Benchmarks reference

The evidence, filterable.

A dated snapshot of what providers, model cards, and third-party eval suites report. We link every claim to its source and measure nothing ourselves.

Open the planner

Scores reflect published claims checked on 2026-08-04. Suites differ in harness, prompting, and tooling, so only compare scores within a single suite.

Filters

Narrow the evidence set.

34 of 34 models

Provider

Weights

Size class

Input types

Output types

GPT-5.6 Luna
OpenAI
closed weights
Params
—
Size
large
Context
1M tokens
in · textin · imageout · text
- Artificial Analysis Intelligence Index (max)51
  Artificial Analysis: GPT-5.6 ↗Checked 2026-08-04
- ARC-AGI-2 (max)59.5%
  ARC Prize: GPT-5.6 results ↗Checked 2026-08-04
Fast/affordable tier of the GPT-5.6 family. Parameter count undisclosed.
GPT-5.6 Terra
OpenAI
closed weights
Params
—
Size
large
Context
1M tokens
in · textin · imageout · text
- Artificial Analysis Intelligence Index (max)55
  Artificial Analysis: GPT-5.6 Terra vs DeepSeek V4 Pro ↗Checked 2026-08-05
- Coding Agent Index (max)77
  Artificial Analysis: GPT-5.6 ↗Checked 2026-08-05
Mid GPT-5.6 hosted tier between Luna and Sol. Parameter count undisclosed.
GPT-5.6 Sol
OpenAI
closed weights
Params
—
Size
frontier
Context
1M tokens
in · textin · imageout · text
- Agents' Last Exam53.6
  OpenAI: GPT-5.6 Sol preview ↗Checked 2026-08-04
- Artificial Analysis Intelligence Index (max)59
  Artificial Analysis: GPT-5.6 ↗Checked 2026-08-04
- Coding Agent Index (max, Codex)80
  Artificial Analysis: GPT-5.6 ↗Checked 2026-08-04
Flagship GPT-5.6 model for long-horizon agentic work. Parameter count undisclosed.
GPT-5.4 mini
OpenAI
closed weights
Params
—
Size
medium
Context
400K tokens
in · textin · imageout · text
- Artificial Analysis Intelligence Index (xhigh)40
  Artificial Analysis: GPT-5.4 mini ↗Checked 2026-08-05
Compact GPT-5.4 generalist for coding and document work. Parameter count undisclosed.
GPT-5.4 nano
OpenAI
closed weights
Params
—
Size
small
Context
400K tokens
in · textin · imageout · text
- Artificial Analysis Intelligence Index (xhigh)38
  Artificial Analysis: GPT-5.4 nano vs Gemini 3.1 Pro ↗Checked 2026-08-05
Lowest-cost GPT-5.4 text tier. Parameter count undisclosed.
DeepSeek V4 Flash 0731
DeepSeek
open weights
Params
284B total · 13B active
Size
frontier
Context
1M tokens
in · textout · text
- Terminal-Bench 2.182.7
  DeepSeek V4 Flash 0731 model card ↗Checked 2026-08-04
- DeepSWE54.4
  DeepSeek V4 Flash 0731 model card ↗Checked 2026-08-04
- Agents' Last Exam25.2
  DeepSeek V4 Flash 0731 model card ↗Checked 2026-08-04
MIT open weights; MoE with ~13B active. Official post-training release superseding the preview.
DeepSeek V4 Pro
DeepSeek
open weights
Params
—
Size
frontier
Context
1M tokens
in · textout · text
- Artificial Analysis Intelligence Index (max)44
  Artificial Analysis: GPT-5.6 Terra vs DeepSeek V4 Pro ↗Checked 2026-08-05
Open-weight DeepSeek Pro tier for harder coding and research. Parameter count varies by reporting source.
Gemma 4 E4B
Google
open weights
Params
4B
Size
small
Context
128K tokens
in · textin · imagein · audioin · videoout · text
- GPQA Diamond58.6%
  Gemma 4 technical report ↗Checked 2026-08-04
- MMMLU76.6%
  Gemma 4 technical report ↗Checked 2026-08-04
Edge/mobile multimodal utility model from the Gemma 4 family.
Gemma 4 26B A4B
Google
open weights
Params
26B total · 3.8B active
Size
medium
Context
256K tokens
in · textin · imagein · videoout · text
- GPQA Diamond82.3%
  Gemma 4 technical report ↗Checked 2026-08-04
- LiveCodeBench v677.1%
  Gemma 4 model overview ↗Checked 2026-08-04
MoE local generalist: ~26B total / ~3.8B active. Fits 24–32 GB consumer GPUs when quantized.
Gemma 4 31B
Google
open weights
Params
31B
Size
large
Context
256K tokens
in · textin · imagein · videoout · text
- GPQA Diamond84.3%
  Gemma 4 model card ↗Checked 2026-08-04
- MMLU Pro85.2%
  Gemma 4 technical report ↗Checked 2026-08-04
- LiveCodeBench v680.0%
  Gemma 4 technical report ↗Checked 2026-08-04
Dense Gemma 4 flagship for higher-quality local multimodal work.
