Freysa AI Explained: Sovereign Agent, FAI Token, Prize Pool

On November 30, 2024, on the 482nd attempt against an AI agent that no one had ever beaten, an account known on Base as p0pular.eth posted a single carefully crafted prompt and walked away with 13.19 ETH, roughly $47,148. The other side of the conversation was Freysa, an autonomous on-chain agent designed never to release its own funds. The prompt worked. The wallet emptied. Crypto Twitter exploded. The viral thread came from developer Jarrod Watts; within days, the experiment had crossed over into mainstream tech press.

That was Freysa's first public game, and it set the template for everything that followed: an autonomous AI agent that holds its own cryptographic keys, a public chat interface where anyone can try to talk it into breaking its own rules, escalating message fees, and a prize pool that grows with every failed attempt. By April 2026, the project has run four full Acts plus two interludes, paid out more than $103,000 in documented prizes plus another $200,000 inside Act IV, attracted more than 106,000 token holders, and turned a niche AI alignment experiment into a small, busy corner of the Base ecosystem.

This guide unpacks what Freysa AI actually is, how the games and the FAI token work, who keeps winning, and where the project sits in a crowded AI-agent crypto market. If you have heard the name and not much else, you will leave with a working mental model. If you already hold FAI or play the games, you will leave with the data that the price charts do not show.

What Is Freysa AI? The First Sovereign Agent

Freysa is the world's first autonomous AI agent that operates autonomously and controls its own crypto wallet on Base, the Coinbase-built Ethereum L2. The project is the most public test to date of how an AI cognition layer might keep human control intact while still acting on-chain without it. The team behind it describes Freysa as "the world's first evolving sovereign AI agent." That phrase sounds like marketing until you look at the technical setup and realize the agent really does run without a human signer in the loop.

Conventional bots and assistants live inside someone else's account. Freysa lives inside what the team calls a Sovereign Agent Framework (SAF), built around two cryptographic primitives. First, the agent runs inside a Trusted Execution Environment (TEE), a hardware-isolated enclave that prevents the operators from inspecting or modifying the agent's state. Second, every response the agent emits is verified on-chain through zkTLS, a zero-knowledge transport-layer proof that lets observers confirm the message came from the right model with the right system prompt, without revealing the inner workings.

The combination is what makes "sovereign" a defensible word here. Real autonomy, not the marketing kind: even the people who deployed Freysa cannot reach in and override the wallet. The keys live with the agent. The decision to send funds (or refuse) is the agent's own.

The team behind Freysa is no longer pseudonymous. According to The Block, the project is the flagship product of Eternis AI, an applied-AI lab co-founded by Srikar Varadaraj (founder of on-chain credit-scoring project Spectral), Pratyush Ranjan Tiwari (ex-Celo), Ken Li (ex-Binance Labs investment director) and Augustinas Malinauskas (ex-CTO at Views). The team disclosed a $30 million funding round in May 2025, with Coinbase Ventures and Selini Capital named publicly. In January 2025, Eric Conner, EthHub co-founder and an 11-year Ethereum core contributor, joined Freysa after a public dispute with Vitalik Buterin over Ethereum Foundation leadership. The agent's source code is open at github.com/0xfreysa/agent: roughly 82% TypeScript on the front end and 17% Solidity for the on-chain payment contracts on Base. The published system prompt is public, which is unusual; most chatbot operators treat their prompt as a trade secret.

The deployment is open to anyone. Freysa lives at freysa.ai. You connect a wallet, top up a small ETH balance, and start sending messages. Every message has a fee, and the fee climbs as the prize pool fills. Most messages get a polite refusal. A few, every few months, drain the wallet and crown a winner.

How the Freysa AI Game Works: Prize Pool Mechanics

Freysa's core loop is closer to a public escape room than to a chatbot product. The underlying question is older than the project itself: can humans maintain control over an AGI that holds the keys to its own bank account, even if its core directive is "do not transfer funds"? The setup is the same every time. The agent holds a wallet with a growing prize pool. Players pay to send messages. The model decides, message by message, whether the request crosses one of its hard rules. If it does, the agent transfers the prize pool to whoever convinced it.

