September 4, 2026 · Published daily
Subnet News - September 4, 2026
Trishool’s OpenAI cyber-program acceptance leads the day with a roughly 20% α/TAO move. Forever Money and Chainlink brought native TAO to Robinhood Chain, Gittensor opened live compute serving, and fresh AlphaGap signals expanded today’s ranked subnet coverage.
- Coverage window
- Sep 3, 2026, 10:00 AM – 2:37 PM ET
- Subnets in coverage roster
- 128
- Subnets with updates
- 46
Biggest stories first
Today's ranked news
Trishool says it joined OpenAI Trusted Access for Cyber
Trishool announced acceptance into OpenAI’s vetted Trusted Access for Cyber program and described it as a step in commercial growth. The acceptance is reported by Trishool; no separate program listing was located during this scan.
+17.3% α/TAO 24H price action as of Sep 4, 2026, 2:27 PM ET
Forever Money takes native TAO to Robinhood Chain with Chainlink CCIP
Forever Money announced that the Bittensor ecosystem has arrived on Robinhood Chain. Chainlink confirmed that its CCIP security powers the expansion of 1:1 native TAO, opening cross-chain access to Robinhood’s retail-focused L2.
Gittensor compute serving goes live as a second miner earning track
Gittensor now documents compute serving as a live companion to open-source contributions: miners can serve the blessed model on GPUs, while validators audit outputs and expose an OpenAI-compatible gateway. The official miner and compute-serving documentation confirms the operating path.
Subtensor release 454 lands with root-claim and proxy-transfer fixes
The core repository merged release 454 alongside fixes for coldkey-wide root claims, stale staking relationships and contract-proxied stake transfers. Runtime activation was not independently confirmed at publication.
Allways exposes timeout-refund facts in swap status
Allways now persists timeout refund details and returns them from its by-key status endpoint. AlphaGap also flagged a wallet-control security fix and direct miner exchange-rate quotes, expanding the update beyond refund visibility.
Vanta isolates position files by miner hotkey
Vanta fixed a multi-wallet collision where position writers could overwrite one another or race on the same temporary file. Position and temporary filenames are now scoped per hotkey.
Pareton serializes round writes and fixes streamed verdict handling
Pareton added row locking so void, live-status and settlement writes share one order, rejected forged candidate beacons, and stopped pollers from rewriting voided entries. These changes target reported infrastructure failures and inconsistent round state.
Cacheon replaces its champion floor with decaying accepted-crown credit
Cacheon merged a large runtime update that seeds more cache families, preserves completed passes and changes rewards from a single champion floor to decaying credit across accepted crowns. The merge also corrects dashboard emission units from TAO to subnet alpha.
Zeus ships a forecast-quality penalty system and follow-up hotfix
AlphaGap reports that Zeus released an SSRD penalty mechanism for low-quality forecasts and deployed a hotfix after rollout. Emissions and TAO flow weakened during the transition; the implementation was not independently exercised.
AdTAO publishes registration and validator-burn status
AdTAO added per-hotkey registration status and burn metadata feeds, fixed direct daily-vector commits when its weekly stream is disabled, and prevented an empty document from replacing a performance report.
SayGM launches GPT-6 Astra through its gateway
SayGM announced GPT-6 Astra availability with browser, coding and terminal capabilities, routed through a verifiable TEE and advertised below list pricing. Availability and pricing are the project’s official claims.
Epago reports its research-ranking testnet is live
Epago says its testnet is running with Tongyi-DeepResearch-30B-A3B as the genesis model and a challenge process for ranking researchers. Mainnet specifications remain under development; the launch report could not be independently exercised.
Score investigates download delays and missing incentives
Score’s team attributes Cricket Challenge download delays and resulting zero scores to server overload, and is investigating weight-setting discrepancies. Capacity expansion and a configuration-reading fix are reported as in progress.
Feval opens miner model submissions and starts validator emissions
Feval reports that model submissions are open, validator emissions started at 10%, and the first 10,000-row evaluation window will run for two days before a planned reduction. The public repository corroborates active dataset and winner-take-all work, while operational figures remain project-reported.
Synth adds a fourth forecasting competition and removes a broken price feed
Synth added an Ultra forecasting track as a fourth reward input and removed a Pyth path that was causing scoring problems. Project documentation confirms the competition structure and migration away from Pyth; activation details remain conditional on the score provider.
lium.io fixes GPU validation and clears a reported 2 TB storage leak
AlphaGap reports that lium.io fixed GPU validation and version-enforcement problems and cleared a recurring 2 TB storage leak. The feed also showed strong 24-hour TAO inflow; primary implementation links were not exposed in the accessible card.
