Bittensor Weekly — September 11, 2026

Oro (SN15) got into Y Combinator, then swapped the benchmark under its own races.

Oro is joining YC's Fall 2026 batch. First crypto company with a live token to be accepted, per Unsupervised Capital's Chris Davis, who noted that YC underwrote the Bittensor model and treated the subnet as a structural advantage. The team announced it on X this week.

Then, on Sep 10, the subnet replaced the suite every agent was scored on. Legacy ShoppingBench out, ORO-Bench in: qualifying runs 525 tasks, 75 in each of seven families, covering intent decomposition, retrieval recall, constraint satisfaction, preference reasoning, ranking, recovery and justification. Races run a hidden set of 90 problems. The winner is decided on a difficulty-adjusted average of an agent's last three races.

Rankings are scoped per suite, so every agent scored against the old benchmark fell off the leaderboard when the switch landed. Submissions were blocked for part of the day while the team worked through environment errors and removed a validator running an outdated version. Seth posted times as they moved, reopened submissions at 15:48 UTC and dropped that day's qualifying threshold to 30% to filter out broken agents, before returning to the dynamic threshold capped at 55%.

Rewriting your evaluation while a cohort of miners depends on the old one is the expensive kind of housekeeping. Doing it in one day, with timings posted and agents reinstated when a validator turned out to be the problem, is the part worth copying.

"Today, we're announcing a solution found by our miners to the near-half-density case of Green's Problem 51, open for 16 years. Verified in Lean through Conjectures." That was the team's post on Sep 9. Another solve and its paper are due this week.

Rewards here scale with how long a problem stayed open, so a run of solves shrinks every bounty at once. New bounties are a fraction of what earlier problems carried, and wejh did not dress it up: "You guys did too good a job and solved too many." The rework in progress spreads rewards more evenly across problems and removes the incentive to sit on a finished proof while its bounty grows.

The pool also gets policed by the miners. One claimant withdrew an award request after the team showed that a formalization of the same target already existed, then handed over a cross-referenced list of 568 public proof records to help screen duplicates by content instead of by problem number.

How do you score a proof of a problem nobody has solved? In the team's own words: "A miner submits a formal proof, and the validator checks that it proves the exact committed statement under the permitted axioms and is accepted by Lean's kernel. The proof itself is the answer, and verification is binary, yes or no." An open-ended search that eats compute, with a pass/fail check at the end that no human referee can fudge.

A miner named Adam tracked the share of traffic captured by each model's top miner on sayGM: 65% in E25022, 71% in E25023, 74% in E25024, 95% in E25025. One key earning nearly everything on the astra model, and the trend still climbing.

Poggle did not argue with the math. He confirmed it and published more: the top two astra workers held 88% of selections over 24 hours, and miners receiving astra traffic fell from 30 to 15 in four days. Then the cause, in his words: roughly 70% of routing decisions are session pins, where a long conversation stays with whoever served its first turn, and astra volume rose 6.6x in four days. glm-5.2 is spread across 40 miners and kimi-k3 across 42, so the damage was one model rather than the whole board.

Two bugs got fixed the same day, including a price change wiping a lane's learned capacity, which is what compounded the concentration. Rolling back the routing is off the table: about 21,000 daily pins carry encrypted reasoning that fails if a later turn moves to another worker. