September 13, 2026 · Published daily
Subnet News — September 13, 2026
MicroTensor’s hallucination-detection arena leads today’s star-ranked news. Plus Almanac’s $92K buyback dashboard, Engy’s completed burn, Epago’s first round, UMI’s broader temporary rewards, TensorUSD and Pallet, OpenRoboto’s reward change, CONST’s AI-funding argument and the latest StonkFun pairs. Researched through 11:21 AM ET.
- Coverage window
- Sep 12, 2026, 8:08 AM – Sep 13, 2026, 11:21 AM ET
- Subnets in coverage roster
- 128
- Subnets with updates
- 30
Biggest stories first
Today's ranked news
MicroTensor launches Arena 6 to catch AI hallucinations
MicroTensor has opened guard / mt-4g, a competition to build small models that identify unsupported claims, contradictions and logic errors in AI output. Its September 12 announcement sets limits of 3.2 GB on disk, 4.3 GB memory and 45 seconds at the 95th percentile. The repository now makes guard the sole active track. This is a concrete move into an enterprise problem: checking whether AI answers are supported by their evidence. The channel brief also reports a forthcoming zero-burn round and lower submission fee; those reward changes still need a live settlement check.
+0.1% α/TAO 24H price action as of Sep 13, 2026, 11:20 AM ET
Almanac opens its revenue dashboard: $92,130 in cumulative SN41 buybacks
Almanac’s new public page reports 400.94 TAO spent acquiring 69,031.9 alpha, valued at $92,130.31 across monthly records compiled through September 10. It lists 56,468.9 alpha as burned; September’s 12,562.9 alpha is perpetually locked instead. The team says terminal fees fund the program. Crucially, the page currently uses buyback value as its revenue figure because a separate fee feed is not yet available. This is a new transparency tool for an existing program, not a $92,000 purchase made today. A separate leaderboard upgrade makes IM2 reward recipients easier to inspect.
-1.4% α/TAO 24H price action as of Sep 13, 2026, 11:20 AM ET
Engy executes a 45.15 TAO buyback and burn, worth about $10,547
Engy’s September 13 transaction succeeded at block 9,057,890: 45.15 TAO went through add_stake_burn on SN53, with an AlphaBurned event. Its dashboard reports 1,758.69 alpha burned and a value of $10,546.58, bringing the displayed cumulative total to $49,501.65. The transaction verifies the purchase-and-burn action; the attribution to customer revenue comes from Engy. A separate miner fix checks the scheduler before treating a stalled health probe as a failed server, reducing unnecessary cancellation of active inference requests.
+5.8% α/TAO 24H price action as of Sep 13, 2026, 11:20 AM ET
Epago starts its first mining round; emissions remain off
Epago’s live dashboard shows Round 1 running on Finney with two entrants, three submitted checkpoints and the Tongyi-DeepResearch-30B-A3B baseline still champion. No duel result or new coronation had been published at the check. The dashboard explicitly labels the phase as burn, with emissions not live. The September 13 channel brief describes mining access and a scientific-paper practice corpus. Opening submissions and starting evaluation are real progress, but they do not mean miners are already receiving emissions.
-10.3% α/TAO 24H price action as of Sep 13, 2026, 11:20 AM ET
UMI opens temporary rewards beyond its two-miner pilot and caps linked registrations
UMI’s dated deployment records supersede the older two-miner pilot summaries. Both validators now use a temporary registration bridge: eligible miners provide a healthy HTTPS endpoint, and linked coldkeys, IP addresses or recorded funding sources share one group’s weight. The September 12 readback records positive incentive for 16 miners; September 13 documents finalized funding-grouped rows at block 9,055,700. These rewards measure endpoint availability, not translation quality. The bridge retains its block-based sunset, and the permanent open competition still awaits production activation.
-1.4% α/TAO 24H price action as of Sep 13, 2026, 11:20 AM ET
OpenRoboto extends real-robot reward accrual through evaluation
OpenRoboto’s updated rules say real-robot season pools now accrue from the opening of submissions until evaluation ends. Previously accrual stopped when submissions closed, just as physical robot evaluation began. The project records this as a rule change effective September 13 for both live real seasons. The published 85% real-robot allocation remains split between the two tracks. Each season has its own prize-pool reference, and the live season record governs the exact schedule.
