Runner H

Runner H is the agentic orchestration product developed by H, a Paris-based AI company, that accepts natural-language objectives from users and dispatches them across browsers, applications, and external tools by combining a foundation language model with the Holo visual-grounding model line. The product entered closed beta in November 2024 and reached public beta on June 3, 2025, alongside the open-weights Holo-1 release and the launch of the Surfer H browser-use agent and the Tester H automated software testing product. Runner H reported a 67 percent success rate on the WebVoyager browser-agent benchmark at launch, ahead of the contemporaneous Anthropic Computer Use baseline at 52 percent, and the integrated Surfer H plus Holo-1 system later reported a 92.2 percent success rate on a public computer-use benchmark.

At a glance

• Lab: H (formally H Company)
• Released: Closed beta in November 2024. Public beta on June 3, 2025. Surfer 2 successor announced subsequently.
• Modality: Agentic orchestration. Inputs are natural-language objectives; the system grounds in vision through Holo and dispatches actions across web browsers, applications, and APIs.
• Open weights: No. Runner H is a commercial product. The underlying Holo-1 visual-language model is open-weights under Apache 2.0; the Runner H orchestration layer is closed.
• Context window: Multi-step task horizons across browser sessions and application chains, with task length effectively bounded by the agent runtime and per-task pricing rather than a fixed token context.
• Pricing: Per-task pricing. H Company reported approximately $0.13 per task for Surfer H paired with Holo-1-7B, compared to $0.54 per task for a comparable GPT-4.1 baseline at the June 2025 launch. Specific tier pricing for Runner H subscriber and enterprise access through the H Company developer surface.
• Distribution channels: Runner H public beta, enterprise contracts through H Company sales, developer access through the H Company API surface.

Origins

H Company was founded in May 2024 in Paris by Charles Kantor (initial CEO), Karl Tuyls, Laurent Sifre, Julien Perolat, and Daan Wierstra. Tuyls, Sifre, Perolat, and Wierstra had been senior research scientists at Google DeepMind with collective contributions to the AlphaGo, AlphaZero, and MuZero research lines. The company raised a $220 million seed round in May 2024, the largest European AI seed at the time, led by Accel with participation from Amazon, UiPath, FirstMark, Bpifrance, and Innovation Endeavors.

The founding thesis was that the next phase of AI commercial value would accrue to systems that can autonomously complete multi-step tasks across digital and operational environments, rather than to conversational chatbots. The company's framing of the model class was "frontier action models," foundation models trained for autonomous task execution rather than conversation, with the broader product positioning of "Holistic and Humane" agentic AI.

Runner H entered closed beta in November 2024 as the company's first product, a "frontier action model" combining LLM reasoning with autonomous task execution across browsers, applications, and APIs. The closed beta evaluation period produced the early WebVoyager benchmark results: 67 percent for Runner H against the contemporaneous Anthropic Computer Use baseline at 52 percent, with Runner H characterized in the company's reporting as faster and more accurate on comparison runs.

The June 3, 2025 public launch was a bundled release that included Runner H public beta, Surfer H (the browser-use agent), Tester H (an automated software testing product), and the open-weights Holo-1 visual-language model under Apache 2.0. The bundled framing positioned Runner H as the orchestration layer with Holo-1 as the visual grounding component and Surfer H and Tester H as application-specific products. The launch reported a 92.2 percent success rate on a public computer-use benchmark for the integrated Surfer H plus Holo-1-7B system, characterized as state-of-the-art for the category at the time and a 5.5 times cost reduction relative to peer GPT-4.1-based baselines.

In June 2025, Charles Kantor was replaced as CEO by Gautier Cloix, formerly the managing director of Palantir's French unit. The leadership change framed the transition from research-led founding to commercial-execution scaling. Three of the five original co-founders (Tuyls, Wierstra, Perolat) had departed in late 2024 over reported "operational differences." Karl Tuyls subsequently became director of Meta's Paris AI research lab.

A successor release line including Surfer 2 (97.1 percent WebVoyager success in subsequent reporting) continues the model line's evolution into 2026.

Capabilities

Runner H is built specifically for multi-step agentic task execution. The system accepts a natural-language objective from a user, plans the steps required to complete the objective, dispatches actions across browsers and applications, observes the results, and iterates as needed.

Three capability features distinguish Runner H from peer agentic systems.

The first is the Holo visual grounding integration. Runner H pairs with the Holo-1 visual-language model line for screen-element identification, allowing the system to see and interact with arbitrary user interfaces rather than only those exposed through structured APIs. The visual grounding capability is the structural reason Runner H can dispatch actions across web applications, native desktop applications, and any UI surface that the model has been trained to read.

The second is the cost-per-task economics positioning. H Company has published per-task cost figures for the integrated Surfer H plus Holo-1 system at approximately $0.13 per task, against $0.54 per task for a comparable GPT-4.1-based baseline at the June 2025 launch. The cost positioning is enabled by the open-weights Holo-1 backbone, which avoids per-token frontier API pricing for the visual grounding component.

The third is the bundled product line. Runner H operates as the orchestration layer that drives Surfer H (browser tasks) and Tester H (software QA testing), in addition to direct Runner H deployments. The bundled approach positions H Company as a platform for agentic AI rather than as a single product.

The orchestration architecture handles task decomposition, action dispatch, observation, and re-planning. The system is positioned for both consumer use cases (research tasks, scheduling, transactional workflows) and enterprise applications (QA automation through Tester H, browser automation for operational workflows through Surfer H).

