Dates and verification labels belong to the original edition. A past verified announcement can now be outdated. Historical project names are retained; an SN filter searches the slot across registrations. All keywords must match the story text or subnet label. Ratings describe editorial priority, not investment quality.
Subnet Summer launches a community-directed validator
Subnet Summer opens delegation to a validator built with Medulla Labs. Its operating model lets delegators weigh in on governance and root-yield allocation, with a published response owed by a precommitted date. The detailed announcement explicitly says the operator can decline a vote with reasons: votes do not cryptographically force a transaction. The new element is a public participation and accountability process around validator decisions.
A live validator introduces a concrete model for delegator input into capital allocation.
AlphaChaser announces version 2.0 of its TAO subnet trading agent
Gordon Frayne announces AlphaChaser 2.0 following feedback from its initial users. The product is presented as an autonomous TAO subnet trading agent. The announcement does not establish independently verified returns, and this edition does not invent a feature-by-feature changelog. Disclosure: AlphaChaser is a Shizzy Unchained partner.
A fresh release directly serves TAO subnet users, with performance claims kept separate.
SN112 THOR: Pylon signs acquisition agreement for a new robotics-data subnet
Pylon Partners says it signed a definitive agreement on September 22 to acquire SN112 and launch THOR Group. Its first proposed business is a marketplace where miners annotate robotics footage and validators check output against customer requirements. The broader plan includes open hardware and manufacturing/distribution relationships in China. The full announcement targets the first incentive mechanism for Q4 2026. This is a major proposed change in ownership and purpose, not a live mining launch or independent confirmation that the on-chain transfer has completed. At 9:58 AM ET, the slot’s alpha was up 6.3% against TAO over 24 hours.
A signed acquisition agreement and a new physical-AI business direction deserve lead coverage, with execution and launch timing clearly separated.
SN44 Score powers 15 newly connected stations in Manako’s latest rollout
SN44 Score’s commercial deployment has a fresh progress report: Manako says it brought 15 stations online around Lyon over the weekend, using their existing camera feeds. Co-founder Max Sebti confirms the vision agents use SN44 Score models. The September 23 update advances the earlier report of 12 stations; 20 was the trip’s target and roughly 400 the wider rollout ambition, not a completed installation count. The important development is subnet-built computer vision operating at multiple real sites. Deployment speed and installation totals remain team-reported, and this update does not disclose revenue.
A concrete multi-site business deployment is stronger evidence of utility than a product teaser or roadmap.
SN66 Conjectures publishes its claimed solution to Erdős Problem 1062(ii)
SN66 Conjectures says its miners solved a problem open since at least 1994: proving that the limiting density of the largest subsets of {1, …, n}, where no element divides two distinct others, is irrational. The team links a paper and a public Lean proof file and reports verification through its system. This is a different result from yesterday’s Erdős 859 announcement. The artifacts are available to inspect; this publication has not rerun the proof checker or established independent mathematical acceptance.
A claimed resolution of another long-standing mathematical question merits top billing, with proof availability distinguished from independent validation.
SN114 SOMA and SN53 Engy put context compression into a live inference workflow
SN114 SOMA says its integration with SN53 Engy is live, and SN53 Engy independently confirms it is powering the service. The workflow compresses the input context through SN114 SOMA before SN53 Engy runs the model, aiming to lower inference costs. This is a concrete integration between two subnet products. No independently measured savings percentage or customer-revenue figure is disclosed. The September 23 channel brief also points to the compressor dashboard and says subscription/API business details will follow.
Both participating projects confirm a live product integration with a clear practical use.
SN6 Numinous ships an equity-impact dashboard and Numinous-1 forecasting access
SN6 Numinous has released an impact interface connecting equities to macroeconomic and geopolitical factors tracked by prediction markets. Separately, release 3.1.3 adds the Numinous-1 forecasting-model endpoint for miner agents and a Jev decision endpoint in the gateway. These are two concrete additions: a customer-facing view of event exposure and new forecasting tools for miners. The released gateway integration is not evidence that model weights are open or that forecasting accuracy has been independently established. Discovering further causal-graph relationships remains a stated next step.
A public product release and inspectable miner-model integration materially expand the subnet’s workflow.
SN15 ORO launches ORO Bench with fresh shopping environments
SN15 ORO has publicly launched ORO Bench, now powering its shopping-agent competition. Instead of relying on a fixed benchmark that agents can memorize, it generates shopping environments across seven task families, including budget constraints, preferences and recovery from unavailable items. Public qualifying tasks and hidden race tasks are separate, and each evaluation is tied to a fixed version of its environment. The practical goal is to reward agents that handle unfamiliar shopping situations. Resistance to gaming is the design objective, not a guarantee that exploitation is impossible.
A new active evaluation system changes what the subnet rewards and how its agents demonstrate useful capability.
SN78 UMI begins technical collaboration with Convo on distribution and integration
SN78 UMI says its team and Convo.io have begun practical work on data structure, validation, delivery requirements and model integration. Convo’s interpreter network gives the initiative a potential route toward real-world Deaf communication services. The announcement describes an early collaboration, not a completed production integration, signed revenue contract or disclosed customer rollout. Counterparty confirmation was not independently located during this scan.
