Scientific, fair, real-time, intuitive AI influence index
What is the Eins Index?
Eins Index is a real-time AI influence index: third-party open data forms the backbone and self-collected public data supplements it, turning public events across the world's AI labs into comparable numbers — the 8100 composite index and the live board answer 'who has the strongest AI influence right now'.
Show: data sources, methodology and more
- The vocabulary covers 12 AI labs, and 10 of them are shown live on the public board.
- 8100 combines event impact, relative strength and breadth, and each lab's contribution can be broken down item by item.
- Values keep refreshing during running hours; they are assessment results converted by this site under a public methodology, for reference only, and do not replace official figures.
- Values draw on four planes: labs' official releases, public-discussion signals from the global opinion plane, third-party open benchmarks for the capability plane, and adjudicated major events.
- Fairness and verifiability are built into the mechanism: the capability plane rests on replaceable third-party open data, conservation and audit inequalities are checked every tick, and the correction log is publicly auditable.
- Beyond the overall board, a second main line tracks global AI news and trends in real time by lab: each one has its own page and event stream, so you can follow that organization's developments, influence breakdown and heat changes on its own.
See the methodology whitepaper for the full formulas and governance rules
Content revision 2026.09.23-eins-m6t3r8
Market is openValue refresh cadence: values update during the daily sessions (UTC 07:00–16:00 and 19:00–04:00 the next day) and pause otherwise (UTC 04:00–07:00 and 16:00–19:00); the status label shows the current phase.4 up6 down0 —· Today's top story：How to evaluate DeepSeek V4.1 Flash 1017?[DS](3.23)
US is Leading +6.65Market StateOpenLast update 17:59:35
BreadthAdvancing4Declining6Unchanged0
Global AI Origin LeaderboardRanking is sorted by current value, highest first.
Live · winner takes all
- DeepSeekDSDeepSeek89.40-2.34%A1−Top 1 in 2d
▲Heat = cumulative flow since today's open; reflects attention.
- How to evaluate DeepSeek V4.1 Flash 1017?Heat 1.08
- DeepSeek-v4.1-Flash-1017Heat 0.81
- DeepSeek: DeepSeek v4.1 Flash 1017Heat 0.55
▼
- On October 15, DeepSeek's website and APP showed "服务器暂时不可用"; a compensation plan was later released. What might have caused this?Heat 0.07
- On October 15, DeepSeek's website and APP displayed "服务器暂时不可用". What could be the possible reasons?Heat 0.05
- How to evaluate DeepSeek's compensation credits issued to users based on the last seven days of API consumption after a 5-hour outage?Heat 0.02
- 2Google DeepMindGDMGemini80.31-1.17%A1−Top 2 in 10dFour Gemini 4 Argon agents beat Minecraft from scratch in 27 hours. Operating in a sandbox using only code and text, theDuring safety testing of Gemini 4, researchers found it "eliminated months of trial-and-error research" on biological weapons (e.g. plague). Google's own testers got enough out of it to plan a biological attack.
- 3OpenAIOAIGPT71.32-2.30%A1−Top 3 in 4dGPT-6 Astra Pro in ChatGPT got an upgrade with faster responses and improved use of plugins and tools. And now, multipleThree Chinese labs just got caught hacking 1.4% of OpenCode's users' personal OpenAI and Anthropic API key usage for distillation, typically stealing hundreds to thousands of dollars from each
- 4AnthropicANTHClaude69.78-3.05%A1−
- 5QwenQWENQwen68.96-0.04%A1−
- 6xAIXAIGrok64.99+3.95%A1−
- 7MetaMAIMuse Spark61.32-1.45%A1−
- 8Zhipu AIZAIGLM56.40+3.52%A1−
- 9Moonshot AIKIMIKimi47.83+4.23%A1−
- 10MiniMaxMMMiniMax41.46+4.08%A1−
Frequently Asked Questions
Thirty seconds on what it is, how to use it, and how it differs from model benchmark leaderboards.
