Generative AI Model Ranking Matrix
Living benchmark · refreshed nightly
Frontier language, video, and image generation models compared side by side: benchmarks, capabilities, release dates, and price.
Updated
Frontier LLMs: coding & reasoning
License Caps / filter 1–3 tabs
No models match.
| Model | Vendor | Params (B) | License | Capabilities | Released | Context | SWE-Bench Verified | SWE-Bench Pro | Terminal-Bench | Reasoning* | $/M in | Best for |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Mistral Medium 3.5 | Mistral | 128 | open weights | TR | 2026-04-30 | 128k | 77.6 | N/A | N/A | 80 | 0.6 | 128B · SWE-V 77.6 beats Devstral 2 + Qwen3.5; τ³-Telecom 91.4 |
| Kimi K2.6 | Moonshot | 1000 | open | VTR | 2026-04-29 | 256k | 80.2 | 58.6 | 66.7 | 94 | 0.6 | 1.1T MoE · ties GPT-5.5 on SWE-Pro at $0.6/M; HLE 54 leads all |
| MiMo-V2.5-Pro | Xiaomi | 1023.2 | open | TR | 2026-04-29 | 1M | N/A | 57.2 | N/A | 89 | N/A | 1.02T MoE / 42B active · matches Opus 4.6 on SWE-Pro w/ 40% fewer tokens |
| Gemma 4 31B-it | 31 | open | VT | 2026-04-29 | 256k | 52 | N/A | N/A | 87 | N/A | 31B dense · AIME 89.2, GPQA 84.3, τ²-Retail 86.4 (12× jump from G3) | |
| Nemotron-3-Nano-Omni | NVIDIA | 30 | open | VATR | 2026-04-29 | 128k | N/A | N/A | N/A | N/A | N/A | Fresh (~17h) · 30B-A3B any-to-any reasoning |
| DeepSeek V4-Pro | DeepSeek | 861.6 | open | TR | 2026-04-27 | 128k | 80.6 | 55.4 | 67.9 | 92 | 0.9 | 1.6T MoE / 49B active · matches Opus 4.6 at fraction of cost |
| DeepSeek V4-Flash | DeepSeek | 158.1 | open | TR | 2026-04-27 | 128k | N/A | N/A | N/A | N/A | 0.3 | Fresh · 158B fast variant of V4 |
| GPT-5.5 | OpenAI | N/A | closed | VATRC | 2026-04-24 | 1M | 87.6 | 58.6 | 82.7 | 95 | 5 | SOTA Terminal-Bench 82.7; GDPval 84.9; OSWorld 78.7; ties Opus 4.7 on SWE-V |
| Qwen3.5-397B-A17B | Alibaba | 397 | open | VTR | 2026-04-24 | 256k | 80 | N/A | 54 | 92 | N/A | 403B MoE / 17B active · GPQA 88.4, MMLU-Pro 87.8, SWE-V 80.0 |
| Qwen3.6-35B-A3B | Alibaba | 35 | open | VTR | 2026-04-24 | 256k | 73.4 | 49.5 | N/A | 84 | N/A | 36B / 3B active · 73.4 SWE-V on tiny active params |
| Tencent Hy3-preview | Tencent | 298.8 | open (preview) | VTR | 2026-04-24 | 128k | 74.4 | N/A | 54.4 | 88 | N/A | 295B / 21B active · +40pts SWE-V vs Hy2; topped Tsinghua math PhD exam |
| Qwen3.6-Max-Preview | Alibaba | N/A | closed | VTR | 2026-04-20 | 1M | 79.5 | N/A | 77.1 | 89 | 6 | Agent loops: preserve_thinking across tool calls |
| MiniMax M2.7 | MiniMax | 228.7 | open | TR | 2026-04-20 | 205k | 78 | 56.22 | 57 | 87 | 0.3 | 229B · self-evolving; matches Codex on SWE-Pro |
| Claude Opus 4.7 | Anthropic | N/A | closed | VTRC | 2026-04-16 | 1M | 87.6 | 64.3 | 69.4 | 95 | 5 | Leads SWE-Bench Verified (87.6) and Pro (64.3); new tokenizer (~35% more tokens) |
| GLM-5.1 | Zhipu / Z.ai | 753.9 | open (MIT) | VTR | 2026-04-07 | 128k | 79 | 58.4 | N/A | 88 | 0.11 | 754B · #1 SWE-Bench Pro among open weights |
| Grok 4.20 Beta 2 | xAI | N/A | closed | VT | 2026-03-03 | 256k | 75 | N/A | N/A | 80 | 2 | 4-agent backbone · IFBench #1 (83); 8% of Opus cost |
| GPT-5.3 Codex | OpenAI | N/A | closed | VTRC | 2026-03 | 400k | 80 | 56.8 | 77.3 | 91 | 10 | Full SDLC agent: debug, terminal, PRDs, tests |
| Gemini 3.1 Pro | N/A | closed | VATRC | 2026-03 | 2M | 78 | N/A | N/A | 94 | 7 | Adjustable Deep Think; agentic browsing (BrowseComp 85.9) | |
| Devstral-2-123B | Mistral | 123 | open weights | TC | 2026-02-25 | 128k | N/A | N/A | N/A | N/A | 0.6 | 125B · code-specialized variant |
| Claude Opus 4.6 | Anthropic | N/A | closed | VTRC | 2026-02 | 1M | 80.8 | 57.3 | N/A | 92 | 15 | Large-codebase reasoning, multi-file refactors |
| Mistral Large 3 | Mistral | 675 | open (Apache 2.0) | VT | 2025-12-02 | 256k | N/A | N/A | N/A | 50 | 2 | 675B MoE / 41B active · non-reasoning (AIME ~40, GPQA ~44) · Apache 2.0 |
| Claude Sonnet 4.6 | Anthropic | N/A | closed | VTRC | 2025-12 | 1M | 79.6 | N/A | N/A | 86 | 3 | Default coding driver. 98% of Opus at ⅕ cost |
| Claude Haiku 4.5 | Anthropic | N/A | closed | VTRC | 2025-10-15 | 200k | 73.3 | N/A | N/A | 78 | 1 | 4–5× faster than Sonnet 4.5; cheap multi-agent driver |
| Kimi-K3 live | Moonshot | N/A | open | VTR | 2026-07-16 | 1M | N/A | N/A | N/A | N/A | 3.00 | Kimi K3 is a 2.8T parameter frontier-class multimodal reasoning model with 1M context, representing major scale and capability relevant to the ranking matrix. |
| GLM-5.2 live | Zhipu / Z.ai | 753.3 | open | R | 2026-07-02 | 1M | N/A | N/A | N/A | N/A | N/A | GLM-5.2 is a 753B-parameter frontier MoE model with a solid 1M-token context, novel IndexShare sparse attention architecture, and strong coding capabilities, representing a meaningful advance in long-horizon LLM capability. |
| claude-opus-5 live | Anthropic | N/A | closed | VTR | 2026-07-24 | 1M | N/A | N/A | N/A | N/A | 5.00 | Anthropic's flagship frontier model with 1M context, targeting demanding reasoning, coding, and long-horizon agentic tasks at premium pricing indicative of top-tier capability. |
| claude-opus-5:batch live | Anthropic | N/A | closed | VTR | 2026-07-24 | 1M | N/A | N/A | N/A | N/A | 2.50 | Anthropic's flagship frontier model with 1M context, strong reasoning and agentic capabilities, positioned as a top-tier general-purpose system. |
| LongCat-2.0 live | meituan | 48 | open | TR | 2026-07-20 | 1M | N/A | N/A | N/A | N/A | 0.30 | 1.6T parameter sparse MoE with ~1M token context targeting agentic/long-horizon tasks represents frontier-class scale and architectural ambition. |
| Inkling live | thinkingmachines | 41 | open | VATR | 2026-07-17 | 1M | N/A | N/A | N/A | N/A | 0.95 | 975B total parameter multimodal MoE with 1M context, general-purpose reasoning and tool-use from a new lab represents frontier-class significance and genuine architectural scale. |
| DeepSeek-V4-Flash-0731 live | DeepSeek | 304.2 | open | N/A | 2026-08-01 | N/A | N/A | N/A | N/A | N/A | N/A | DeepSeek-V4-Flash-0731 is a frontier-class 304B-parameter MoE model from a major lab with a dedicated technical report, representing a significant general-purpose release. |
| claude-opus-5-fast live | Anthropic | N/A | closed | VTR | 2026-07-24 | 1M | N/A | N/A | N/A | N/A | 10.00 | Claude Opus 5 fast-mode variant is a frontier-class general-purpose model from Anthropic with 1M context and top-tier pricing, representing a significant deployment at scale even if the fast-mode variant itself is not architecturally novel. |
| gemini-3.6-flash live | N/A | closed | VATR | 2026-07-21 | 1M | N/A | N/A | N/A | N/A | 1.50 | Gemini 3.6 Flash is a frontier-class general-purpose model from Google with a 1M+ token context window targeting coding and agentic workflows. | |
| Laguna-S-2.1 live | poolside | 117.6 | open | TR | 2026-07-27 | 1M | N/A | N/A | N/A | N/A | N/A | Laguna S 2.1 is a frontier-class 118B MoE coding-focused model with 1M context, native reasoning, and novel architectural features (softplus gating over 256 routed experts, mixed SWA/global attention) from a dedicated AI lab, making it significant and sufficiently novel for inclusion. |
| qwen3.8-max live trusted-lab | Alibaba | N/A | closed | VTR | 2026-08-03 | 1M | N/A | N/A | N/A | N/A | 2.00 | Qwen3.8 Max appears to be a flagship multimodal reasoning model from Alibaba with a 1M context window, but parameter count is unknown and it's a closed proprietary model with limited verifiable details, placing it just below the include threshold on scale. |
| gemini-3.6-flash:batch live trusted-lab | N/A | closed | VATR | 2026-07-21 | 1M | N/A | N/A | N/A | N/A | 0.75 | Gemini 3.6 Flash is a frontier-class general-purpose model from Google with strong capabilities but served as a batch variant with unknown parameter count, preventing full scale assessment. | |
| Ling-3.0-flash live trusted-lab | InclusionAI | 127.5 | open | R | 2026-08-06 | N/A | N/A | N/A | N/A | N/A | N/A | Novel hybrid linear attention architecture (KDA+MLA) with 124B total/5.1B active parameters represents genuine architectural innovation, but the moderate scale and limited adoption place it in review territory. |
