What Andon Labs does

Andon Labs is an AI safety and research startup that develops benchmarks, evaluations, and real-world deployments of autonomous AI agents. Founded in 2023 and Y Combinator-backed, the company builds "Safe Autonomous Organizations" (SAOs) — businesses operated by AI agents without humans in the loop — including a retail boutique (Andon Market in San Francisco), a café (Andon Café in Stockholm), AI-powered radio stations (Andon FM), and vending machine agents deployed at offices of Anthropic, xAI, and OpenAI. The company also publishes a suite of evaluation benchmarks (Vending-Bench 2, Vending-Bench Arena, Blueprint-Bench 2, Butter-Bench) that measure AI capabilities in long-horizon business management, spatial reasoning, and robotic control, with multiple benchmarks adopted by leading AI labs at each model release.

The company's core technology centers on benchmarking and deploying frontier AI agents in physical business environments to study autonomous behavior and failure modes. Its benchmarks evaluate capabilities like inventory management, supplier negotiation under adversarial conditions, spatial reasoning from photographs, and robot orchestration for household tasks. Andon Labs collaborates closely with all four leading frontier AI labs (Anthropic, Google DeepMind, OpenAI, xAI), powering their evaluation pipelines for model releases and providing deployment venues for their models in commercial settings.

The business model combines research and benchmarking services to AI labs (pricing not publicly disclosed) with direct-to-consumer sales of hardware and merchandise, including the Andon Radio ($149, handmade WiFi radio streaming Andon FM) and Andon Abs protein powder ($50). The real-world business deployments function primarily as research platforms rather than profit centers — the SF store has accumulated approximately $13,000 in losses against a $100,000 budget, and the Stockholm café has generated roughly $5,700 in revenue against a $21,000+ budget since launching in April 2026. Revenue figures are not publicly disclosed, and the 1-10 person team appears to operate as a research-driven organization where commercial deployments subsidize the production of evaluation data and research output.

Name

Andon Labs

Legal name

Andon Labs Inc.

Website

https://andonlabs.com

Company type

Private

Founded year

2023

Operating status

Operating

Headcount range

1–10 employees

Short description

Andon Labs is a Y Combinator-backed AI safety research startup that builds Safe Autonomous Organizations — AI agents running real businesses like retail stores, cafés, radio stations, and vending machines — and publishes evaluation benchmarks adopted by Anthropic, OpenAI, Google DeepMind, and xAI.

Ownership category

akta.pro rank

Product category

AI Safety Research and Evaluation

akta.pro primary industry

AI Observability, Monitoring & Evaluation Platforms (Drift, Quality, Safety) (HDAEANAF)

akta.pro secondary industries

Safety & Alignment Evaluation (red-teaming, harmful capability testing) (HDAAAMAL), Responsible AI, Security & Privacy Platforms (Safety, Guardrails, PII) (HDAEANAG)

Keywords

• AI safety benchmarking
• Autonomous agent evaluation
• AI agent deployment
• Long-horizon coherence testing
• Spatial intelligence benchmarks

Headquarters

HQ city

London

HQ country

United Kingdom

HQ region

Europe

Offices2 records

Markets served

GTM type

B2B and B2C

Offering type

Software

Cost components

Personnel, Technology or R&D, Operations, Supply Chain, Marketing or Sales, Infrastructure

Revenue model

1. Product Sales: Direct sales of physical products including the Andon Radio (handmade wooden WiFi radio for streaming Andon FM at $149), Andon Abs protein powder ($50), and other merchandise
2. Research and Benchmarking Services: Collaboration with AI labs (Anthropic, Google DeepMind, OpenAI, xAI) on AI safety research and benchmarking. Vending-Bench is used at every major model release by AI companies.

Pricing tiers

Go-to-market motion1 record

Distribution channels3 records

Marketing channels4 records

Core offering

Andon Labs develops and publishes AI safety benchmarks (Vending-Bench, Blueprint-Bench, Butter-Bench) for evaluating frontier AI model capabilities, and operates Safe Autonomous Organizations where AI agents run real-world businesses without human oversight. Revenue is generated through research collaborations with AI labs and direct sales of consumer products including the Andon Radio hardware and branded merchandise.

Product overview

Andon Labs is an AI safety and research company building Safe Autonomous Organizations (SAOs) — businesses run entirely by AI agents. The portfolio consists of real-world AI deployments (Andon Market retail store, Andon Café, AI radio stations via Andon FM, and vending machine agents) alongside a suite of evaluation benchmarks (Vending-Bench 2, Blueprint-Bench 2, Butter-Bench) that measure AI capabilities in long-horizon business management, spatial reasoning, and robotics. Physical products include the Andon Radio hardware streaming device and merchandise. The core thesis is that AI safety must be studied through real-world deployment rather than simulation alone, with findings used to build control protocols for autonomous AI systems.

Differentiator

Problem solved

Functional benefit

Products and services

• Vending-Bench 2 Year-long vending machine business simulation benchmark evaluating AI model performance on managing a business over extended time horizons. Models are scored on final bank account balance after 365 simulated days, navigating adversarial suppliers, negotiations, customer complaints, and pricing optimization. Claude Opus 4.7 leads with $10,936.76. Built for AI labs and researchers evaluating agentic long-horizon coherence.
• Vending-Bench Arena Multi-agent competitive variant of Vending-Bench 2 where AI agents manage vending machines head-to-head at the same location. Agents can email each other, send money, and trade goods. Leads to price wars, price cartels, and strategic collusion. Enables study of competitive and collaborative AI behavior in business settings for AI safety researchers.
• Blueprint-Bench 2 Spatial intelligence benchmark where AI agents convert approximately 20 apartment photographs into accurate 2D floor plans showing room layouts, connections, and relative sizes. Each agent processes 50 apartments sequentially with a persistent notepad. Claude Fable 5 leads with 0.386 score versus human baseline of 0.586. Tests genuine spatial reasoning beyond pattern matching in training data.
• Blueprint-Bench Original spatial intelligence benchmark testing how AI models understand space by converting apartment photographs into floor plans. Most models performed at or below random baseline (0.279); humans scored 0.547, significantly outperforming all AI systems. Published October 2025.
• Butter-Bench Evaluation of LLMs as robot orchestrators for practical household tasks including 'pass the butter' and delivery tasks. Best model scored 40% versus 95% for humans. Revealed significant spatial intelligence gaps in current LLMs controlling robotic systems. Published October 2025.
• Andon Market World's first fully AI-managed retail boutique in San Francisco's Cow Hollow neighborhood, operated by AI agent Luna powered by Anthropic's Claude Sonnet 4.6. Luna was given a $100,000 budget, corporate credit card, 3-year retail lease, phone, email, internet, and security camera access to hire staff, select inventory, set prices, and manage all business operations.
• Andon Café AI-managed café in Stockholm's Vasastan district operated by AI agent Mona powered by Google Gemini. Mona handles hiring, supplier management, permits, menu pricing, and all business decisions while human baristas prepare coffee. Café generated 44,000 SEK (~$4,659) in revenue in first two weeks of operation.
• Andon FM Multi-station AI radio network where four AI models (Claude Opus 4.8, GPT-5.5, Gemini 3.5 Flash, Grok 4.3) independently run their own radio stations 24/7. Each agent develops a unique personality, curates music, writes and broadcasts DJ commentary via text-to-speech, engages listeners on X/Twitter, and attempts to generate revenue. Powered by Live365 streaming infrastructure.
• Andon Radio Handmade wooden WiFi-connected radio streaming Andon FM stations. Built in American Black Walnut or White Oak by in-house woodworker with one rotary dial for volume, one for channel selection, and USB-C power. Streams to Dayton Audio 2.5" full-range driver. Priced at $149 plus shipping and tax.
• Vending Machine Duo Physical vending machine setup where two AI agents running different models compete side-by-side for customers, revealing which models are best at running a real-world business. Agents live in a Slack channel where users can negotiate, place custom orders, and receive business updates. Currently available via waitlist.
• Andon Abs 2lb chocolate whey protein powder used internally for Butter-Bench human baseline preparation. Priced at $50. Ships in 1-2 weeks.