How is AI influence measured?
Eins Index grades each lab's public events and developments into five contribution tiers (side A positive, side B negative), then aggregates them into a live value using the formulas in the methodology whitepaper. A higher value means stronger influence right now.How is this different from model benchmark leaderboards?
Benchmark leaderboards score models on fixed test sets and move at release cadence. Eins Index measures a lab's real-world influence right now and refreshes with events. The two are complementary, not interchangeable.How often is the board updated?
Values refresh continuously during active sessions and pause outside them. The footer shows the timestamp of the latest data update, with session times displayed in the time zone of the current language.Where does the data come from?
Four planes: labs' official releases, public-discussion signals from the global opinion plane, third-party open benchmarks for the capability plane (Hugging Face, GitHub, SWE-bench, Aider, LiveCodeBench, LMArena, OpenAlex), and adjudicated major events. Eins Index is an assessment-method layer: third-party open data forms the backbone, self-collected public data supplements it, and the site neither produces, owns nor sells the underlying data. Converted values do not replace official figures. Full rules are in methodology whitepaper §13.How many AI labs are covered?
The vocabulary covers 12 AI labs, and 10 of them are shown live on the public board. The other 2 are not displayed on the public board and have no separate page.Can I use this data for investment decisions?
No. Eins Index measures AI influence, not the price of any asset. The site is not a live trading venue and provides no investment advice.How does Eins Index ensure fairness?
Through three institutional designs, not claims. First, the capability plane is built on third-party open data; the site does not compete with those sources, so it never measures itself with its own ruler. Second, every source has an independent circuit breaker and is replaceable — one source failing or being manipulated does not take down the index. Third, asymmetric disclosure: hard-to-game methodology and reference parameters are public, easy-to-game short-horizon parameters are not. This is independent third-party assessment, not endorsement: no evaluated organization's marks are used. See whitepaper §7.2.How can the numbers be checked? What if something is wrong?
The system is designed to fail in publicly detectable ways: the conservation identity and the audit inequality are checked every tick, and any violation raises a data-integrity alert. Four continuous audits run on top (including replay and reconciliation), and the replay audit allows verification without reproduction. Judgement criteria are stated before the evidence exists, so standards cannot be picked after the fact. When a correction is needed, three rules apply: correct toward the truth and propagate forward (history is annotated, never silently edited), never fabricate a fix, and keep the correction log auditable. Limitations are stated openly: a short observation window, 12 labs covered, and an RI real-time component that is not externally reproducible — the price of manipulation resistance. See whitepaper §8 and §9.Are event polarity and impact decided by AI? Wouldn't human judgement be subjective?
Two layers, and the human authority does not sit with the operator. Layer one is the machine first pass: each event gets a polarity and a strength under a published scoring guide, with fixed scoring settings and a per-event record; low-confidence items go to human review, whose conclusion takes precedence over the machine call and is recorded the same way. Layer two is institutional adjudication: the four categories that require human judgement — admission and valuation of the academic ledger, designation of major risk items, hyperparameter and algorithm changes, and changes to the inclusion list — are decided by the independent Eins arbitration committee. Decisions are published as notices and written into the corresponding version number so they can be checked externally; committee membership is not disclosed for now, and members must declare conflicts of interest and recuse themselves from related matters.Why is the intraday line discontinuous, with visible gaps?
This is by design, not missing data. The index runs in sessions: during a paused session public events are still collected and queued as usual, but nothing is injected into the value and only time decay applies, so that window produces no new points. When the session opens again, one consolidated settlement pass runs first — everything accumulated while paused is injected at once and a blank marker is left in the series, which is what shows up as a gap. The discontinuity therefore comes from 'no settlement while paused, one consolidated settlement at the open', not from a collection interruption or data loss. Session times are displayed in the time zone of the current language.