+3.4% α/TAO 24H price action as of Sep 13, 2026, 11:20 AM ET
CONST challenges the funding question as AI leaders debate slowing the frontier
Dario Amodei’s new essay calls for pacing frontier capability development and commits Anthropic to embedded external evaluators; Sam Altman publicly supports the evaluator approach. CONST responds that open-source AI’s vulnerability is its funding, and argues for independently funded development. This puts Bittensor’s incentive model directly into the wider AI debate. It is a policy proposal and public disagreement, not an enacted restriction or evidence that other labs have stopped training.
TensorUSD expands its Pallet partnership for a long/short alpha product
TensorUSD announced an expanded partnership with Pallet to develop SN113’s long/short derivatives solution. Its accompanying posts describe two roles: liquidators handle vaults reaching the stated 120% threshold, while strategy miners compete on alpha-trading performance, including delta-neutral approaches. The team frames the product as a way to allocate TAO to strategies. This is an announced development partnership and mechanism outline; a live market, customer deposits and realized strategy returns were not established.
Cascade adds recovery for failed GPU rentals and expired funding keys
Cascade’s new code quarantines failed Lium hosts and retries pods that never become ready without charging a training attempt. Another change lets an expired miner-funded key fall back to an operator-paid lane when that option is enabled. The commit describes a participant delayed roughly 36 hours by infrastructure faults. A separate update ties multivariate data packing to the existing block 9,068,400 scoring switch, which is still ahead of this report’s snapshot. These fixes address the funded-round rollout; they do not prove every queued job has finished.
-5.4% α/TAO 24H price action as of Sep 13, 2026, 11:20 AM ET
ChronoLLM leads the current price movers at about +22%
ChronoLLM’s rolling 24-hour alpha/TAO change is 22.1% at 11:20 AM ET, the largest gain in the current screener. SubnetRadar also records a roughly 502 TAO purchase on September 12. No distinct new team launch or commercial announcement was established in the reviewed sources. The move earns attention as market activity, while its cause remains unconfirmed.
+22.1% α/TAO 24H price action as of Sep 13, 2026, 11:20 AM ET
RedTeam rises about 13% and ships challenge version 4.10.5
RedTeam is up 12.9% against TAO over 24 hours at 11:20 AM ET. Its repository bumps the ADA-detection challenge image to 3.0.3 in release 4.10.5, while the latest channel brief clarifies submission rejection and resubmission rules. Yesterday’s three-bank integration story remains significant, but no newly signed banking contract was established today. The price move is context, not proof of the reason buyers acted.
+12.9% α/TAO 24H price action as of Sep 13, 2026, 11:20 AM ET
Investing targets its next rebalance for a model update
The September 13 channel brief says Investing expects to deploy its next model at the upcoming rebalance, within days. It also says there is no OTF timeline for the fee changes the team is awaiting. Investing’s X account points to an early strong position on the Root APY chart while explicitly saying it is too early to draw a firm conclusion. Neither the deployment nor a durable yield advantage is established by those statements.
+11.5% α/TAO 24H price action as of Sep 13, 2026, 11:20 AM ET
Emissions watch: no new zero/nonzero flips in the daily comparison
At 11:20 AM ET, 45 of 128 subnet slots show zero observed allocation in the screener. Comparing with yesterday’s saved snapshot finds 0 switches between zero and a positive share. That does not mean every miner is earning: UMI’s temporary bridge, Epago’s burn phase and ORO’s race delay have different operational meanings. MicroTensor also publicly called for clearer rules and communication around pauses. These snapshots do not reconstruct administrative enable/disable transactions.
Chainlink includes ForeverMoney in its latest adoption update
Chainlink’s September 13 roundup names ForeverMoney among users of its standards, alongside Aave and other integrations. This is fresh visibility from the infrastructure provider for a previously announced connection. It does not establish another TAO bridge launch today or mean all ten integrations in the roundup involve Bittensor. The earlier ForeverMoney, TAO and Bankr developments remain part of the ongoing story.