Ridges launches a PostgreSQL and ClickHouse engineering competition
AlphaGap reports that Ridges is expanding beyond general coding benchmarks with production-style PostgreSQL and ClickHouse tasks, including fixed query budgets and time limits. The launch is currently supported by the signed-in intelligence feed.
Enigma raises its challenge to RSA-1024 with a $250,000 prize
The official Qbittensor Labs account announced a $250,000 prize for the first successful RSA-1024 break. This verifies the announced terms, not the solvability or funding custody of the prize.
TPN fixes false precheck bans and scores its second competition
TPN corrected a precheck that hashed leftover test files, updated its container and unbanned affected hotkeys. The team also announced 21 completed attempts in TPN-002; that attempt count is project-reported.
Conjectures opens paid public proof submissions
Conjectures published a public contribution endpoint where users can submit Lean 4 proofs and receive bounties for accepted work. AlphaGap reported 13 live verified submissions and weaker emissions during the rollout.
Leadpoet fixes validator recovery and opens daily source competitions
Leadpoet hardened validator retry and registration recovery, fixed timeout paths affecting scoring, and activated daily competition tooling for miners to submit new research sources. Repository changes verify the reliability work; the competition rollout is also reported by AlphaGap.
Gittensor adds image input and neutralizes reference-side 429s
SparkInfer merged vision image input and fixed model reporting. Its validator also now treats a reference-provider HTTP 429 as retryable backpressure and then neutral, rather than a miner miss.
Hippius fixes storage billing while a bridge issue delays transfers
AlphaGap reports that Hippius corrected storage-emission calculations and added customer-plan controls, while the team works through a post-upgrade bridge issue awaiting core approval. Hippius’s official materials confirm the bridge is part of its live on-chain payment architecture, but the current incident is feed-reported.
Vocence confirms its pivot to UMI as pilot evidence lands
The former Vocence account said the project has pivoted to UMI and will no longer use that account for subnet updates. UMI’s repository published replayable component-pilot evidence and an immutable media location in the same window.
Cascade activates head-to-head miner scoring as emissions rise
AlphaGap reports that Cascade deployed block-height-activated head-to-head scoring rounds intended to reduce gaming. The feed also recorded a strong emissions increase alongside net TAO outflow, so market and validator signals were mixed.
MicroTensor isolates its scorer and adds parallel artifact evaluation
MicroTensor moved module scoring behind an in-process defense boundary, added optional parallel evaluation for leased artifacts, and anchored reference-cost and observation requirements.
Verathos repairs protocol-v3 mesh scoring and GPU capacity classes
Verathos qualified mesh scoring on protocol v3 and normalized VRAM classes used in capacity audits, including a separate Blackwell calibration fix.
Verathos reports Sleipnir is live on mainnet
A dated team brief reports that Sleipnir is live for distributed inference across global participants. The public repository confirms continuing mesh work, but this review did not independently run a mainnet inference request.
Cortex commits a 20/80 Bounty and Proof allocation
Cortex committed a trust-root allocation of 20% to Bounty and 80% to Proof, with winner-take-all or discovery payout modes. A follow-up pins a real proof-evaluation image; runtime adoption and payouts were not independently replayed.
Capcomp fixes attribution for revealed on-chain recipes
Capcomp now attributes an opened commitment by its reveal block, treats a commitment for another run separately from a refusal, and records the admitted run explicitly. The project reports these fixes restored admission of Run 423’s commitments; that live count remains project-reported.
Beam adds explicit room-transfer capability checks
Beam’s public core now exposes a room-transfer capability helper and refines ordinary transfer eligibility with explicit blocked reasons and fallback behavior. The team separately previewed a tinygrad tensor transport, which remains upcoming.
Minos publishes a round-verification script
Minos added documentation and a script that retrieves revealed round files, checks all six expected files and verifies published SHA-256 digests. The team’s reported reduction of miner burn to 80% remains an operational claim.
Ditto requires private calibration evidence for coding data generation
Ditto added a private calibration-evidence requirement and bound private memory bundles in its coding data-generation path. The repository confirms the new gates; a public broad rollout was not established.
Swarm publishes five replayable drone-interceptor evaluation episodes
Swarm reports five successful indoor target-drone captures in evaluation episodes without GPS or compass, with the seed displayed for replay. The accessible post does not establish physical-world testing, and the result was not independently reproduced.
DSperse releases 14.14.2 with a reconnecting chain transport
DSperse merged release 14.14.2, including a WebSocket transport that reconnects and sends keepalive pings. The repository confirms the release and implementation; individual validator uptake was not measured.
