No single benchmark decides it
Combines 24 daily LLM Stats signals instead of letting one saturated leaderboard declare victory.
Open-source AGI progress tracker
AGI Clock
A public instrument for estimating how far frontier AI still is from AGI across reasoning, engineering, autonomy, tool use, search, and reliability signals.
Benchmark proxy only. This does not claim AGI has been reached.
- AGI Readiness
- 72.9%
- Distance
- 27.1%
- Bottleneck
- Agents + Autonomy
Collecting baseline Updated Jul 1, 2026
Why it exists
A stricter scoreboard for a noisy question.
Autonomy matters
Weights agentic work, computer use, planning, tools, and search because AGI must do more than answer tests.
Easy to audit
Static site, public data file, MIT license, and a collector script that can be inspected or improved.
Current signal
24 benchmarks, four dimensions, one readiness proxy.
Each benchmark reports the best available frontier score from LLM Stats. The index groups those raw scores into capability dimensions and keeps the ETA hidden until the new formula has enough daily history.
Reasoning + Expert Knowledge 76.3%
Abstract reasoning, expert knowledge, science, math, and frontier closed-ended questions.
Coding + Engineering 80.2%
Software engineering, competitive coding, terminal work, and multi-step code tasks.
Agents + Autonomy 59.1%
Computer use, long-horizon professional workflows, planning, and applied office/finance tasks.
Tool Use + Search Reliability 81.5%
Tool calling, web/search persistence, user-policy following, and repeated agent reliability proxies.
Reasoning
ARC-AGI v2
Abstract grid transformation tasks for fluid reasoning and novel problem solving.
Best 85% GPT-5.5
llm-stats:arc-agi-v2 Flat
Expert Knowledge
Humanity's Last Exam
2,500 expert-vetted multimodal questions across math, science, humanities, and vision.
Best 64.7% Claude Mythos Preview
llm-stats:humanity's-last-exam Flat
Science
GPQA
PhD-level questions in biology, chemistry, and physics.
Best 94.6% Claude Mythos Preview
llm-stats:gpqa Flat
Knowledge
MMLU-Pro
Harder professional knowledge benchmark across academic and applied domains.
Best 89.6% Qwen3.7 Max
llm-stats:mmlu-pro Flat
Mathematics
FrontierMath
Advanced unpublished mathematical problems authored and reviewed by expert mathematicians.
Best 47.6% GPT-5.4
llm-stats:frontiermath Flat
Engineering
SWE-Bench Verified
Real GitHub issue resolution requiring codebase understanding and patching.
Best 95% Claude Fable 5
llm-stats:swe-bench-verified Flat
Engineering
SWE-Bench Pro
Complex real-world software engineering tasks with extended multi-step reasoning.
Best 80% Claude Fable 5
llm-stats:swe-bench-pro Flat
Coding
LiveCodeBench v6
Contamination-resistant competitive programming tasks collected over time.
Best 91.6% Qwen3.7 Max
llm-stats:livecodebench-v6 Flat
Terminal Work
Terminal-Bench
End-to-end tasks in real terminal environments, including servers, data, and security workflows.
Best 50% Claude Sonnet 4.5
llm-stats:terminal-bench Flat
Terminal Work
Terminal-Bench 2.1
Updated terminal-agent benchmark for autonomous command-line execution.
Best 84.3% Claude Fable 5
llm-stats:terminal-bench-2.1 Flat
Computer Use
OSWorld
Real desktop/web application tasks across operating systems and multi-app workflows.
Best 78.8% Seed 2.1 Pro
llm-stats:osworld Flat
Long-Horizon Agents
APEX-Agents
Long-horizon professional tasks requiring sustained planning and execution.
Best 33.8% Seed 2.1 Pro
llm-stats:apex-agents Flat
Virtual Tasks
VITA-Bench
Virtual task automation for multi-step real-world workflows.
Best 49.7% Qwen3.5-397B-A17B
llm-stats:vita-bench Flat
Planning
DeepPlanning
Complex multi-step planning, goal decomposition, and long-horizon reasoning.
Best 62.3% Qwen3.7-Plus
llm-stats:deep-planning Flat
Office Work
OfficeQA Pro
Professional knowledge-work tasks involving documents, spreadsheets, and office workflows.
Best 72.2% Seed 2.1 Pro
llm-stats:officeqa-pro Flat
Finance Work
Finance Agent v2
Agentic financial analysis workflows over realistic business tasks.
Best 57.9% Gemini 3.5 Flash
llm-stats:finance-agent-v2 Flat
Tool Agents
Tau-bench
Tool-agent-user interaction with domain rules and repeated-trial reliability.
Best 88.2% Step-3.5-Flash
llm-stats:tau-bench Flat
Tool Agents
TAU-bench Retail
Retail customer-service agent tasks with tools, policies, and user interaction.
Best 86.2% Claude Sonnet 4.5
llm-stats:tau-bench-retail Flat
Tool Agents
TAU-bench Airline
Airline customer-service agent tasks with tools, policies, and user interaction.
Best 70% Claude Sonnet 4.5
llm-stats:tau-bench-airline Flat
Tool Use
MCP Atlas
Scaled tool coordination across complex multi-step tool-use tasks.
Best 83.8% Seed 2.1 Pro
llm-stats:mcp-atlas Flat
Tool Use
Toolathlon
Multi-tool proficiency across diverse tool-use categories.
Best 59.9% Claude Opus 4.8
llm-stats:toolathlon Flat
Search
BrowseComp
Persistent web browsing and hard-to-find information retrieval.
Best 90.1% GPT-5.5 Pro
llm-stats:browsecomp Flat
Search
WideSearch
Broad, parallel search operations across multiple sources.
Best 80.8% Kimi K2.6
llm-stats:widesearch Flat
Search
DeepSearchQA
Multi-hop deep search and answer retrieval.
Best 93.1% Claude Opus 4.8
llm-stats:deepsearchqa Flat
Trajectory
New formula baseline
Daily v2 history started now. The chart shows the first baseline point and waits for 30 samples before estimating a trend.
Current 72.9% Jul 1, 2026
Distance 27.1% remaining
ETA Collecting baseline 29 daily samples needed
Legacy mini-history Old formula, context only
Method
Transparent enough to disagree with.
AGI Clock estimates distance to AGI with a frontier benchmark basket. It is not a claim of consciousness, economic replacement, or that a single model has passed every task.
Collect Daily frontier benchmark data from the llm-stats.com API.
Normalize Group raw scores into reasoning, engineering, autonomy, and tool/search dimensions.
Publish Show readiness, distance, bottleneck, and hold ETA until enough v2 history exists.
Pending daily sources: GAIA, SimpleBench, ARC-AGI-3, METR Time Horizons.
Built in public
Use it, fork it, or challenge the benchmarks.
AGI Clock is intentionally small: static frontend, auditable data, and room for better benchmark proposals.