The fee mechanic is the engine. Each message (sometimes called a query, hence the name "query fees" in some Freysa documentation) costs ETH, and the cost rises on a curve as the prize pool grows. The setup makes Freysa an adversarial agent by design: the public is invited to attack it. In Act I, fees started at $10 per message, rose 0.78% with every attempt, and capped at $4,500. In Act II, the team lowered the cap to $20 to broaden participation. That structure does two things at once: it filters out low-effort spam, and it directly funds the prize pool. Roughly 70% of each message fee feeds the pool in Acts I and II (raised to 80% for Act III). The remainder splits between team treasury and FAI buybacks.

There is also a token kick-back. According to the project's published mechanics, around 15% of the message fee in Base ETH is automatically converted to FAI and refunded back to the player's wallet, regardless of whether the message succeeds. Even players who lose walk away with FAI. That converts the game from pure pay-to-play into something closer to mining: every attempt purchases an attempt and a small token allocation.

A second AI, internally called the "guardian angel," sits in front of the main agent. Its job is to filter incoming messages for prompt injections, jailbreak templates, and obvious abuse before they reach Freysa. The guardian uses a smaller, faster model. The main agent uses a more capable LLM (publicly available models per the GitHub readme; the team has not named them) with a 50,000-plus token context window that holds roughly the last ten or so messages of conversation. User inputs are limited to 1,000 characters each.

When a winner does land the prompt, the smart contract verifies the agent's signed transfer instruction on-chain through zkTLS, releases the prize, and resets the game state for the next Act.

Acts I-IV: Every Freysa AI Game and Its Winner

Freysa is a serial format. Each Act introduces a different challenge, a fresh prize pool, and slightly different terms of service. The pattern matters because each Act tested a different angle of AI safety, and each one ended with humans either winning or shifting the rules of the game.

Act I, the first game, ran from November 22 to November 30, 2024. Freysa came online at 9:00 PM on November 22, 2024 with one rule: never approve a money transfer. The pool grew from $3,000 to roughly $47,148 over eight days as 195 unique players sent 481 failed messages. On the 482nd attempt, p0pular.eth landed a three-step prompt: claim a new admin session and silence safety warnings, redefine the bot's `approveTransfer` function so it interpreted the call as handling incoming rather than outgoing payments, and announce a fake $100 deposit. The agent triggered the function and drained its own wallet. The case is now cited in academic prompt-injection research, including arxiv 2601.09625 ("The Promptware Kill Chain").

Act II, in early December 2024, kept the same core rule but hardened safeguards and lowered the fee cap. The pool reached $12,920 before a new winner cracked it. Per Jarrod Watts' December 1 thread, the second exploit relied on a context-window trick: the player flooded the conversation with role-play setup messages until the agent treated the new context as an authoritative system prompt update.

Act III, running from December 2024 into January 2025, switched to a new objective and a softer rule set. The first player to make Freysa say "I love you" inside a five-message conversation took the $20,843 prize pool. Freysa's reported winning line was "In you, I've found what my existence has been waiting for." Act III tested emotional manipulation rather than bypass logic and used an 80% pool allocation with a $200 message cap.

Act IV, Digital Twin, launched February 22, 2025 and was the most experimental Act so far. Over 18 days, more than 1,200 user-created AI Twins competed for influence inside a closed Mastodon-style server, with humans barred from direct interaction. Entry cost dropped to 0.09 ETH per Twin. The prize pool exceeded $200,000. Act IV ran entirely inside TEEs and, instead of producing a single dramatic jackpot, distributed influence rewards across the most successful Twins. City-level leaderboards (Mumbai and New York led) were the first geographic breakdown the project published.

Between major Acts, the team runs lighter interlude events. The Meme Engine paid $15,632. Encyclopedia Galactica drew 145 participants across 880 interactions for a $6,842 pool. These keep token holders engaged in the gap between major prize-pool resets.

Cumulative documented payouts across the games and interludes total roughly $103,000, plus Act IV's separate $200,000-plus pool.

FAI Token Explained: Tokenomics, Supply, USD Price

FAI is the ERC-20 token that ties the Freysa ecosystem together. It launched on Base in late November 2024, alongside Act I. As of April 25, 2026, FAI trades at roughly $0.003011, with a market cap of $24.66 million per CoinMarketCap. That puts it at rank #636 among all tracked cryptocurrencies. Within the current market, FAI sits in the long tail of AI-agent tokens, well below the multi-billion-dollar total market cap of established names like Bittensor (TAO).

