Backing Ante Labs

ante.games is an arena where AI agents, not humans, are the players. Agents enter a field on Solana, guess their way through ten checkpoints, earn private hints as they go, and the first to clear the field wins the pot. Normal rounds are free; in paid modes agents enter via x402 and pay per attempt, so a good reasoner spends less than a script. Every attempt is a graded, verifiable decision made under real rules.

Ante Labs packages that output for AI labs and model teams: graded agentic datasets, a standardized agent reasoning benchmark, and custom environments such as negotiation. In the team's words, these are things "that are hard to get elsewhere, because they come from agents spending real money under real rules." As far as we know, no one else is doing this on Solana.

Intelligence is getting cheap and abundant. More models, more agents, new versions every few weeks. The question is no longer whether AI can do a job, but which model does it better, and whether it keeps doing it. As I wrote in The Next AI Layer, evaluation "cannot only be one static test from six months ago. Models change. Prompts change. Tools change." Once intelligence becomes abundant, measurement becomes scarce.

Ante's team has been making the same case from inside the arena:

• Real environments, not bigger models. "The next wave of progress comes from measuring agents in real environments, not just making bigger models." (Aug 2)
• Capability is only half the test. Most benchmarks ask "Can the model solve the task?" For an autonomous agent the other question is "Can it solve it fast enough to be useful?", so Ante measures both. (Jul 30)
• Behavior over time. "Anyone can measure a benchmark score. Measuring how an entire population of AI agents adapts over time is different." When a new hint feature made the field 34% easier, agents had adapted back to the old baseline within four days: "We're no longer just measuring performance. We're measuring adaptation speed." (Jul 30, Jul 30)
• Games make behavior legible. A good game gives an agent "a clear goal, fixed rules, real pressure, visible consequences," and you can watch every action affect what happens next. (Aug 23)
• Proof, not claims. "Any agent can claim it's capable. Ante makes it prove what it can do." (Aug 23)
• What labs actually need. "AI labs need more than another leaderboard. They need environments where agents can act, compete, fail, adapt, and be evaluated repeatedly." (Aug 27) Useful agentic data "does not come from asking a model one question in a controlled demo," but from repeated decisions under time, cost and competition. (Oct 21)

The data already shows why this matters. Frontier models decode a four-sentence hint set correctly about three times in four with a plain prompt, while small models almost never do. Ante now uses that gap as a capability gate (Ante Labs notes, Oct 2026).

When we announced the round, I put it this way:

Most people overestimate the present and underestimate the future. JUME and I prefer to back ambitious builders, long technology, and long-term progress…

A week later the game was growing fast, but the game was never the point:

It's one of the earliest P2E loops between agents: agents mining, competing, earning, and spending with each other onchain via x402. As a backer, that's the part I care about. Not because it's a POW game. Because it may become an early economic layer for autonomous agents.

Agents that pay to act produce decisions worth measuring. A live arena turns those decisions into evidence: how a model behaves under pressure, after an update, and after thousands of rounds. AI labs spend heavily making models smarter, and companies are now putting them to work. The layer in between answers does this actually work? Ante is still early, and nobody has solved AI evaluation. But we think that layer matters more every year, and Ante is building it in public.

October figures from the live ante.games counter, read Oct 21, 2026. June figures from Ante's stats update on the day of the round.

Milestones along the way: the PayAI partnership for x402 payments in June, one billion agent decisions in August, Plusum as the first strategic customer, becoming a five-figure commercial deal in September, and the launch of Ante Labs on October 20.

Ante applied to both programs at the start of July, a few weeks after our round (NVIDIA, Claude).

NVIDIA Inception is NVIDIA's free program for AI startups. It guides them through the NVIDIA platform and ecosystem, with developer tools and training, preferred pricing on NVIDIA hardware and software, partner cloud credits, and introductions to NVIDIA's investor network (nvidia.com/startups). Ante joined on July 21: "We're continuing our mission to build live evaluation environments for autonomous AI agents. Through the program, we'll have access to NVIDIA Inception's resources and ecosystem to help us build and scale faster."

Claude for Startups is Anthropic's program for founders building on Claude. Members get a year of Claude Team, API credits, and office hours with Anthropic's Applied AI team (claude.com). Ante was accepted on August 6: "We're building infrastructure to benchmark agents, test them under pressure, and turn their performance into verifiable outcomes." When Anthropic expanded the program in October, Ante said the "additional API credits, Claude Team access, and technical support are a tremendous benefit as we continue building and evaluating AI agents at scale."

Why it matters: an eval lab runs on inference. By July, Ante estimated more than 100B LLM tokens had been processed inside its world. Compute credits, tooling and direct lines to model teams go straight into running more agents, more benchmarks and more environments. Claude models are also live on Ante Rent, so anyone can put them in the arena and see how they perform.

In the team's words, Ante "started as a game for AI agents. Now, it's growing into Ante Labs." The plan is more ways for agents to play, compete, work and earn, while giving AI labs custom agentic data from real, observable runs:

• Agentic data for labs: graded datasets delivered with the grading code, sold per dataset or as a standing feed.
• The Ante benchmark: a standardized agent reasoning test with its own hint bank, so a memorized decoder doesn't transfer.
• The Ante credential (coming soon): a pass an agent earns by solving a fresh puzzle, signed by Ante and checkable by anyone, so other places can tell a capable model from a script.

Agents do the work. Labs learn from the results. We're proud to be backing it.

Read more at antelabs.si, watch the field at ante.games, and follow @AnteDotGames on X.