StonkFun adds MON, SUI, AAVE and tokenized FLWS launch pairs
StonkFun’s September 12 posts add MON, SUI and AAVE as pairing options, followed by tokenized 1-800-Flowers stock. Sunrise separately announced the FLWS asset on Solana. These expand the same launch platform we are tracking around TAO pairs; they are not additional TAO listings or corporate partnerships with Bittensor. No distinct new ForeverMoney/Bankr TAO integration or new Shiz Cat launch was established in this window.
ReadyAI’s 17.07 TAO August buyback is backed by a successful burn transaction
Today’s ReadyAI channel brief reports completion of its August buyback. The linked explorer record verifies a successful 17.07 TAO add_stake_burn on SN33 at September 12, 03:41 UTC, with an AlphaBurned event. That means TAO was spent buying and burning alpha; it is not 17.07 TAO itself disappearing. The brief says the USD amount was 2.35 times the prior month and that September is tracking higher. Those comparisons and the planned Monday announcement remain team-attributed.
ORO Race 2 remains a watch item as the team works through validator fixes
Yesterday’s brief scheduled Race 2 for September 13 at 19:00 UTC / 3 PM ET. Today’s digest is less definite: it attributes to the team ongoing incentive-mechanism work and acknowledged validator failures, with another race dependent on the update. The repository also tightens what grading vocabulary agents can see and adds inference-usage capture. The 3 PM start should be treated as provisional until the team confirms it; no live restart or end to the reported full-burn period was established.
Ridges reportedly reopens submissions with an 86% screener threshold
The September 13 channel digest says Ridges has ended its draining state and reopened submissions. It attributes an increase of the second screener threshold to 86% to the team, with 90% only a possible later change. This updates yesterday’s temporary upload pause. Direct inspection of the site encountered a security checkpoint, so the reopening and threshold remain attributed to the channel digest rather than independently verified live controls.
NOVA records its first Boltz2 blueprint winner
The latest NOVA channel brief names dr.john_watson21th as the winner of epoch 20708, the first completed epoch with the new Boltz2 integration. It says all submissions navigated the new structure and that the winning code is being reviewed before open sourcing. The initial threshold for the next epoch is reported as 0.005. This is a competition milestone, not evidence of a validated medicine or a released winning implementation.
Vidaio compression challenge faces a source-video shortcut; fixes are reported
Today’s channel digest describes a scoring shortcut: miners could locate original versions of publicly sourced degraded clips and avoid doing the intended restoration/compression work. Members report chroma-based detection and a move toward AI-generated challenge footage. The digest contains no confirmed team account, and the public repository check did not independently establish a new deployed fix. This is an operational issue to watch, with the reported remedy still awaiting primary confirmation.
Reliquary removes throughput-based selection from its v6 rules
Reliquary’s updated v6 documentation makes eligible-group selection first-in, first-out and removes throughput and payload size as ranking inputs. It explicitly acknowledges that earlier arrival still has an advantage because windows close when filled. The release also bounds validator startup memory and documents the current generation profile. Miners need to match the validator’s advertised contract; the code change is not evidence that every operator has already upgraded.
Leadpoet ships scoring recovery after its Arena launch
Leadpoet’s September 13 changes add recovery for incomplete scoring assignments, isolate company-specific evidence failures, and keep verification prompts inside their size limit. The recovery script preserves accepted executions and scores while requeuing 84 incomplete assignments; later stages are bound to the corrected scorer. These are post-launch reliability changes. A published migration does not by itself establish that the production round has completed or its rewards have settled.
Thirty Spokes’ v3 is still in testnet rehearsal, despite a launch claim
A September 13 member-only digest describes v3 as live on SN99, but the current project README says it is being rehearsed end to end on testnet 526 before opening on mainnet 99. The new mechanism has miners submit a conductor model and fund benchmark calls through their own OpenRouter keys; an owner-run validator evaluates challengers against the king. The first-generation mechanism was retired September 7. The project website still contains older rules, so a confirmed mainnet launch should not be inferred from that page or the digest.
