Which LLM has the largest context window?

As of July 2026, Meta's open-weights Llama 5 leads with a 5 million token context window. Among proprietary models, GPT-5.5, GPT-5.4, Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5, and the Gemini 3 family all offer roughly 1 million tokens. Note that some providers bill long prompts at premium rates — GPT-5.x doubles input pricing past 272K tokens, and Gemini 3.1 Pro steps up past 200K.

What does input vs output pricing mean?

LLM APIs charge separately for prompt tokens (input) and generated tokens (output). Output typically costs 3-5x more than input because generation is computationally heavier. Pricing is quoted per 1 million tokens.

What is the difference between open-source and closed-source LLMs?

Open-weights LLMs (Llama 5, Mistral Large 3, DeepSeek V4) publish their weights publicly — you can run them locally or fine-tune them without per-token API costs. Licenses vary: DeepSeek V4 is MIT, Mistral Large 3 is Apache 2.0, and Llama 5 uses Meta's community license. Closed-source models (GPT-5.x, Claude, Gemini) are accessed only through paid APIs.

What are modalities in the context of LLMs?

Modalities refer to the types of input a model supports: text (all models), image input (GPT-5.x, Claude, Gemini, Llama 5), audio (GPT-5.5, Gemini 3), and video (Gemini 3 family only among the compared models). Multimodal models process more than one type; DeepSeek V4 remains text-and-code only.

LLM Model Comparison

Compare models side by side on context length, pricing, modality and benchmark results.

LLM Model Comparison

13 models · Updated 2025 · Prices per 1M tokens

13 models

Model	Context	Input $/1M	Output $/1M	Released	Modalities	License	Strengths
Claude Fable 5 Anthropic · Claude 5	1M	$10.00	$50.00	2026	textimagecode	Prop.	Deepest reasoningLong-horizon agentsWriting
Claude Haiku 4.5 Anthropic · Claude 4	200K	$1.00	$5.00	Oct 2025	textimagecode	Prop.	Fastest ClaudeLow costCoding
Claude Opus 4.8 Anthropic · Claude 4	1M	$5.00	$25.00	2026	textimagecode	Prop.	Long-horizon agentsCodingKnowledge work
Claude Sonnet 5 Anthropic · Claude 5	1M	$3.00	$15.00	2026	textimagecode	Prop.	CodingAgentic work1M context
DeepSeek V4 Flash DeepSeek · DeepSeek V4	1M	$0.14	$0.28	Apr 2026	textcode	Open	284B MoE (13B active)High volumeSpeed
DeepSeek V4 Pro DeepSeek · DeepSeek V4	1M	$0.44	$0.87	Apr 2026	textcode	Open	1.6T MoE (49B active)Code & mathUltra-low cost
GPT-5.4 OpenAI · GPT-5	1M	$2.50	$15.00	2026	textimagecode	Prop.	CodingTool searchStructured output
GPT-5.5 OpenAI · GPT-5	1M	$5.00	$30.00	2026	textimageaudiocode	Prop.	Flagship reasoningAgentic tool useMultimodal
Gemini 3.1 Pro Google · Gemini 3	1M	$2.00	$12.00	2026	textimageaudiovideocode	Prop.	Multimodal reasoningComputer useVideo input
Gemini 3.5 Flash Google · Gemini 3	1M	$1.50	$9.00	2026	textimageaudiovideocode	Prop.	Agentic codingSpeedLong context
Llama 5 Meta · Llama 5	5M	Free / Self-host	Free / Self-host	Apr 2026	textimagecode	Open	5M contextOpen weights600B params
Mistral Large 3 Mistral · Mistral	256K	Free / Self-host	Free / Self-host	Dec 2025	textimagecode	Open	Multilingual (200+ languages)675B MoE (41B active)Apache 2.0
o3 OpenAI · o-series	200K	$2.00	$8.00	Apr 2025	textimagecode	Prop.	Deep reasoningMathScience

Prices shown are standard API rates as of July 2026. Batch/cached rates may be lower; some providers charge premium rates above a long-context threshold. Open-source models shown as "Free / Self-host" — inference costs vary by provider.

Head-to-head comparisons

Detailed side-by-side pages with pricing math, verdicts, and use-case guidance.