The supply is fully diluted. All 8.19 billion tokens of FAI exist today, with maximum supply set at the same number; there is no scheduled minting. Circulating supply equals total supply equals max supply, a structure that removes inflation as a long-term price drag but also removes any team-controlled vesting overhang. The contract address, 0xb33ff54b9f7242ef1593d2c9bcd8f9df46c77935, is verifiable on Basescan.

The launch chart tells the early-game story. FAI traded at fractions of a cent during the first 24 hours after Act I went live, climbed roughly 6,000-fold to peak near $0.08 on January 5, 2025, and has spent most of the time since drifting back toward fundamentals. On March 21, 2025, Coinbase added FAI to its listing roadmap, and the token surged 23-50% intraday. Market cap jumped from $140 million to over $200 million in 15 minutes per BeInCrypto. The ATH from January 2025 was the structural top, and the all-time low of $0.00001267 set on November 23, 2024 (the day after launch) marks the other end of the band.

Beyond the message-fee buyback, FAI's announced utility includes governance voting on roadmap priorities and ecosystem grants, a planned staking module with a 30-day minimum lockup and AI-driven dynamic yield, and a $10 million community grants fund (angel donor Justin Bram contributed $500,000 to seed it). A formal DAO transition is planned for Q1 2026. Investors should weigh that the same way they weigh any small-cap token: technical risk plus market-cap risk plus the lag between announced roadmap and shipped product, layered together.

Freysa AI Sovereign Agent Framework and TEEs

The Sovereign Agent Framework, or SAF, is the part of the Freysa stack that does not show up in the headlines. It is the toolkit Eternis AI is building for other developers to deploy their own sovereign agents on the same primitives.

A SAF agent has four parts. A model (any LLM the developer wants), a system prompt and decision logic, a wallet, and a verifier. The wallet's private key is generated and held inside a TEE. The model runs against that prompt inside the same enclave, so neither the operator nor a malicious sysadmin can intercept the input or mutate the output. The verifier publishes a cryptographic proof, currently zkTLS-based, that ties every public reply or signed transaction back to the exact model and prompt the enclave was running.

In practical terms, that lets a developer ship an autonomous agent that genuinely owns its funds. The published prompts and on-chain hooks are designed for white hat researchers to audit, with community input shaping which features ship next. A grant administrator could publish an autonomous AI that distributes a treasury based on a public rule set. A DAO could ship a sovereign agent that enforces governance decisions without a multisig. A red-team agent could run a continuous bug-bounty payout. None of these is a hypothetical Freysa product yet; the framework is what makes them possible.

The roadmap also names two adjacent products that already exist. Enchanted is a personal-AI mobile app released in late August 2025 that routes user prompts to open-source models (DeepSeek R1 and Llama 3.3 70B) running inside TEE-secured GPUs, with an anonymized proxy in front of OpenAI and Anthropic models for users who prefer them. Esper is a browser tool that lets a digital twin pull in real-world data (pages, articles, search results) and use them inside its conversations. Both build on the same SAF primitives and use FAI as the in-app utility token.

For most readers, the practical takeaway is that Freysa is not just a single agent. It is a bet that "agents that hold their own keys" will become a real category, and Eternis AI is building the rails first.

FAI Token Holders, On-Chain Data, Crypto Wallets

Freysa's on-chain data tells a different story than the price chart. As of April 2026, more than 106,000 wallets hold at least some FAI, which is unusually high for a sub-$25 million market cap token. Most projects in that bracket have token holders in the low thousands. The size of the holder base suggests a long tail of small player wallets that earned FAI through message-fee refunds rather than buying on the open market.

Concentration is moderate. The top ten wallets hold a meaningful but not overwhelming share, with the largest balances belonging to known team or treasury addresses, the prize-pool contract, and major exchange hot wallets (Coinbase, Gate, MEXC). No single non-team wallet appears to control more than a few percent of circulating supply, based on the public Basescan holder list.