Gemini 3.1 Pro
Google
closed weights
Params
—
Size
frontier
Context
1M tokens
in · textin · imagein · audioin · videoout · text
- Artificial Analysis Intelligence Index46
  Artificial Analysis: GPT-5.4 mini vs Gemini 3.1 Pro ↗Checked 2026-08-05
Hosted Gemini Pro multimodal tier. Parameter count undisclosed.
Gemini 3.6 Flash
Google
closed weights
Params
—
Size
large
Context
1M tokens
in · textin · imagein · audioin · videoout · text
- Artificial Analysis Intelligence Index (high)50
  Artificial Analysis: Gemini 3.6 Flash vs Claude Sonnet 5 ↗Checked 2026-08-05
Fast multimodal Flash generalist on the current Gemini 3.6 line.
Gemini 3.5 Flash-Lite
Google
closed weights
Params
—
Size
medium
Context
1M tokens
in · textin · imagein · audioin · videoout · text
- Artificial Analysis Intelligence Index36
  Artificial Analysis: Gemini 3.5 Flash-Lite ↗Checked 2026-08-05
Low-cost multimodal Flash-Lite route for high-volume document and vision work.
Claude Sonnet 5
Anthropic
closed weights
Params
—
Size
large
Context
1M tokens
in · textin · imageout · text
- Artificial Analysis Intelligence Index (max)53
  Artificial Analysis: Anthropic models ↗Checked 2026-08-05
Premium Claude generalist. Parameter count undisclosed.
Claude Haiku 4.5
Anthropic
closed weights
Params
—
Size
medium
Context
200K tokens
in · textin · imageout · text
- Artificial Analysis Intelligence Index30
  Artificial Analysis: Anthropic models ↗Checked 2026-08-05
Economy Claude route for focused coding and document work.
Claude Fable 5
Anthropic
closed weights
Params
—
Size
frontier
Context
1M tokens
in · textin · imageout · text
- Artificial Analysis Intelligence Index60
  Artificial Analysis: Claude Fable 5 ↗Checked 2026-08-04
- Humanity's Last Exam53%
  Artificial Analysis: Claude Fable 5 ↗Checked 2026-08-04
Anthropic frontier model; parameter count undisclosed. Some evals fall back to Opus 4.8 under safety guardrails.
Claude Opus 4.8
Anthropic
closed weights
Params
—
Size
frontier
Context
1M tokens
in · textin · imageout · text
- Terminal-Bench 2.185.0
  DeepSeek V4 Flash 0731 comparison table ↗Checked 2026-08-04
- DeepSWE58.0
  DeepSeek V4 Flash 0731 comparison table ↗Checked 2026-08-04
Prior Anthropic flagship still used as a fallback and comparison baseline.
gpt-oss-20b
OpenAI
open weights
Params
21B total · 3.6B active
Size
medium
Context
128K tokens
in · textout · text
- GPQA Diamond (high, no tools)71.5%
  OpenAI gpt-oss capability findings ↗Checked 2026-08-05
- Artificial Analysis Intelligence Index (high)15
  Artificial Analysis: gpt-oss-120b vs gpt-oss-20b ↗Checked 2026-08-05
Open-weight MoE (~21B total / 3.6B active). Entry-point coding model for 16 GB-class machines.
gpt-oss-120b
OpenAI
open weights
Params
117B total · 5.1B active
Size
large
Context
128K tokens
in · textout · text
- GPQA Diamond (high, no tools)80.1%
  OpenAI gpt-oss capability findings ↗Checked 2026-08-05
- SWE-Bench Verified (high)62.4%
  OpenAI gpt-oss capability findings ↗Checked 2026-08-05
- Artificial Analysis Intelligence Index (high)24
  Artificial Analysis: gpt-oss-120b vs gpt-oss-20b ↗Checked 2026-08-05
Open-weight MoE (~117B total / 5.1B active). Capacity-class coding route for 80 GB+ machines.
Llama 4 Scout
Meta
open weights
Params
109B total · 17B active
Size
large
Context
10M tokens
in · textin · imageout · text
- MMLU Pro (instruct)74.3%
  Meta Llama 4 model card ↗Checked 2026-08-04
- GPQA Diamond (instruct)57.2%
  Meta Llama 4 model card ↗Checked 2026-08-04
- MMMU (instruct)69.4%
  Meta Llama 4 announcement ↗Checked 2026-08-04
Open-weight MoE (17B active / 109B total) with a 10M-token context window.
Llama 4 Maverick
Meta
open weights
Params
400B total · 17B active
Size
frontier
Context
1M tokens
in · textin · imageout · text
- MMLU Pro (instruct)80.5%
  Meta Llama 4 model card ↗Checked 2026-08-04
- GPQA Diamond (instruct)69.8%
  Meta Llama 4 model card ↗Checked 2026-08-04
Larger Llama 4 MoE (17B active, 128 experts). Total parameter count commonly reported near 400B.
Qwen3.6 27B
Alibaba
open weights
Params
27B
Size