A named collaboration could connect subnet capability with an established distribution channel, while commercial execution remains developing.
SN78 UMI reports provisional challengers ahead of its reference model
SN78 UMI reports cohort-three scores of 43.3% for its leading provisional challenger versus 35.1% for the reference model: an 8.2 percentage-point gain, or roughly 23% relative improvement. The team explicitly calls these small, pre-calibration test windows rather than production rankings. Its live competition endpoint checked today still reports rewards_active=false and chain_submission_authorized=false. The team is finalizing a policy to renew certified weights until a newer certified result is ready; that plan should not be mistaken for confirmed reward activation.
New measured competition results matter, provided provisional scoring and reward activation remain distinct.
SN28 SayGM adds Opus 5.5 and GPT-6 Sol/Luna to its model offering
SN28 SayGM announces Claude Opus 5.5 with an 18.2% discount and GPT-6 Sol/Luna with a 12.9% discount against their respective list prices at launch. The provider presents the models as live through its gateway. This expands the available model catalog; the percentages are SN28 SayGM’s published pricing claims at announcement time, not a continuously checked cheapest-provider ranking. This update is separate from yesterday’s image-comparison launch.
New model availability and specific published prices are useful product changes for inference customers.
SN8 Vanta fixes carry-forward of unpaid miner rewards
SN8 Vanta’s merged payout fix carries previously unpaid amounts into the next weekly calculation and corrects how frozen ledgers and partial checkpoints are handled. This addresses reward accounting, not a claim that every affected payment has already settled. Miners should distinguish the released correction from completed wallet receipts.
A payout-accounting correction has direct consequences for miners.
SN10 Pareton discusses an inference-provider pilot and recurring-revenue plans
The TAO Daily’s September 23 account of Xavier Lu’s TAO Pill interview reports an active pilot with an unnamed inference provider. SN10 Pareton wants to sell ongoing optimization of customers’ model-serving infrastructure. The provider may be named if the pilot progresses. Proposed returns to alpha holders remain under legal and accounting review; no finalized buyback structure or paying contract is established by this report.
A reported external-provider pilot is a meaningful commercial lead, with revenue and token-value plans kept provisional.
SN19 Blockmachine changes archive base rewards from September 24
SN19 Blockmachine’s updated documentation schedules archive base payments for the five highest-traffic providers per chain, allocated proportionally among them, from the September 24 epoch. Providers outside that group can still earn for served requests and recent-block service. This concerns one base-reward component, not elimination of all rewards outside the top five.
A dated reward-allocation change materially affects archive operators.
SN23 Trishool releases guard deobfuscation and audit improvements
SN23 Trishool’s v2.0.30 changes improve guard handling of obfuscated inputs and audit information. Its September 23 team-attributed channel brief also announces Phase 3 Challenge Card 10 and reiterates one prompt submission per miner per challenge card. These are new operational changes, separate from yesterday’s reported F1 improvement.
A security-model release and specific challenge rules are actionable for participants.
SN25 UR: indexer flags 100,000-alpha lock switching to decay
SubnetRadar records a September 22 change of a 100,000-alpha SN25 UR lock from perpetual to decaying mode. This changes the lock’s conviction behavior; it does not establish an immediate unlock, an executed sale or the holder’s motive. No relationship to today’s SN112 THOR announcement is inferred from the timing.
A substantial lock-mode change merits attention without presenting it as a completed sale.
SN32 It’s AI launches a paid referral program
SN32 It’s AI says its referral program is available inside user accounts. Referred customers receive 20% off and the referrer earns $7 in USD under the program’s terms. This is a customer-acquisition feature; the announcement does not establish how many users have joined or how much revenue it will produce.
A live, clearly specified customer feature merits a concise update.
SN50 Synth clarifies its current timeout and duplicate-submission rules
SN50 Synth’s team-attributed September 23 brief corrects its timing guidance: the current timeout is one second, while the five-millisecond budget will be enforced gradually. It also says duplicate models submitted across hotkeys will be disqualified. Miners should not treat the future latency target as the active limit today.
The clarification prevents miners building against the wrong timing limit.
SN51 Lium releases an inspector fix after a service interruption
SN51 Lium’s September 23 channel brief reports executor-v1.134 with an inspector-payload fix and a confidential-VM attestation update. It also records a server interruption that disrupted miner pods and a team-reported restoration. Exact outage duration and full recovery across every workload were not independently established.
A release and service interruption directly affect compute providers and customers.
SN54 Yanez schedules a sandbox phase after the September 24 cycle
SN54 Yanez’s team-attributed brief says the current cycle ends September 24, followed by a sandbox phase and a new task. During that phase, only the KAV reward track is active while UAV submissions undergo manual review. This is a planned subnet reward transition, not the whole network switching its allocation off.
A near-term reward-phase change is important to active miners.
Searches published headlines and summaries, not private bookmarks, unpublished leads or source-page contents. Duplicate headlines within an edition appear once; separate editions remain separate so follow-ups are preserved. Share the page URL to retain your filters.