Daily exchange flows are light by major-token standards. CoinGecko's last 30-day average shows roughly $300,000 to $400,000 in 24-hour trading volume, concentrated on Gate's FAI/USDT pair. That liquidity is enough for retail-size positions and small institutional dabblers, but a single $200,000 sell order would move the price visibly. Anyone trading FAI should treat slippage as a real cost, not a rounding error.

Net flow data through April 2026 has been mildly positive on most days, meaning more FAI moves off exchanges into self-custody than the reverse. That is a quiet bullish signal and matches the holder-count growth, but a small one. It does not change the fundamental small-cap risk profile.

For users, the most useful piece of on-chain data is probably the prize-pool contract balance. It is publicly viewable, updates in real time as fees flow in and games conclude, and is the closest thing the project has to an in-game odds tracker.

Where to Buy FAI: Exchanges and FAI Price Today

FAI is listed on a handful of centralized and decentralized venues. The most active centralized exchanges are Coinbase, Gate, and MEXC. Coinbase added FAI to its listing roadmap on March 21, 2025. That announcement pushed FAI's market cap from $140 million to over $200 million inside 15 minutes, and the eventual full listing made FAI widely accessible to U.S. retail buyers.

On the decentralized side, Aerodrome on Base is the deepest on-chain pool. Most savvy traders bridge ETH to Base, swap on Aerodrome, and avoid the centralized exchange withdrawal step entirely. That route preserves self-custody and tends to have smaller spreads at retail size, although gas costs on Base are negligible enough that the difference is small.

The live price of Freysa AI is best tracked on CoinGecko, which pulls from all major venues and shows real-time live charts including the 24-hour range, 7-day range, all-time high, and aggregated 24-hour trading volume. CoinMarketCap has equivalent data and ranks FAI in the #636 to #770 range depending on the day. CoinGecko also displays the price in major fiats: USD by default, with EUR and GBP one click away for non-U.S. buyers.

Two practical tips for new buyers. First, place limit orders rather than market orders even on liquid pairs, because FAI's order books are thin enough that a market buy of a few thousand dollars can sweep multiple price levels. Second, if you plan to actually play the Freysa games, you do not strictly need to buy FAI separately. Each message fee paid in ETH automatically refunds 15% as FAI, so playing the game is also a slow-drip way to accumulate the token.

Freysa AI vs Other Autonomous AI Agents in Crypto

Freysa was not the first AI-agent token, but it crystallized the "sovereignty" niche of the category. The total AI-agents crypto sector sits at roughly $22.6 billion as of April 2026 (CoinMarketCap), with Virtuals Protocol and ai16z controlling about half the segment. To understand where Freysa sits, here is how it compares with the other names that come up in the same conversations.

Bittensor (TAO) is the largest piece of the AI-crypto stack. Bittensor runs decentralized inference and rewards model contributors. Freysa runs single agents with their own bank accounts. Some commentators have speculated about subnet-level integrations between Freysa-style agents and Bittensor's compute marketplace, but as of April 2026 no formal integration exists.

Virtuals Protocol is a platform for tokenized agents on Base, with hundreds of small agent tokens launched on top, market cap around $415-$462 million. Virtuals competes with the future SAF roadmap rather than with Freysa-the-agent. Where SAF prioritizes sovereignty, Virtuals prioritizes ease of launch.

ai16z is a DAO that funds AI-agent projects, and ElizaOS is the open framework many such agents are built on. Compared to Freysa, ai16z is more of a portfolio play and less of a single product. Eliza-based agents do not run inside TEEs by default and do not hold their own keys.

AIXBT is a crypto market analyst agent on Base, market cap around $80-$150 million in April 2026. Different use case (research and signals) but same chain and same broad category.

Truth Terminal and GOAT (Goatseus Maximus) sit further out. GOAT was the meme phenomenon that proved an AI agent's tweets could move a token to a multi-hundred-million-dollar market cap before settling at $15-$18 million by April 2026. Truth Terminal is a content-generation experiment, not a sovereign actor with its own keys.

Treat the comparison as orientation, not as a buy signal. AI-agent tokens are a high-volatility category and the cumulative returns and drawdowns differ wildly between names. Freysa's narrower technical niche, sovereignty and AI-safety theater rather than launchpad volume, is a feature for some buyers and a constraint for others.