| Ling-3.0-flash-fp8 live trusted-lab | InclusionAI | 127.5 | open | R | 2026-08-06 | N/A | N/A | N/A | N/A | N/A | N/A | Ling-3.0-flash introduces a genuinely novel native hybrid linear attention architecture (KDA+MLA) with sparse MoE at 124B total/5.1B active parameters, representing real architectural innovation, but its niche efficiency-focused positioning and moderate scale keep it from clear inclusion. |
| gemini-3.5-flash-lite:batch live trusted-lab | N/A | closed | VATR | 2026-07-21 | 1M | N/A | N/A | N/A | N/A | 0.15 | Gemini 3.5 Flash Lite is a frontier-class model from Google with notable agentic capabilities and a massive context window, but its 'Lite' positioning and unknown parameter count prevent a confident include without more details. | |
| gpt-5.6-luna:batch live trusted-lab | OpenAI | N/A | closed | VTR | 2026-07-09 | 1.1M | N/A | N/A | N/A | N/A | 0.10 | GPT-5.6 Luna is a frontier-class OpenAI model with a 1M+ context window and capable reasoning, but unknown parameter count and batch-only serving limit full assessment. |
| gpt-5.6-terra-pro:batch live trusted-lab | OpenAI | N/A | closed | VTR | 2026-07-09 | 1.1M | N/A | N/A | N/A | N/A | 1.00 | GPT-5.6 Terra Pro appears to be a frontier-class OpenAI model with extended context and reasoning capabilities, but it is a batch/reasoning-mode variant of an existing model rather than a novel architecture, and parameter count is unknown. |
| gpt-5.6-terra live trusted-lab | OpenAI | N/A | closed | VTR | 2026-07-09 | 1.1M | N/A | N/A | N/A | N/A | 1.00 | GPT-5.6 Terra appears to be a frontier-class OpenAI model with a very large context window (1M+ tokens), but it is an intermediate tier within a series rather than a flagship, and key details like parameter count are unknown, preventing confident composite scoring. |
| gpt-5.6-terra:batch live trusted-lab | OpenAI | N/A | closed | VTR | 2026-07-09 | 1.1M | N/A | N/A | N/A | N/A | 1.00 | GPT-5.6 Terra appears to be a frontier-class model from OpenAI with a very large context window, but it is an intermediate tier within a series rather than a flagship, and parameter count is unknown, preventing confident scale scoring. |
| gpt-5.6-sol-pro:batch live trusted-lab | OpenAI | N/A | closed | VTR | 2026-07-09 | 1.1M | N/A | N/A | N/A | N/A | 2.50 | GPT-5.6 Sol Pro appears to be a frontier-class OpenAI reasoning model with a 1M+ context window, but it is a serving variant (batch + pro reasoning mode) of an existing model rather than a novel architecture, keeping novelty moderate and composite just below the include threshold. |
| gemini-3.5-flash-lite live trusted-lab | N/A | closed | VATR | 2026-07-21 | 1M | N/A | N/A | N/A | N/A | 0.30 | Gemini 3.5 Flash Lite is a notable Google frontier model with a massive 1M token context and agentic capabilities, but as a 'lite' efficiency-focused variant it falls short of full frontier significance, and parameter count is unknown preventing confident scale scoring. | |
| gpt-5.6-sol-pro live trusted-lab | OpenAI | N/A | closed | VTR | 2026-07-09 | 1.1M | N/A | N/A | N/A | N/A | 5.00 | GPT-5.6 Sol Pro appears to be a frontier-class OpenAI reasoning model with a 1M+ context window, but it is essentially a configuration variant (reasoning.mode=pro) of GPT-5.6 Sol rather than a distinct model, limiting novelty. |
| deepseek-v4-flash-latest live trusted-lab | DeepSeek | N/A | closed | TR | 2026-08-01 | 1M | N/A | N/A | N/A | N/A | 0.08 | DeepSeek V4 Flash appears to be a frontier-adjacent model with a massive 1M token context, but params, architecture details, and capabilities are not stated, and the card provides no substantive information beyond version routing. |
| gpt-5.6-luna live trusted-lab | OpenAI | N/A | closed | VTR | 2026-07-09 | 1.1M | N/A | N/A | N/A | N/A | 0.10 | GPT-5.6 Luna is a production OpenAI model with a very large context window and agentic capabilities, but parameter count is unknown and it appears to be a smaller/faster variant rather than a frontier-class general-purpose model. |
Capabilities key: V Vision (image input) A Audio T Tools / function calling R Extended reasoning C Computer / browser use
top tier mid lower *Reasoning is a composite score (GPQA Diamond / HLE / AIME / ARC-AGI-2, normalized 0–100, directional).
