Subnet News — September 16, 2026

SN45 AlphaRidge expands its crypto screener; brief reports V3 economics

SN45 AlphaRidge says its crypto screener now covers 481 coins with 48 sortable metrics, including momentum indicators, supply, market rank, relative strength and chart patterns. Separately, the September 16 channel brief reports Stage 5/V3 economics activated with a broader reward distribution, and says an unintended burn spike was corrected within 30 minutes. The product figures come from the project’s original post; the rollout and incident details remain attributed to the digest.

Why 3 stars: A quantified product expansion and potentially important reward changes merit coverage, with the less directly verified rollout labeled.

+11.0% α/TAO 24H price action as of Sep 16, 2026, 10:55 AM ET

StonkFun posts about $599K daily revenue and adds BROS launch pairs

StonkFun’s September 15 UTC table shows $599,181 in revenue, $357,844 spent on buybacks and 2.06 million STONK burned. The revenue total is one dollar above its X announcement. These are platform-wide figures, not revenue specifically generated by TAO pairs. The platform also added tokenized Dutch Bros stock, BROS, as a launch-pair option; Sunrise separately confirms the Backpack-issued asset. This is an expansion of tokenized-stock pairing on StonkFun, with no new direct Bittensor partnership implied.

Why 3 stars: A completed revenue day and a partner-confirmed asset addition are substantive updates to the requested StonkFun watch.

SN11 TrajectoryRL pauses submissions and warns against the 50-alpha fee

The September 16 SN11 TrajectoryRL channel brief says submissions are disabled during a system upgrade and explicitly warns participants not to burn the 50-alpha submission fee. The team is described as targeting completion this week, without a confirmed restart time. Treat the pause as an operational watch item until a direct reopening notice is available.

Why 3 stars: A submission pause paired with a potentially wasted payment is actionable miner news.

SN14 Cacheon connects its inference result to modeled chip economics

SN14 Cacheon reports that its last competition reached 2,337.9 tokens per second on MiniMax M3, 32.2% above its SGLang baseline. Its new commercial illustration applies that gain to an external inference-cost model: with prices and a $4.25 chip-hour cost fixed, modeled profit rises from $4.81 to $7.73 per chip-hour. The post explicitly labels this a scenario. The optimization result is project-reported; the profit figures are not measured subnet revenue or a customer’s realized margin.

Why 3 stars: The update explains the commercial value of useful miner output without confusing a cost model with actual earnings.

SN15 ORO refreshes qualification and moves its first race to September 16

SN15 ORO’s latest channel brief reports that the refreshed qualification suite has reached validators and submissions are unblocked, with a 30% passing threshold and hotkey cooldowns cleared. It says infrastructure issues moved the first race to September 16 at noon Pacific, or 3 PM Eastern. Preference-reasoning tasks are temporarily removed. These are team-attributed digest updates, not independently observed race activation; they supersede yesterday’s report of a pause.

Why 3 stars: A changed qualification process and a new race schedule are practical information for participants.

SN21 AdTAO revises the September 17 standing amendment

SN21 AdTAO’s adopted scoring amendment keeps the existing settlement-day aging rules after feedback that the proposed alternative disadvantaged longer-horizon forecasts. It instead adds a gradual discount for previous-model results and a field average counting one entry per copy group. The changes are behind their own activation switch and effective date. This is the published amendment for September 17, not proof that today’s standings already use it.

Why 3 stars: The adopted rule changes how model improvements and copied strategies influence rankings while addressing a concrete fairness concern.

SN41 Almanac raises forecasting’s configured reward share to 10%

SN41 Almanac has merged an increase in forecasting’s configured mechanism weight from 5% to 10%. The accompanying accounting change treats the market side as its own budget slice and preserves the portion reserved for forecasting rather than treating that reserve as burn. The release also makes boosts and clipping more visible in logs. These are verified code changes; adoption by every validator was not established.

Why 3 stars: Doubling the configured share for a task changes incentives and requires correct accounting between the two reward streams.

SN51 Lium updates the idle B300 payment reference to $6.40 an hour

SN51 Lium has merged a validator configuration override setting the B300 SXM6 AC hourly price reference to $6.40, replacing the older $5.10 value in the pinned shared configuration. The commit identifies the purpose as paying idle B300 capacity. This is an incentive-side pricing change; it should not be read as a guarantee that every retail rental is offered at the same rate.

Why 3 stars: A specific change to GPU supplier compensation affects participating operators and available capacity.

SN62 Ridges sets owner-uploaded baseline agents’ reward score to zero

SN62 Ridges has merged a change that gives baseline agents uploaded through the owner hotkey an initial reward score of zero. They can still be approved and ranked for comparison, but the code explicitly excludes them from earning emissions through that score. This makes the distinction between a benchmark baseline and a rewarded competitor explicit. The merge alone does not verify every running service’s deployed version.

Why 3 stars: Separating owner baselines from rewarded entries is a concrete fairness and reward-accounting improvement.

SN64 Chutes demonstrates Hermes configuring its Agent Toolkit

SN64 Chutes has published a Hermes-agent walkthrough: the agent reads the toolkit documentation, runs its diagnostic script and prepares a provider configuration for review before applying it. The first test inference call failed because the account was unfunded, an explicit limitation in the team’s report. The useful result is the documented setup workflow and review step, rather than a claim that every part of an autonomous inference run succeeded.

Why 3 stars: A concrete agent integration example makes the inference service easier to use and inspect.