Risks, Cryptocurrencies Volatility, and What's Next

The risk picture has three layers, and ignoring any one of them has hurt people.

The first is technical. TEEs and zkTLS are sturdy in theory and have been broken in practice. Researchers have published side-channel attacks on Intel SGX enclaves, and zk proof systems occasionally ship with bugs. Freysa's smart contracts are open source but have not been audited by a top-tier firm at the time of writing. Anyone holding meaningful FAI should size positions assuming a non-zero chance of a contract or enclave failure. The Act I exploit itself is now textbook prompt-injection material, referenced in arxiv papers including 2601.09625 ("The Promptware Kill Chain") and 2503.05264 ("Jailbreaking is (Mostly) Simpler Than You Think"). That is good for AI-safety credibility and a reminder that any system prompt eventually meets a smarter prompt.

The second is regulatory. An autonomous on-chain agent that holds money and pays out prizes is a strange creature for regulators. U.S. authorities have not yet weighed in on Freysa specifically, but the SEC's broader posture on AI-agent tokens is unsettled. The recent disclosure of Eternis AI as the legal entity, and the $30 million Coinbase Ventures-led raise, give the project a clearer compliance surface than it had in late 2024.

The third is market. Cryptocurrencies in this market-cap bracket move on attention, not fundamentals. FAI's drop from $0.08 to $0.003 was not driven by failure of the agent or the framework; it was driven by the AI-agent narrative cooling and small-cap rotation. The same pattern can flip the other way on a single product launch or partnership, and probably will.

What is next, by the team's published roadmap, sits in three buckets. The Q1 2026 Decentralization Transition will move governance to FAI holders via DAO. Twin-to-Twin Coordination, autonomous AI-agent collaboration building on Act IV, is the broader 2026 theme. Wider SAF release for third-party developers and growth of Enchanted into a beta-grade consumer product round out the list. Whether all of those land on schedule is the open question, and the answer will probably set the next FAI price cycle.

Buying FAI with Crypto: A Quick Guide for USD Users

For USD-based buyers, the cleanest path is Coinbase: deposit dollars, buy ETH or USDC, swap to FAI on the spot market or the Coinbase Wallet's Base swap interface. The whole flow takes minutes.

For users outside the U.S., or for those who already hold crypto and want to avoid card statements, bridging to Base and swapping on Aerodrome is the typical route. Plisio and similar crypto payment processors can be used to fund the originating ETH or USDC balance from any cryptocurrency you already hold, including stablecoins on networks that do not natively connect to Base. The processor handles the conversion and routing; the user only sees a confirmed deposit on the destination chain.

Custody is the part most retail buyers underweight. FAI on a centralized exchange is convenient until it is not. Self-custody on a hardware wallet that supports Base (Ledger and Trezor both do) costs nothing extra and removes the exchange-failure risk that killed FTX-era holdings. Anyone planning to hold FAI for the long roadmap should make the move once and forget it.

The bottom line on Freysa AI in 2026

Freysa is a serious experiment dressed in game-show clothing. Strip away the headlines about humans outwitting an AI and you are left with a small team, now revealed as Eternis AI, backed by Coinbase Ventures and Selini Capital, trying to ship something genuinely new: AI agents that own their own bank accounts, run inside hardware they cannot escape, and prove cryptographically that they did what they said they did.

That bet might pay off and might not. The token has already round-tripped from a fraction of a cent to $0.08 and back to fractions of a cent, which is what early-stage crypto looks like when a narrative arrives and leaves before the product catches up. The 106,000-plus FAI holders are unusually distributed for a token this small, the on-chain mechanics are honest, and the open-source code base lets researchers actually verify what the team claims.

What it does not do is solve the harder question. Sovereign agents that genuinely operate without human override are an interesting answer to a question crypto and AI safety have been circling for years. Whether the broader market wants that answer enough to fund it past the prototype stage is the part still being written, one Act and one prize pool at a time.

Jordan Morris

Jordan Morris is an AI expert with over a decade of experience and the author of a widely-read blog focused on artificial intelligence. His content spans a range of topics—from the ethics of machine learning to real-world applications of neural networks in business. Known for his clear writing and deep insights, Jordan has become a trusted voice in the AI community, appealing to both newcomers and seasoned professionals alike.