LLM Intelligence Dashboard

5 live pumps Live snapshot: Jul 15, 2026 Models tracked: 360+
LiveBench Vellum Vellum Assistants Vellum Coding Vellum Open-Source Onyx YDN Arena Snapshots LLM Stats Swfte Arena WhatLLM OpenLM Arena Artificial Analysis Arena.ai OpenRouter LiteLLM X Signals Jina Reader RSS Releases
Daily Weekly Monthly Bi-monthly
Overall Leader
Claude Fable 5
LiveBench 80.8 · Arena #1 · SWE 95.0
Best Open-Weight
GLM 5.2
HLE 54.7 · GPQA 91.2% · Terminal 81%
Best Value (Cost/Task)
Grok Build 0.1
$0.024/task · Overall 67.8
Fastest Inference
Step-3.5-Flash
629 tok/s output throughput
Largest Context
Grok 4 Fast
2,000,000 tokens
Highest Arena Elo
Claude Fable 5
1525 Elo · Swfte composite 100/100

Frontier Models — Cross-Source Consensus

Model Provider License HLE LiveBench Arena Elo SWE-Bench GPQA Input $/1M Output $/1M Context Sources
GPQA Diamond Leader
Claude Sonnet 5
96.2% — Graduate-level science
HLE Leader
Claude Mythos 5
64.5% — Humanity's Last Exam
ARC-AGI 2
Kimi K2.5 / Claude Opus 4.6
12% — Visual fluid intelligence

Reasoning Benchmarks Compared

Model Provider HLE % GPQA Diamond % ARC-AGI 2 % LiveBench Reasoning MMMU-Pro Source
SWE-Bench (Agentic)
Claude Mythos 5
95.5% — Real GitHub issues resolved
LiveCodeBench
DeepSeek V4 Pro
93.5% — Live competitive programming
Terminal-Bench 2.1
GLM 5.2
81% — Multi-step terminal tasks
BFCL (Tool Use)
GPT-4.5
69.9% — Berkeley Function Calling

Coding & Engineering Benchmarks

Model Provider SWE-Bench % LiveCodeBench % Terminal-Bench % BFCL % MATH 500 % LiveBench Coding Source
Computer Use (OSWorld)
Claude Fable 5
85% — Real desktop GUI tasks
Browsing (BrowseComp)
Claude Fable 5
88% — Agentic web search
Work Automations
Claude Fable 5
17.4% — AutoBench multi-step
LiveBench Agentic Coding
GPT-5.6 Sol Max
65.6%

Agentic & Tool-Use Benchmarks

Model Provider OSWorld % BrowseComp % AutoBench % LiveBench Agentic License Source
LiveBench Mathematics
GPT-5.6 Sol Max
96.2%
AIME 2025
Gemini 3 Pro
100.0%
LiveBench Data Analysis
GPT-5.5 Thinking xHigh
81.6%

Mathematical & Analytical Benchmarks

Model Provider LiveBench Math LiveBench Data Analysis MATH 500 % AIME 2025 % Source
Text Arena Leader
Claude Fable 5
1525 Elo · 6M+ votes
Swfte Composite #1
Claude Fable 5
100/100 · 357 models ranked
Chatbot Arena (Onyx)
Claude Opus 4.6
1503 Elo
LLM Stats Score #1
Claude Fable 5
2138 composite score

Human Preference Arena Rankings

Arena Elo ratings from blind human evaluation (OpenLM Arena), aggregated by Swfte, LLM Stats, and Onyx. Updated daily from fresh matchups.

