	4,096
Llama 3.1 405B	Meta	128,000	4,096
DeepSeek V3	DeepSeek	128,000	8,192
Mistral Large 2	Mistral	128,000	8,192

Frequently asked questions

What counts as context vs output?: Context is the input budget — everything the model reads before generating: system prompt, prior conversation turns, tool/function definitions, retrieved documents, and the latest user message. Output is what the model writes back. They're billed and limited separately, even though both consume the same underlying attention budget. A 200K-token context window with an 8K output cap means you can feed ~200K tokens in but only get up to ~8K tokens back per call.
Do system prompts count toward the window?: Yes. Every token the model sees consumes the input budget — system prompt, developer instructions, few-shot examples, tool schemas, prior assistant turns, retrieved RAG chunks. The only tokens that don't count are the ones the model emits as output (those count toward the separate output cap). Long system prompts on multi-turn conversations are the #1 cause of mysterious context-overflow errors.
What is RAG and how does it interact with context size?: Retrieval-Augmented Generation (RAG) means fetching relevant passages from a vector store at query time and pasting them into the prompt so the model can ground its answer in fresh or domain-specific facts. Each retrieved chunk eats context budget. A typical RAG setup retrieves 5–20 chunks of 500–1000 tokens each, so 5K–20K tokens of context vanish before the user message is even appended. Plan for that — large windows let you retrieve more aggressively, but precision matters more than volume.
Are reasoning tokens included in the output budget?: For OpenAI's o-series and similar reasoning models, yes — internal chain-of-thought tokens count against the output cap, which is why those caps are dramatically larger (e.g. 100K for o1) than non-reasoning models. The user only sees the final answer, but the model may have generated 10–50× that many hidden tokens to get there.
Why does the same model show different context limits at different vendors?: Open-weight models (Llama, DeepSeek, Mistral) are hosted by many providers (Together, Groq, Fireworks, Bedrock, Vertex), and each picks its own deployment configuration — KV-cache sizing, max sequence length, batching strategy. The headline 128K Llama 3.3 70B context might be 32K on one host and 128K on another. Check your specific provider's docs.
How accurate is the token estimate?: Within roughly ±10% for English prose and code. Real tokenization needs each model's exact BPE or SentencePiece table — Convertitive doesn't ship those in the browser because they'd add tens of megabytes of JavaScript. For final cost or limit sign-off, run your prompt through the vendor's own tokenizer (OpenAI tiktoken, Anthropic's count-tokens endpoint, Google's count_tokens).
Should I always pick the largest window?: No. Bigger windows let you cram more in, but model quality typically degrades past 30–50% utilization — attention becomes diffuse and the model misses details in the middle of long contexts (the 'lost in the middle' effect). For retrieval-heavy workloads, fewer high-quality chunks beat dumping the whole corpus. For analysis tasks, summarize-then-process often beats one giant prompt.

About

Why visualize rather than just list numbers

A bar chart makes the trade-off instant: at a glance you can see that a 40K-token prompt eats 31% of GPT-4o, 4% of Gemini 1.5 Pro, and overflows nothing. The bar's color encoding (green / amber / red) flags trouble before you ship — the same numbers in a table require mental arithmetic. The 'what fits' panel converts the bars into a routing decision: which models can accept this call.

How vendors count tokens

OpenAI uses cl100k or o200k BPE depending on the model. Anthropic Claude uses a custom tokenizer broadly similar in compression rate to cl100k. Google Gemini uses SentencePiece. Meta Llama uses its own tiktoken-derived BPE. All four land within ~10% of the 4-char/token rule for English; non-Latin scripts and code diverge more. The estimate here is good enough for sizing — but cross-check against the vendor tokenizer before you sign a contract that depends on a specific token count.

Long-context isn't free

Vendors charge per token, so a 1M-token call to Gemini 1.5 Pro costs roughly 10× a 100K-token call — and latency grows roughly with sequence length squared during prefill. Long-context tiers are a capability, not a default mode. Use them when you need them; otherwise compress, summarize, or shard.

Sources & references

Authoritative references behind the math, constants, and tables on this page. Verified by Buğra Sözeri on the dates shown and re-checked at every deploy.

OpenAI — Models reference — Per-model context window and max-output limits for GPT-4o, GPT-4 Turbo, and the o-series(as of 2026-05-30)
Anthropic — Models overview — Context window and output token limits for Claude Opus, Sonnet, and Haiku families(as of 2026-05-30)
Google — Gemini models — Input token limits (1M / 2M tiers) and per-response output caps for Gemini 1.5 and 2.0(as of 2026-05-30)
Meta — Llama model cards — Reference context length for Llama 3.1 / 3.3 deployments (host-dependent in practice)(as of 2026-05-30)
DeepSeek — API documentation — Context window and output limits for DeepSeek V3(as of 2026-05-30)
Mistral — Models documentation — Context window for Mistral Large 2 and adjacent production models(as of 2026-05-30)
Liu N et al. — Lost in the Middle (2023) — Empirical evidence that long-context models lose recall for information in the middle of the window(as of 2026-05-30)