LLM Model Comparison
Flagship large language models side-by-side — context window, pricing per million tokens, MMLU-Pro / GPQA / SWE-Bench Verified, modalities, and open-weights status.
15 models · Last reviewed: · Pricing is published list rate; effective cost varies with caching and volume.
Related framework dimensions: Data & Infrastructure, Agentic AI.
Showing 15 of 15 models
| Modalities | Tags | |||||||
|---|---|---|---|---|---|---|---|---|
Claude Haiku 4.5 Anthropic · 2025-10 low-latency tool routing · high-volume inference | 200K | $0.80 cached $0.08 | $4 | 74 | 62 | 48 | text, image | |
Claude Opus 4.7 Anthropic · 2026-02 long-context reasoning · agentic tool use · coding | 1.0M | $15 cached $2 | $75 | 87 | 84 | 72 | text, image | reasoning |
Claude Sonnet 4.6 Anthropic · 2026-01 balanced cost/quality · coding · agent orchestration | 500K | $3 cached $0.30 | $15 | 82 | 78 | 65 | text, image | reasoning |
Command R+ Cohere · 2025-06 RAG optimization · enterprise privacy | 128K | $3 | $10 | 71 | — | — | text | |
DeepSeek-R1 DeepSeek · 2025-01 open weights reasoning · exceptional cost · math | 128K | $0.55 cached $0.14 | $2 | 84 | 71 | 49 | text | reasoningopen weights |
Gemini 2.5 Flash Google · 2026-02 ultra-low cost · long context · multimodal | 1.0M | $0.30 cached $0.07 | $3 | 76 | 65 | 50 | text, image, audio, video | |
Gemini 2.5 Pro Google · 2026-02 2M context · multimodal · grounded search | 2.0M | $4 cached $0.90 | $15 | 84 | 80 | 63 | text, image, audio, video | reasoning |
GPT-5 OpenAI · 2026-03 multimodal · reasoning · general purpose | 1.0M | $10 cached $1 | $30 | 86 | 82 | 70 | text, image, audio | reasoning |
GPT-5 mini OpenAI · 2026-03 balanced cost · multimodal | 400K | $2 cached $0.15 | $6 | 78 | 68 | 55 | text, image | |
Grok 3 xAI · 2025-11 real-time X data · reasoning mode | 1.0M | $3 | $15 | 80 | 74 | — | text, image | reasoning |
Llama 4 405B Meta · 2025-11 open weights · self-hostable · strong general purpose | 256K | — | — | 79 | 70 | 52 | text, image | open weights |
Llama 4 70B Meta · 2025-11 open weights · good cost/perf | 128K | — | — | 72 | 58 | 42 | text, image | open weights |
Mistral Large 3 Mistral · 2025-09 EU-hosted option · European data-residency | 128K | $2 | $6 | 76 | 62 | 45 | text | |
o3 OpenAI · 2025-12 deep reasoning · math · complex research | 200K | $60 | $240 | 89 | 88 | 71 | text | reasoning |
Qwen 3 235B Alibaba · 2025-08 open weights · strong multilingual (esp. Chinese) | 128K | — | — | 77 | 66 | 47 | text, image | open weights |
Prices are USD per million tokens, published list rates as of the last-reviewed date. Cached-input pricing where offered (Anthropic, OpenAI, Google). Benchmarks pulled from vendor-reported and independent sources; blank cells indicate no comparable published number.
FAQs
How often is the model comparison updated?
Quarterly, and immediately after major model releases. Last reviewed: 2026-08-27.
Which benchmarks are listed?
MMLU-Pro (general knowledge + reasoning), GPQA Diamond (graduate-level science), and SWE-Bench Verified (real-world software engineering). We deliberately avoid benchmarks that are either saturated or gameable.
Why does open-weights pricing show —?
Open-weights models (Llama, Qwen, DeepSeek partially) are self-hostable or run on many providers. Cost depends on the provider (Together, Groq, Fireworks) or your own inference setup, so there is no single list price.
What does cached-input price mean?
Anthropic, OpenAI, and Google offer prompt caching that drops the input token price by roughly 10× for reused prefixes. Cached-input price is what you pay for the cached tokens. Order of magnitude matters most for RAG and agent workloads.