Model Provider Arena Elo Swfte Quality LLM Stats Score Onyx Arena License
Fastest Throughput
Step-3.5-Flash
629 tok/s output
Lowest Latency (TTFT)
Baseten (hosting)
0.27s time-to-first-token
Cheapest Frontier
DeepSeek V4 Flash
$0.14 input · $0.28 output per 1M
Best Cost/Task
Grok Build 0.1
$0.024 per successful task

Inference Performance & Pricing

Model Provider Input $/1M Output $/1M Speed (tok/s) TTFT (latency) Context Cost/Task Source
Open HLE Leader
GLM 5.2
54.7% — Humanity's Last Exam
Open SWE-Bench
DeepSeek V4 Pro
80.6% — Agentic coding
Open BrowseComp
DeepSeek V4 Flash
85.9% — Web browsing
Open Terminal-Bench
GLM 5.2
81% — Terminal use

Open-Weight Model Rankings

Model Provider HLE % GPQA % SWE-Bench % OSWorld % BrowseComp % Terminal % Speed Input $/1M
AI Assistants Compared
4 Platforms
Vellum, Hermes, Claude Cowork, OpenClaw
Most Channels
Hermes
iOS, macOS, CLI, Voice, Email, TG, Slack
Largest Skill Ecosystem
OpenClaw / Hermes
Self-improving skills, MCP, plugins

AI Assistant Platform Comparison (Vellum 2026)

Feature Vellum OpenClaw Hermes Claude Cowork

Provider × Benchmark Heatmap

Normalized scores (highest in each column = 100). Green = top performer, amber = mid, red = lowest. Aggregated from LiveBench, Vellum, Onyx, and LLM Stats.

OpenRouter Model Tracker

Live pricing & capabilities from openrouter.ailoading...0 models

Model Provider Context Input $/1M Output $/1M Modality

Price vs Intelligence — Frontier & Budget Models

Three visualization types comparing OpenRouter blended pricing ($/1M tokens) against benchmark-derived intelligence scores. Prices from OpenRouter API (Jul 15, 2026 snapshot). Intelligence = composite of HLE, LiveBench, GPQA, SWE-Bench, Arena Elo.

Best Value (Highest Intell / $)
Smartest Frontier Model
Cheapest Capable Model
Models Compared
with both price & intelligence data
0
$100

1. Scatter — Blended Price ($/1M) vs Intelligence Score (log scale)

Bottom-left = best value. Top-right = premium frontier. Hover for details.

2. Bubble — Price vs Intelligence (Bubble Size = Context Window)

Larger bubbles = bigger context windows. Color by provider family.

3. Radar — Top 8 Models Multi-Dimensional Comparison

Normalized scores across 6 dimensions. Higher = better performance per axis.

Raw Data

Model Provider Intelligence Score Prompt $/1M Completion $/1M Blended $/1M Context Value Ratio
LiveBench Release
public CSV · contamination-free
Models in Release
auto-refreshed by hr / refresh_dashboard_data.sh
Top Overall

LiveBench Live Categories

Source: livebench.ai public table_.csv (no auth). Category means computed client-side from task scores.

Rank Model Overall Reasoning Coding Agentic Coding Mathematics Data Analysis Language Instruction Following Mean Cost
Artificial Analysis
via Jina Reader (keyless)
Vellum Categories
via Jina Reader (keyless)
Fetched
r.jina.ai · no API key

Artificial Analysis — Intelligence Index

Live scrape of artificialanalysis.ai via Jina Reader. Intelligence score, blended $/1M, output speed (tok/s).

#ModelProviderContext IntelligenceBlended $/1MOutput tok/s

Vellum — Per-Benchmark Leaders

Release Items
keyless RSS/Atom
Feeds Live
labs + inference engines
Fetched
feedparser · no API key

Model / Engine Release Feed

Newest-first from OpenAI, Google DeepMind, Hugging Face, Together, Mistral, Qwen, vLLM, SGLang. Anthropic has no public RSS (omitted). Keyless.

X Feed Status
via xurl → X API v2
Posts Captured
0
preferred lab + arena accounts
Account
local xurl default app

Model / Leaderboard Signals from X

Preferred accounts: ArtificialAnlys, lmsysorg, OpenRouterAI, xai, AnthropicAI, OpenAI, GoogleDeepMind, deepseek_ai, Alibaba_Qwen. When credits are depleted the panel stays wired and surfaces the block clearly.

All Data Sources — Methodology & Update Cadence