Anza is targeting about 150 milliseconds to finality, down from the roughly 12.8 seconds it attributes to TowerBFT, but Firedancer and Frankendancer cannot run the code yet. No mainnet date has been announced, and September 28 is a feature-gate processing date rather than a confirmed Alpenglow launch.
Anza began moving Alpenglow, the rewrite of Solana's consensus rules, onto the network's public testnet on Wednesday. The upgrade runs only on Agave, the validator client Anza maintains, and the network's two alternative clients cannot take part in the test.
Alpenglow's first component is Votor, which replaces TowerBFT. Under TowerBFT a block becomes irreversible once validators have accumulated votes on chain across 32 slots. Votor has validators send those votes to each other directly and finalize a block after one or two rounds. Anza is targeting roughly 150 milliseconds to finality against the approximately 12.8 seconds it attributes to TowerBFT, and crypto.news reports simulations reaching 100 milliseconds in favorable conditions. Those are design targets. No measured finality figure from the testnet has been published, and the point of the test is to produce one.
The software carrying the change is Agave v4.3, tagged on September 18 and described in its release notes as stable for testnet, devnet and mainnet-beta. It reached Solana's mainnet that day with the Alpenglow code inactive. Anza had staged the migration by asking validators controlling 10 percent of stake to upgrade by September 8 and 25 percent by September 14, then called v4.3 the recommended version for every validator at 15:37 UTC on September 21.
Client support is the thinnest part of the rollout. Firedancer and Frankendancer, the alternative validator programs meant to reduce Solana's reliance on any single codebase, do not yet support Alpenglow, so the initial testnet migration runs entirely through Agave. Client diversity exists so that a fault in one implementation does not stop the chain. For the duration of this test, the network's newest consensus code has none of it, and every validator taking part is running the same software from the same team that wrote the upgrade.
Rotor, the block propagation half of the Alpenglow design, is not part of this activation. What is being tested is the voting layer alone.
The change that reaches outside the validator set is what happens to vote transactions. They stop appearing in blocks, because validators no longer write them there. Anza's upgrade guidance puts vote transactions at about 59 percent of all transactions on the network, which means transaction counts, block contents and anything downstream that reads them will change shape when Alpenglow activates. Providers that stream blocks, track votes or depend on Solana's existing commitment levels have to adjust before that happens rather than after.
Anza has not announced a mainnet activation date. September 28 appears on its schedule as the day mainnet resumes processing queued feature gates, and both crypto.news and Solana Compass state directly that this is not a confirmed date for Alpenglow itself; CoinDesk described it as tentative and unconfirmed. The published order is testnet first, then devnet, then mainnet-beta after an observation period, with each step waiting on the one before it.
Alpenglow first ran on a community test cluster in May, four months before the code reached a network anyone else depends on. The public testnet is the first environment where it meets the validator set, the RPC layer and the tooling at something close to production scale.
The finality numbers deserve one qualification. The 12.8-second baseline describes 32 slots at the 400-millisecond slot target Solana ran on when that figure was set. The network's slot target fell to 250 milliseconds on September 18 under a separate staged schedule, and none of the current reporting carries a revised TowerBFT figure for the shorter slots. Slot time and finality are different schedules moving at the same time, and the headline comparison is drawn against the older one.
Solana has several protocol changes in flight at once, including a staged reduction in what the network charges for state that runs through its own feature gates. Alpenglow is the largest of them, and the testnet phase is where the cost of running it on one client becomes measurable rather than theoretical.