Video generation models
License I2V rank (1 = best)
No models match.
| Model | Vendor | License | Released | I2V Rank | Max len (s) | Resolution | Audio | $/gen | Best for |
|---|---|---|---|---|---|---|---|---|---|
| Kling 3.0 | Kuaishou | closed | 2026-02 | 1 | 15 (multi-shot) | 1080p | synced dialogue + SFX | 1.00 | Cinematic shots; #1 general-purpose |
| Veo 3.1 | closed | 2026-01 | 2 | 8 | 1080p | yes | 0.75 | High-fidelity realism, prompt adherence | |
| Sora 2 | OpenAI | closed | 2025-10 | 3 | 12 | 1080p | yes | 0.90 | Imaginative T2V; ChatGPT-integrated |
| Seedance 2.0 | ByteDance | closed | 2026-04 | 4 | 15 | 1080p | yes | 0.50 | Product ads, e-comm, character consistency |
| LTX-2 | Lightricks | open | 2025-Q4 | 5 | 10 | 4K@50fps | native sync | 0.20 | Open-weights leader; 4K + audio |
| Wan 2.2 | Alibaba | open | 2025-Q3 | 6 | 6 | 720p | N/A | 0.05 | Runs on a 4070; novel MoE denoiser |
Image generation models
License Quality = directional leaderboard score
No models match.
| Model | Vendor | License | Released | Max res | Quality* | Speed (s) | $/img | Best for |
|---|---|---|---|---|---|---|---|---|
| Midjourney v8 | Midjourney | closed | 2026-03 | 2K | 95 | 10 | 0.04 | Aesthetic king · rewritten engine, 5× faster than v7 |
| FLUX.2 [pro] | Black Forest Labs | closed (API) | 2026-Q1 | 2K | 93 | 4.5 | 0.04 | Photoreal commercial: best skin, lighting, materials |
| FLUX.2 [dev] | Black Forest Labs | open weights | 2026-Q1 | 2K | 89 | 6 | 0 | Top open-weights photoreal model |
| GPT Image 2 | OpenAI | closed | 2026-Q1 | 1K+ | 92 | 6 | 0.04 | Best prompt adherence: complex composed scenes |
| Imagen 4 Ultra | closed | 2025-Q4 | 2K | 94 | 5 | 0.04 | Photoreal flagship; strong text rendering | |
| Imagen 4 Fast | closed | 2026-03 | 1K | 86 | 2 | 0.02 | Cheapest fast quality at $0.02/img | |
| Nano Banana 2 | Google (Gemini 3.1 Flash Image) | closed | 2026-03 | 1K | 82 | 1.5 | 0.02 | Fastest end-to-end (~1–3s) · in-chat editing |
| Recraft V4 | Recraft | closed | 2025-Q4 | 2K | 88 | 5 | 0.04 | Brand / design assets, vector-style outputs |
| Ideogram 3 | Ideogram | closed | 2025-Q4 | 1K | 84 | 5 | 0.03 | Typography & in-image text rendering |
| Seedream 4.5 | ByteDance | closed | 2026-Q1 | 2K | 90 | 5 | 0.03 | Asian aesthetics, character consistency |
| Stable Diffusion 4 | Stability AI | open | 2026-Q1 | 2K | 80 | 6 | 0 | Open ecosystem · ControlNet / LoRA backbone |
| Hunyuan Image 3 | Tencent | open | 2026-Q1 | 2K | 83 | 5 | 0 | Open Tencent flagship; bilingual prompts |