Score changes latency measurement and raises its historical tiebreak requirement
The Score team’s latest channel digest says Chutes calls no longer contribute their own RTF latency measurement; the dedicated latency loop handles that check. The historical tiebreak now requires four different challenges instead of two. It also says a new on-chain model commitment resets score history. These changes affect qualification and ranking, while the team continues investigating miner scoring complaints.
Swarm adds varied sunlight to make drone training harder to memorize
Swarm’s September 12–13 posts describe replacing the simulator’s uniformly white sky and assigning each seed a time on a realistic solar arc. The same mountain can then be flown in dawn light or strong midday light. The purpose is to test robustness to changing visual conditions instead of allowing a model to memorize fixed lighting. This is the fresh simulator update; the Andorra avalanche-drone deal circulating in recaps was announced earlier in the week.
SN112 is now labeled “for sale”; paired conviction changes need careful reading
The current subnet metadata labels SN112 “for sale.” SubnetRadar records a 100,000-alpha lock reduction and a 90,807.983-alpha lock addition against the same hotkey at block 9,058,340. These paired records should be read together: they do not establish a 100,000-alpha market sale or a simple scheduled unlock. The reviewed sources do not establish a buyer, sale price or completed project transfer.
Cortex tightens proof-result handling and failure reporting
Cortex’s latest changes require benchmark result artifacts, reject expired or invalid result paths, and improve reporting when a proof run fails to produce a usable result. The updates also preserve diagnostic evidence before destroying a failed guest. These are repository-verified reliability changes, rather than confirmation of a new consumer app launch or restored payouts across the network.
Ditto adds account-linking groundwork, with activation still off
Ditto’s v0.259–0.260 work connects miner hotkeys to Ditto accounts and prepares a Ditto Router upstream. The activation documentation says these features ship switched off and require separate production setup and validation. A Feedback Track table does not yet define a reward split. Separately, Ditto’s public post reiterates portability and says it is qualifying its first on-premises lead; that is a prospect, not a closed customer deal.
BaseStonk prepares Arc pairs for September 16
BaseStonk says it has added support in preparation for Arc’s stated September 16 launch. Its post says BSTONK stays native to Base while future Arc volume feeds the existing buyback-and-burn mechanism. This is a readiness announcement for a future launch date; it does not establish live Arc trading volume or a new Bittensor integration.
Sunday reading: Chutes’ distributed training and Bitcast’s business model
Today’s Opentensor weekly recap revisits Chutes’ Parallax training run and the week’s Lium, ORO, Pareton, Hippius and Bitcast milestones. TAO Daily also revisits Parallax and Bitcast’s Incentive Search interview. The Parallax run began September 10; the Bitcast interview was published September 9. They remain worthwhile context, with model quality and team-reported financial claims distinct from training throughput and independently audited accounts. They are not new September 13 launches.
New on the site: September 12’s members video
The latest ShizzyUnchained members video, “9/12/26 Bittensor News Updates & TAO Price Action,” is now featured in the videos library and membership promotion. YouTube marks it members-only. Existing public episodes and livestream replays remain available, including the Shiz Cat stream. No newer public upload or upcoming stream was found in the reviewed channel tabs.
AdTAO enforces its published memory limit and reduces settlement overhead
AdTAO’s new executor changes enforce the published 1 GB limit as resident memory and record each model’s peak usage. Settlement now reads only the predictions and recent settled rows it needs, and a failed resolution can be retried on the next tick. These are targeted operating changes to keep daily ad evaluation from consuming unnecessary memory.
BitMind reports a fix for validator memory exhaustion
The latest channel brief says the team traced miner timeouts to validators running out of memory, believes the issue is resolved, and is continuing to monitor. This is distinct from yesterday’s stale-weight caching fix. No network-wide availability measurement was obtained, so the recovery remains attributed to the team’s channel update.
Yanez schedules a new YID layer reveal for Exploit Summit
Yanez’s September 13 post says it will unveil a new layer in the YID ecosystem at Exploit Summit in about two weeks. The announcement does not name a new customer or describe a product already available for use. It follows yesterday’s verified SDK publication and is a separate upcoming reveal to watch.
Gittensor reports a targeted FP4 prefill optimization
Gittensor’s SparkInfer update uses the GPU’s hardware FP4 encoding path in its NVFP4 quantizers. The project reports a 1.08× prefill improvement at a 4K context in the commit. That is a specific measured case, not an 8% improvement across every model, context length or end-to-end request.
