Key takeaways (as of May 2026)
- Largest context windows: GPT-4.1 and Gemini 2.0 Flash / 2.5 Pro — all at 1M tokens.[1][5]
- Most open: DeepSeek-V3 (MIT, 671B MoE) and Llama 3.3 70B (community license, 70.6B) — both weights downloadable.[10][7]
- Multimodal leaders: GPT-4o (text/image/audio in+out) and Gemini 2.0 Flash (text/image/audio/video in, text/image out).[1][5]
- Parameters undisclosed: OpenAI (GPT-4 series) and Anthropic (Claude 3 series) have not published parameter counts. Cells read “Undisclosed” — not estimated figures.
Comparison table
All data as of May 2026.[*] “Undisclosed” means the vendor has not published the value — not that it is unknown to us. Context window definitions vary; see methodology for details.
| Model | Vendor | Parameters | Context window | Input modalities | Output modalities | License | Knowledge cutoff |
|---|---|---|---|---|---|---|---|
| GPT-4o | OpenAI | Undisclosed[1] | 128K | Text, Image, Audio | Text, Audio | Proprietary | Oct 2023 |
| GPT-4.1 | OpenAI | Undisclosed[2] | 1M | Text, Image | Text | Proprietary | Undisclosed |
| Claude 3.5 Sonnet | Anthropic | Undisclosed[3] | 200K | Text, Image | Text | Proprietary | Apr 2024 |
| Claude 3.7 Sonnet | Anthropic | Undisclosed[4] | 200K | Text, Image | Text | Proprietary | Undisclosed |
| Gemini 2.0 Flash | Google DeepMind | Undisclosed[5] | 1M | Text, Image, Audio, Video | Text, Image | Proprietary | Aug 2024 |
| Gemini 2.5 Pro | Google DeepMind | Undisclosed[6] | 1M | Text, Image, Audio, Video | Text | Proprietary | Undisclosed |
| Llama 3.3 70B | Meta AI | 70.6B[7] | 128K | Text | Text | Llama 3 Community License (open weights) | Dec 2023 |
| Mistral Large 2 | Mistral AI | Undisclosed[8] | 128K | Text | Text | Proprietary | Undisclosed |
| Grok-3 | xAI | Undisclosed[9] | 131K | Text, Image | Text | Proprietary | Undisclosed |
| DeepSeek-V3 | DeepSeek | 671B total / 37B active (MoE)[10] | 128K | Text | Text | MIT (open weights) | Undisclosed |
| Qwen2.5-72B | Alibaba Cloud | 72.7B[11] | 128K | Text | Text | Qwen License (Apache 2.0 base with usage restrictions) | Undisclosed |
| Command R+ | Cohere | Undisclosed[12] | 128K | Text | Text | Proprietary | Undisclosed |
How we compile this table
Our full methodology covers column definitions, source tiers, the “Undisclosed” policy, recency policy, COI disclosure, and our correction process. Column structure intentionally mirrors Wikipedia's LLM comparison article for editor cross-reference.
Read the methodologyFrequently Asked Questions: LLM Comparisons
- What is the best large language model in 2026?
- The best LLM depends on your use case. Claude claude-sonnet-5 and GPT-4.1 lead on instruction following, long-context coding, and complex reasoning. Gemini 2.5 Pro excels at multimodal tasks and document analysis. Llama 4 Scout and DeepSeek-V3 are the top open-weight alternatives for self-hosted deployments. This comparison table ranks each model by context window, input modality, and license type so you can match the right model to your workload.
- What does context window mean for LLMs?
- The context window is the maximum number of tokens (roughly 0.75 words each) an LLM can process in a single request — both the input prompt and the generated output combined. Models with larger context windows (200K–1M tokens) can process entire books, long codebases, or lengthy conversation histories without losing earlier context. GPT-4.1 and Gemini 2.5 Pro offer the largest publicly documented context windows in this comparison at 1M tokens.
- What is the difference between proprietary and open-weight LLMs?
- Proprietary LLMs (GPT-4o, Claude, Gemini) are accessible only via API or product interface; their weights are not publicly released. Open-weight LLMs (Llama 4, Mistral, DeepSeek, Falcon) publish model weights for download, allowing self-hosting, fine-tuning, and offline deployment. Open-weight models are preferred for privacy-sensitive workloads, cost-sensitive high-volume use, and research contexts requiring full reproducibility.
- How are LLM benchmarks like MMLU and HumanEval used to compare models?
- MMLU (Massive Multitask Language Understanding) measures knowledge breadth across 57 academic subjects — a proxy for general knowledge quality. HumanEval measures code generation accuracy on 164 Python programming problems. GPQA tests graduate-level science reasoning. These benchmarks allow apples-to-apples comparison across labs, but real-world performance often diverges from benchmark scores, so this table pairs them with modality, context, and license data for a fuller picture.
- How often is this LLM comparison table updated?
- We update this table when major model releases or significant specification changes occur — typically monthly during active periods in the AI landscape. The page header and JSON-LD dateModified field show the date of the most recent data revision. Each row links to the primary vendor documentation source so you can verify specifications independently.
Sources
- GPT-4o (OpenAI). GPT-4o system card (opens in new tab); OpenAI models API reference (opens in new tab) Accessed May 2026.
- GPT-4.1 (OpenAI). OpenAI models API reference (opens in new tab) Accessed May 2026.
- Claude 3.5 Sonnet (Anthropic). Anthropic models list (opens in new tab) Accessed May 2026.
- Claude 3.7 Sonnet (Anthropic). Anthropic Claude 3.7 Sonnet announcement (opens in new tab); Anthropic models list (opens in new tab) Accessed May 2026.
- Gemini 2.0 Flash (Google DeepMind). Google AI Gemini model page (opens in new tab) Accessed May 2026.
- Gemini 2.5 Pro (Google DeepMind). Google AI Gemini 2.5 Pro model page (opens in new tab) Accessed May 2026.
- Llama 3.3 70B (Meta AI). Meta Llama 3.3 announcement (opens in new tab); Llama 3.3 70B HuggingFace model card (opens in new tab); Llama 3 technical report (arXiv 2407.21783) (opens in new tab) Accessed May 2026.
- Mistral Large 2 (Mistral AI). Mistral AI models (opens in new tab) Accessed May 2026.
- Grok-3 (xAI). xAI Grok-3 announcement (opens in new tab) Accessed May 2026.
- DeepSeek-V3 (DeepSeek). DeepSeek-V3 technical report (arXiv 2412.19437) (opens in new tab); DeepSeek-V3 GitHub (opens in new tab) Accessed May 2026.
- Qwen2.5-72B (Alibaba Cloud). Qwen2.5 technical report (arXiv 2412.15115) (opens in new tab); Qwen2.5-72B HuggingFace model card (opens in new tab) Accessed May 2026.
- Command R+ (Cohere). Cohere Command R+ model page (opens in new tab) Accessed May 2026.
- * Data reflects information available from primary vendor sources as of May 2026. This is a fast-moving field; check the dateModified stamp above and consult vendor documentation for the latest specifications.