Benchmarks and standing

Runner H's principal disclosed benchmarks are agentic-task success rates on browser-use evaluation suites.

On WebVoyager, the standard browser-agent benchmark released by Emergence and the academic web-agent research community, Runner H reported 67 percent at the November 2024 evaluation, against the Anthropic Computer Use baseline at 52 percent on the same evaluation set run from the same geographic region in November 2024. The comparison was characterized in H Company's reporting as evidence of both higher accuracy and faster execution.

On the integrated Surfer H plus Holo-1-7B configuration at the June 2025 launch, the company reported a 92.2 percent success rate on a public computer-use benchmark (after 10 attempts), characterized as state-of-the-art for the category at the time. The reported per-task cost was approximately $0.13, against $0.54 per task for the comparable GPT-4.1-based baseline.

On WebVoyager specifically, the Surfer 2 successor system has reported a 97.1 percent success rate, surpassing Magnitude's previous state-of-the-art of 93.9 percent. The benchmark trajectory across H Company's product line indicates continued capability improvement through 2025 and into 2026.

Independent verification across third-party leaderboards has been mixed. The browser-agent and computer-use benchmark category is fragmented, and direct head-to-head comparisons with peer agent products from OpenAI, Anthropic, and Google DeepMind on shared evaluation infrastructure have not been published consistently.

The standard horizontal language model benchmarks (Artificial Analysis Intelligence Index, LMArena, GPQA Diamond, AIME, SWE-bench) do not directly apply to Runner H's agentic positioning, although SWE-bench is a relevant evaluation for the Tester H software-testing component of the H Company product line.

Access and pricing

Runner H is accessible through the H Company developer surface and through enterprise contracts.

Public beta access is available through the Runner H product page at hcompany.ai/runner-h. Enterprise pricing handled through direct H Company sales engagement, with the Tester H product targeting QA workflows, Surfer H targeting browser-automation tasks, and Runner H operating as the orchestration layer that can incorporate either or both.

The cost-per-task economics published at launch (approximately $0.13 per task for Surfer H plus Holo-1-7B against $0.54 for comparable GPT-4.1-based baselines) indicate the pricing positioning, though the company has not published a fully transparent tiered pricing schedule.

The underlying Holo-1 visual-grounding model is open-weights under Apache 2.0 (7B variant), which means enterprises can self-host the visual-grounding component if they choose. The orchestration layer of Runner H itself is closed and accessed through the H Company surface.

Comparison

Direct competitors and adjacent agentic AI systems:

• OpenAI Operator and ChatGPT agent (OpenAI). Principal US competitor on consumer and enterprise agentic tasks.
• Claude Computer Use (Anthropic). Closest direct competitor on browser-and-application automation. Claude scored lower than Runner H on the November 2024 WebVoyager comparison, though the benchmark category has evolved since.
• Gemini agent capabilities and Project Astra (Google DeepMind). Principal frontier-lab agentic offering. Direct competitor.
• Cognition AI Devin, Imbue. Adjacent agentic startups; Devin is positioned for software engineering rather than general orchestration.
• Mistral AI, Aleph Alpha. European AI peers emphasizing foundation models rather than agentic products.
• Surfer 2 (same lab). Successor cross-platform agent. Runner H operates as the orchestration layer, while Surfer 2 represents the next generation of cross-platform computer-use execution.

Runner H's distinctive position among 2024 to 2025 agentic AI products: integration with the open-weights Holo-1 visual grounding line, the cost-per-task pricing advantage relative to GPT-4.1-based baselines, the European founding pedigree from the DeepMind-derived team, and the bundled orchestration platform that includes Surfer H and Tester H as adjacent products.

Outlook

Open questions for Runner H over the next 6 to 18 months:

• Commercial customer traction. H Company's pricing positioning is durable only if the underlying capability advantage holds and named enterprise customers convert into contracted volume. Disclosed enterprise reference customers are a watchable signal.
• Successor capability cadence. Runner H public beta released June 2025, and the Surfer 2 successor extended the line shortly after. The pace of subsequent capability disclosures, including against frontier-lab agent products from OpenAI, Anthropic, and Google DeepMind, will affect commercial standing.
• Open-weights Holo cadence. The Apache 2.0 Holo-1-7B release was a structural part of the Runner H value proposition (per-task cost economics). Whether H Company sustains the open-weights cadence on the Holo line, given the Holo-1.5-72B research-only license shift, will signal commercial strategy.
• Leadership stability. H Company experienced significant founder churn in 2024 and a CEO change in June 2025. Continued senior leadership stability under Cloix is a risk factor characterized in industry coverage.
• Series A or growth round close. The implied valuation at the May 2024 seed (just under $2 billion) has not been refreshed publicly. Whether the company closes a follow-on round and at what terms will affect the operating runway and competitive positioning.

Sources

• Surfer-H Meets Holo1: Cost-Efficient Web Agent Powered by Open Weights (arXiv). Technical paper covering Surfer H plus Holo-1 architecture and benchmarks.
• BusinessWire: H Company Launches Next-Generation Autonomous AI Agents. June 2025 product suite launch.
• H Company blog: Charting a New Route. Technical detail on Runner H WebVoyager results.
• MarkTechPost: H Company Releases Runner H Public Beta Alongside Holo-1 and Tester H. Product launch coverage.
• H Company blog: Surfer 2. Successor release reference.
• H Company Runner H product page. Public beta access surface.
• TechCrunch: French AI startup H raises $220M seed round. Founding context.
• Steel.dev AI Browser Agent Leaderboard. Benchmark reference for browser agents.