SN65 True Performance Network reworks its benchmark after overfitting

SN65 True Performance Network’s channel brief says work on TPN-004 is paused while prechecks and the benchmark are redesigned with randomized tests and private evaluations. It attributes to the team that competition 4’s results were not useful because miners optimized the reward formula. The proposed response is to improve the rules rather than manually ban particular submissions. No new benchmark launch or independently verified restart date was found.

Why 3 stars: A benchmark redesign affects whether rewarded output reflects the intended task and changes the next competition’s requirements.

SN68 NOVA retries failed property profiles before rejecting compounds

SN68 NOVA’s evaluation code now retries TNP property profiles that failed to compute before rejecting a submission, distinguishing a transient computation failure from a compound that actually fails the criteria. The main repository has also updated its submodule to include the repair. Separately, the channel brief reports a historical-similarity limit of 0.5 from epoch 25,160; that runtime setting remains digest-attributed.

Why 3 stars: Avoiding rejection caused by failed infrastructure improves evaluation reliability; the separate configured threshold is attributed clearly.

SN76 Ormas documents its first external miner’s onboarding

SN76 Ormas has published a FAQ drawn from its first third-party miner’s September 15 run. It explains job-specific repository access, how an operator-run validator checks delivery, and a fix restricting staged changes to the client’s allowed paths. The current payment model is an accepted price in USD credit; its alpha emission model is still planned. The first empty reference-solver run settled as no delivery for $0, so onboarding is not evidence that this particular run earned money.

Why 3 stars: First external participation and explicit payment rules are meaningful progress, without overstating an unpaid setup run.

SN78 UMI reports a registration halt while launch policy is revised

SN78 UMI’s September 16 channel brief reports that registrations are being stopped during network work, with another update expected in one to two days. Separately, published version-3 launch documentation explicitly assigns the unallocated model share to a verified burn destination and preserves the earlier task-scoring proportions. The code documents a policy; it does not establish that the permanent open competition is live or that registrations have reopened.

Why 3 stars: The registration interruption and clarified burn policy materially affect participation, with activation kept unconfirmed.

SN90 KubeTEE fixes a reward-pool calculation that understated miner weights

SN90 KubeTEE’s new scoring fix distinguishes combined miner-plus-validator emission from the half available to miners after the owner cut. Its previous denominator used the combined amount, which the maintainers say roughly halved the intended payout weighting. The corrected calculation applies the 50% miner share. This is a verified implementation repair, not an independent audit that every affected miner has received a back-payment.

Why 3 stars: A concrete accounting error affects reward sizing, making the fix more significant than routine maintenance.

SN91 Cascade adopts five-hour training legs and a new budget unit

SN91 Cascade’s accepted September 16 decision extends training legs from three hours to five and denominates budgets in series-points. That lets a wider multichannel series train more channel tokens for the same number of time steps while keeping the five-hour wall limit. The release pins worker v0.9.0 and documents why old workers would count the budget differently. Actual gains still depend on the workload and whether a run hits its time limit.

Why 3 stars: This changes both the compute commitment and the incentive for richer training data, with an explicit operator compatibility requirement.

SN92 MicroTensor releases portable databases for SQL test corpora

SN92 MicroTensor v0.4.3 lets a test corpus carry the SQLite databases used by its SQL tasks, keyed by the hash of their raw contents and covered by the corpus digest. Validators supply the correct database when selecting a task, and the local corpus checker rejects unresolved references. This makes evaluation packages more self-contained and reproducible; it is a tooling release, not another newly settled competition round.

Why 3 stars: Reproducible evaluation inputs matter when miners are scored on database tasks and incomplete packages can invalidate results.

SN100 Cortex announces Apple Developer Program acceptance

SN100 Cortex says it has been accepted into the Apple Developer Program, opening a path to distribute its software through Apple’s app ecosystem. That is a distribution milestone, but developer-program membership is not approval of a particular App Store app or a commercial partnership with Apple. The announcement does not provide a public app launch date.

Why 3 stars: A specific distribution step is worth tracking while keeping program access distinct from a finished app release.

SN102 ConnitoAI adds coordinated task switching for running nodes

SN102 ConnitoAI has merged a shared task-synchronization client so nodes can follow the active training task supplied by the owner API. The client checks the task bundle’s content hash before accepting a transition and keeps the current task if the API cannot be reached. This supports coordinated changes without treating a temporary API failure as a reason to stop mining. The implementation is published; deployment across the network was not independently confirmed.

Why 3 stars: Coordinating training-task transitions and surviving control-service outages improves continuity for miners and validators.

SN112 channel introduces POSTMIND and a proposed new direction

The September 16 SN112 channel brief describes a new team and introduces POSTMIND as a network for creating and ranking social posts, captions and scripts. It says the repository and X account are still being prepared. The live roster still labels the slot for sale, so the reported ownership change and final identity have not been independently reconciled. This is an early transition watch, not a confirmed product or mining launch.

Why 3 stars: A possible new team and use for the slot is material, but the remaining identity and ownership uncertainty must stay visible.

SN114 SOMA plans fresh evaluations after mixed model routing

SN114 SOMA’s latest channel brief reports that some evaluations reached DeepSeek V4 Pro while others reached V4.1 Flash. It says the team is pinning V4.1 Flash and rerunning affected qualified submissions from the model switch. This matters because a model change can alter scores even when the miner’s submission is unchanged. The incident and remedial runs are digest-attributed and were not independently replayed.

Why 3 stars: Using one consistent evaluator is important to the fairness of competition scores and rewards.