GPT-5.5 vs Claude Opus 4.8 Claude Sonnet 5 vs GPT-5.4 Gemini 3.1 Pro vs GPT-5.5 Claude Opus 4.8 vs Gemini 3.1 Pro Claude Fable 5 vs GPT-5.5 DeepSeek V4 Pro vs GPT-5.5 DeepSeek V4 Pro vs Claude Sonnet 5 Gemini 3.5 Flash vs Claude Haiku 4.5 GPT-5.4 vs Gemini 3.5 Flash Llama 5 vs DeepSeek V4 Pro Mistral Large 3 vs Llama 5 o3 vs DeepSeek V4 Pro

Choosing a model, in a sensible order

Benchmark tables are the least useful input to this decision. A workable sequence:

Start with the constraints, not the leaderboard —Latency ceiling, data residency, whether you can send data to a third party at all, and budget per request. These eliminate most of the field before quality enters the picture.
Build an evaluation set from your own traffic —50–100 real examples with known-good outputs. This is the single highest-value thing on this list and the one most often skipped. Public benchmarks measure general capability; your eval measures whether the model does your job.
Start with a small model and escalate —Most production tasks — classification, extraction, routine generation — are handled well by small models at a fraction of the cost and latency. Reach for a frontier model when your eval shows the small one failing.
Test the failure modes, not the happy path —Ambiguous inputs, adversarial phrasing, missing fields, unusually long documents. Models differ far more in how they degrade than in how they perform on clean input.
Check structured-output reliability separately —If you need valid JSON or a specific schema, measure that as its own metric. Models that score similarly on reasoning can differ substantially in format adherence.

What benchmark scores do not tell you

Contamination. Benchmark questions and answers circulate online and end up in training data. A high score can reflect memorisation rather than capability, and there is rarely a way to tell from the outside.

Saturation. Once every model scores above 90%, the remaining differences are mostly noise and test-set errors. Several long-standing benchmarks are now in this state.

Prompt sensitivity. Published figures use tuned prompts, sometimes with many-shot examples or self-consistency sampling. Your single zero-shot call will not reproduce them.

Aggregate scores hide the distribution. A model that is excellent at code and weak at multilingual work can post the same average as one that is uniformly adequate. If your workload is narrow, the average tells you almost nothing.

Use benchmarks to build a shortlist. Use your own evaluation set to choose from it.

About

This tool provides a side-by-side comparison of 13 leading large language models (as of July 2026) including GPT-5.5, GPT-5.4, o3, Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5, Claude Haiku 4.5, Gemini 3.1 Pro, Gemini 3.5 Flash, DeepSeek V4 Pro, DeepSeek V4 Flash, Llama 5, and Mistral Large 3. Each model entry includes context window size, input/output pricing per million tokens, supported modalities (text/image/audio/video/code), knowledge cutoff date, open/closed source status, license, key strengths, and release date. Filter by provider, open-source status, or modality. Sort by any column. Dedicated head-to-head pages add verdicts and workload cost math. All data is embedded client-side — no API calls or server needed.

How to use

1 Browse the table to compare all 13 models at a glance.
2 Type in the search box to filter by model name, provider, or keyword.
3 Click a column header to sort by that property (click again to reverse).
4 Use the filter chips to show only open-source models or filter by modality.
5 Click any model row to expand its full details including strengths and license.

Which LLM has the largest context window?: As of July 2026, Meta's open-weights Llama 5 leads with a 5 million token context window. Among proprietary models, GPT-5.5, GPT-5.4, Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5, and the Gemini 3 family all offer roughly 1 million tokens. Note that some providers bill long prompts at premium rates — GPT-5.x doubles input pricing past 272K tokens, and Gemini 3.1 Pro steps up past 200K.
What does input vs output pricing mean?: LLM APIs charge separately for prompt tokens (input) and generated tokens (output). Output typically costs 3-5x more than input because generation is computationally heavier. Pricing is quoted per 1 million tokens.
What is the difference between open-source and closed-source LLMs?: Open-weights LLMs (Llama 5, Mistral Large 3, DeepSeek V4) publish their weights publicly — you can run them locally or fine-tune them without per-token API costs. Licenses vary: DeepSeek V4 is MIT, Mistral Large 3 is Apache 2.0, and Llama 5 uses Meta's community license. Closed-source models (GPT-5.x, Claude, Gemini) are accessed only through paid APIs.
What are modalities in the context of LLMs?: Modalities refer to the types of input a model supports: text (all models), image input (GPT-5.x, Claude, Gemini, Llama 5), audio (GPT-5.5, Gemini 3), and video (Gemini 3 family only among the compared models). Multimodal models process more than one type; DeepSeek V4 remains text-and-code only.

In-depth guides

More in AI Tools

Token Counter Count tokens for any text using the real tokenizers, with a per-token breakdown showing exactly where a passage splits.LLM Cost Calculator Estimate API spend from input and output token volumes across providers, and compare the same workload model by model.Context Window Planner Check whether a system prompt, retrieved documents, history and expected output all fit inside a model’s window.Prompt Builder Compose structured prompts from role, context, task and format sections, with a live token count as you write.Prompt Diff Compare two prompt versions token by token to see exactly what changed between runs.System Prompt Analyzer Analyse a system prompt for length, contradictory instructions, ambiguity and structural problems.

See all ai tools.