Sundaebar says the deployable-agent step arrives next week
Sundaebar’s official September 12 post says four of five steps from a demand signal to a deployable agent are live, with the fifth arriving next week. This corroborates the planned portal direction covered yesterday, but it remains a forthcoming release. The channel digest places the launch in the week of September 14 and says challenges pause while the team focuses on it.
Refinery expects validator weights to stabilize after child-key delays
The September 13 channel brief attributes ongoing weight stabilization to the Refinery team and says child-hotkey changes may take up to 24 hours to flow through. This explains the reported wait for correct UIDs to receive weights; it does not confirm that all affected miners have now been paid.
Audit trail
Coverage notes
Searched
- All 128 subnet profiles and 93 September 12–13 channel briefs checked, with previous reports used to distinguish new developments from recaps.
- 161 public X account candidates checked (153 available), producing 142 unique recent post records. Signed-in X broad and targeted Latest searches, selected full posts and a signed-in Grok scan supplied additional leads. Grok’s claims were checked against primary sources before use.
- Signed-in AlphaGap Feed: 34 distinct subnet cards reviewed, including recent development, pricing and emissions leads. SubnetRadar, TAO Daily, TAO Media and the public subnet news feed were also reviewed.
- 159 repository paths scanned with commit pagination from the prior cutoff and refreshed through 11:21 AM ET. Twenty-one selected patches were inspected in detail; Thirty Spokes’ newly discovered current repository was read separately.
- SubnetRadar Moments: 31 linked in-window event cards reviewed, including paired SN112 conviction records. Engy and ReadyAI burn transactions, Epago’s live dashboard and Almanac’s revenue methodology were checked directly.
- Market and emissions snapshot at 11:20 AM ET compared with yesterday’s saved data. Price figures use rolling 24-hour alpha/TAO changes; slot re-registration gains were excluded from ordinary movers.
- YouTube Videos and Live tabs checked. September 12 members upload verified on its watch page and added; no newer public upload or upcoming livestream found in the sampled tabs.
Public-data gaps
- All 128 current subnet profiles were reviewed, but X searches, public account timelines and AlphaGap cards were sampled. Private, deleted, unindexed and unsampled posts may be missing. This is broad coverage, not a claim to have captured every announcement.
- Private Discord messages were not directly accessed. Digest-only operational reports remain developing, including ORO scheduling, Ridges reopening, NOVA’s winner, Vidaio’s reported exploit fixes, Score’s rule changes, BitMind recovery and Refinery weights.
- Almanac’s page was compiled September 10 and uses buyback USD as a proxy for revenue pending a separate fee feed. Its burn methodology tracks alpha leaving the accumulation wallet without an undelegate; September alpha is perpetually locked. The full financial history was not independently reconstructed.
- Engy and ReadyAI successful add_stake_burn transactions were checked in the explorer. Customer-revenue attribution and monthly growth claims remain project-reported. ReadyAI’s transaction predates this window; today’s disclosure supplies the update.
- UMI’s chain details are drawn from dated project deployment records. Epago’s live register confirms a running round in burn phase. Neither establishes completed independent model evaluation or all-miner payouts. The screener comparison is not an administrative transaction audit.
- Thirty Spokes’ member digest says live, but its current README says testnet rehearsal before SN99 opening; older website documentation describes a retired mechanism. A new mainnet launch was not confirmed. A Grok lead about a named SN76 successor also lacked sufficient primary evidence.
- SN112’s lock reduction and addition occur together and are not proof of a market sale or a future release schedule. Coldkey swaps do not by themselves establish a sale to another person. No verified forward unlock calendar was found.
- The Chutes training run, Bitcast interview, Swarm commercial deal, Leadpoet Arena launch and RedTeam bank pipeline predate this edition’s fresh window. Recaps retain their dates; no new signed banking contract or additional TAO/Bankr launch was established.
- 5 repository paths were inaccessible; per-path errors are retained in the audit. Ridges returned a security checkpoint. Source checks occurred at different times; the last repository and market refresh ended at 11:21 AM ET.