medium
Context
262K tokens
in · textout · text
- Artificial Analysis Intelligence Index (reasoning)37
  Artificial Analysis model comparison ↗Checked 2026-08-04
Open-weight mid-size Qwen tier; strong value for local or self-hosted runs.
Mistral Small 3.3
Mistral
open weights
Params
24B
Size
medium
Context
128K tokens
in · textin · imageout · text
- GPQA Diamond (5-shot CoT)46.13%
  Mistral Small 3.2 model card ↗Checked 2026-08-05
- HumanEval Plus (Pass@5)92.90%
  Mistral Small 3.2 model card ↗Checked 2026-08-05
Dense 24B multimodal open weights. Scores cite the published Small 3.2 instruct card for the same size class.
Gemini 3.5 Flash
Google
closed weights
Params
—
Size
large
Context
1M tokens
in · textin · imagein · audioin · videoout · text
- Artificial Analysis Intelligence Index (high)50
  Artificial Analysis: Gemini 3.5 Flash ↗Checked 2026-08-04
- Artificial Analysis Intelligence Index (launch v4.0)55
  Artificial Analysis: Gemini 3.5 Flash launch ↗Checked 2026-08-04
Hosted multimodal Flash tier. Index scores shifted between methodology versions; both dated claims are linked.
Kimi K3
Moonshot
open weights
Params
—
Size
frontier
Context
1M tokens
in · textout · text
- Artificial Analysis Intelligence Index57
  Santage LLM comparison (citing AA) ↗Checked 2026-08-04
Open-weight long-horizon coding/reasoning model. Parameter count varies by reporting source.
GLM-5
Z.AI
open weights
Params
744B total · 40B active
Size
frontier
Context
200K tokens
in · textout · text
- Artificial Analysis Intelligence Index50
  Artificial Analysis: GLM-5 ↗Checked 2026-08-05
Open-weight MoE (~744B total / 40B active). Useful open frontier baseline beside Kimi.
GLM-5.2
Z.AI
open weights
Params
744B total · 40B active
Size
frontier
Context
200K tokens
in · textout · text
- Artificial Analysis Intelligence Index54
  Artificial Analysis: GLM-5.2 ↗Checked 2026-08-05
Successor release on the GLM-5 base with stronger agentic coding results.
Grok 4.5
xAI
closed weights
Params
—
Size
frontier
Context
500K tokens
in · textin · imageout · text
- Artificial Analysis Intelligence Index (high)54
  Artificial Analysis: Grok 4.5 ↗Checked 2026-08-05
- Artificial Analysis Coding Agent Index (Grok Build)76
  Artificial Analysis: Grok 4.5 launch ↗Checked 2026-08-05
Current xAI flagship reasoning tier. Parameter count undisclosed.
Mistral Large 3
Mistral
open weights
Params
675B total · 41B active
Size
frontier
Context
256K tokens
in · textin · imageout · text
- Artificial Analysis Intelligence Index16
  Artificial Analysis: Mistral Large 3 ↗Checked 2026-08-05
Open-weight MoE (~675B total / 41B active). Current Mistral Large API tier.
Amazon Nova Pro
Amazon
closed weights
Params
—
Size
large
Context
300K tokens
in · textin · imagein · videoout · text
- MT-Bench (median, Claude 3.7 Sonnet judge)8.5
  AWS: Benchmarking Amazon Nova ↗Checked 2026-08-05
Bedrock multimodal Pro tier with 300K context. Parameter count undisclosed.
Command A
Cohere
open weights
Params
111B
Size
large
Context
256K tokens
in · textout · text
- Artificial Analysis Intelligence Index8
  Artificial Analysis: Command A ↗Checked 2026-08-05
Cohere enterprise Command tier. Open weights under a non-commercial research license.
MiniMax-M2
MiniMax
open weights
Params
230B total · 10B active
Size
frontier
Context
200K tokens
in · textout · text
- Artificial Analysis Intelligence Index28
  Artificial Analysis: MiniMax-M2 ↗Checked 2026-08-05
Open-weight MoE (~230B total / 10B active) aimed at agentic coding.
Doubao Seed Code
ByteDance
closed weights
Params
—
Size
large
Context
256K tokens
in · textin · imageout · text
- Artificial Analysis Intelligence Index26
  Artificial Analysis: Doubao Seed Code ↗Checked 2026-08-05
ByteDance Seed coding/reasoning API model. Parameter count undisclosed.
ERNIE 4.5 300B A47B
Baidu
open weights
Params
300B total · 47B active
Size
frontier
Context
128K tokens
in · textout · text
- Artificial Analysis Intelligence Index9
  Artificial Analysis: ERNIE 4.5 300B A47B ↗Checked 2026-08-05
Open-weight MoE (~300B total / 47B active) from Baidu Ernie.