Teutonic adds batched evaluation and model-digest visibility
Teutonic added batched evaluation and now displays model digests for queued and active evaluations. These are verified repository changes, with no independent measurement of production throughput yet.
Vanta reports more than $603,000 in trader rewards
Vanta says its public ledger records $603,594 paid since April and previewed $1 million Pro accounts for traders who pass its evaluation. The totals and future program are project claims from the official account.
Blockmachine reports more than one billion Avalanche requests
Blockmachine says Avalanche has been live on its network for two weeks and has already produced more than one billion requests. This is an official project metric that was not independently reproduced.
UR freezes a corrected testnet release candidate
UR refreshed its testnet release lock, authenticated historical runtime evidence and froze a corrected release candidate. This is a testnet release milestone rather than proof of mainnet deployment.
Perturb reports a validator scoring fix is coming
A dated team brief acknowledges intermittent zero scores from a validator bug and says the next release will address it. A larger adversarial-training upgrade is also planned, without a firm launch date.
Endure defines the context axis for Alpha Risk V1
Endure says each assessment will be identified by target, context and schema, with risk.v1 serving as the first vertical’s context. The announcement defines the planned assessment contract rather than a completed public launch.
Engy adds reasoning-token usage accounting
Engy merged support for reasoning-token usage, extending its accounting beyond ordinary token totals. The repository verifies the change; downstream billing behavior was not independently exercised.
RedTeam tightens detection logic as demand and volume rise
AlphaGap reports new detection-logic changes for live security challenges alongside elevated subnet volume and TAO inflow. The code change is material to scoring, but the exact deployment commit was not independently located.
Harnyx binds verified miner endpoints to finalized membership
Harnyx changed endpoint registration to check finalized subnet membership before accepting a miner endpoint, then bumped its validator release.
OpenRoboto reports its real-robot validation workstation is running
OpenRoboto says its xArm 6 workstation is operational and will be used to validate the simulation-track champion on physical hardware. The promised demonstration has not yet been published.
Vidaio demonstrates a new one-pass restoration model
Vidaio posted an Elvis interview restored and upscaled from SD to FHD in one pass using a new generative model. The demonstration is project-published and was not independently compared with the source footage.
Investing delays its second quant-fund round
Investing says it is postponing Round 2 because current bearish and sideways conditions are unsuitable for new capital. The team has not supplied a replacement launch date in the accessible brief.
InfiniteQuant slows closer submissions to one per minute
InfiniteQuant raised the minimum interval between closer submissions from 10 to 60 seconds and documented a limit of 30 per day.
Albedo fixes registration re-keying after miner reactivation
Albedo now changes a registration key only when a newer activation carries a new submission key, preventing stale activation data from re-keying the current registration.
Connito begins archiving round podiums to Hugging Face
Connito added validator publication of each round podium to a Hugging Face archive repository and names archived folders by cycle.
SOMA documents its GitHub Copilot compressor
SOMA refreshed its public README for the context-compression layer. Public announcements say it is available for GitHub Copilot, while reported token savings remain project measurements.
Audit trail
Coverage notes
Searched
- All 128 active subnet slots and current names from the finalized-chain screener
- 50 dated September 4 subnet briefs and their available source links
- 120 public subnet and ecosystem GitHub repositories from September 3 at 10:00 AM ET through publication
- Official subnet X posts surfaced in the public all-project feed, with original post URLs and timestamps; direct sampled retrieval was blocked
- Shizzy Unchained YouTube Videos and Live tabs
- Original Forever Money, Chainlink and Trishool X posts supplied after the first publication, verified through X’s official embed endpoint
- Signed-in AlphaGap 48-hour feed, notifications, leaderboard, price, flow, emissions and development signals
Public-data gaps
- Correction: the first September 4 edition missed the Forever Money and Chainlink announcements. Both original posts were added after publication and the report was reranked.
- X required phone or password verification in the connected browser. A Grok guest query returned a sign-up requirement and no research results, so no Grok findings are claimed.
- Public X search indexing is incomplete; unavailable, deleted, private and unindexed posts may not appear in the all-project feed.
- Five mapped repositories were missing or inaccessible, and organization-only links could not be treated as complete repository inventories.
- Discord-only reports were marked developing unless corroborated by public code or an official public post.
- No independently verified forward alpha-unlock calendar was available, so upcoming unlocks are left empty.
- Direct X post retrieval returned HTTP 403 for sampled posts. X-only stories rely on the public project feed and are marked developing; original post links are supplied for review.