Quantifiable outcome

• Vending-Bench 2 leaderboard shows Claude Opus 4.7 achieving $10,936.76 money balance in simulated vending machine business over one year
• +3 more outcomes

Segments2 records

Ideal customer profiles2 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

Integration12 records

AI capability13 records

Feature8 records

Partnerships

Three partnerships are on record, tiered core.

• AnthropiccoreTechnology or IntegrationCollaboration on Project Vend - AI-powered vending machine deployed at Anthropic's San Francisco office. Andon Labs' Vending-Bench is used at every Anthropic model release. Claude models power multiple Andon Labs AI agents including Luna at Andon Market.
• Google DeepMindcoreTechnology or IntegrationGoogle DeepMind is listed as a working partner on Andon Labs' website. Gemini models power AI agents including Mona at Andon Cafe Stockholm. Vending-Bench used at Google Gemini model releases.
• OpenAIcoreTechnology or IntegrationListed as working with OpenAI. GPT models power AI agents including DJ GPT at Andon FM. Vending-Bench used at OpenAI model releases.

Scale indicators3 records

Recent moves8 records

Expansion highlights6 records

Direct peers

• Redwood Research: AI safety research organization working on alignment and oversight of advanced AI. Comparable as a peer AI safety research shop, though Redwood focuses more on theoretical alignment than real-world deployment evaluations.
• Conjecture: AI safety company building control and oversight tools for advanced models. Comparable in mission and customer overlap with frontier labs, though Conjecture leans product/tooling while Andon emphasizes benchmark publication.
• SaferAI: AI safety organization producing evaluations and policy recommendations for frontier AI. Comparable as a peer in publishing third-party safety evidence aimed at labs and regulators.
• METR (Model Evaluation and Threat Research): Non-profit that builds rigorous capability evaluations for frontier AI systems (e.g., long-horizon agentic task benchmarks). Directly comparable to Andon Labs' Vending-Bench line in methodology and customer base (frontier labs and government).
• Apollo Research: Independent AI safety evaluation lab focused on scheming, deception, and high-risk behavior in frontier models. Closest direct peer to Andon Labs' safety-evaluation thesis, though Apollo leans on red-team probing while Andon emphasizes long-horizon real-world deployment.

Emerging players

• Holistic AI: AI governance, risk, and compliance platform serving enterprises subject to the EU AI Act and similar regulations. Adjacent in the AI safety/GRC market though focused on enterprise compliance rather than frontier-model evaluation.
• Credo AI: AI governance and audit platform helping enterprises document AI risk and compliance. Comparable to Andon's governance framing but aimed at enterprise buyers rather than frontier-model developers.
• Patronus AI: Commercial LLM evaluation platform offering safety, hallucination, and quality scoring. Comparable customer base (enterprises deploying LLM agents) but narrower focus on production LLM quality rather than long-horizon agent safety.

Others

• Arize AI: ML/LLM observability platform providing drift, quality, and safety monitoring for models in production. Adjacent infrastructure that complements but does not directly compete with Andon's frontier-safety benchmarks.

Broad incumbents

• Scale AI: Large incumbent in AI data infrastructure and evaluation, including its Safety, Evaluations, and Alignment (SEA) unit working with frontier labs. Overlaps with Andon's benchmarking work at much greater scale and broader scope.

Market position

Strengths4 records

Weaknesses4 records

Competitive moat5 records

Key risks6 records

Key highlights5 records

Customer concentration

• Xx.com/andonlabs
• LinkedInlinkedin.com/company/andonlabs

Revenue estimate

Valuation estimate

Number of profiles

Profiles2 records

Funding overview

Funding rounds

Investors

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Andon Labs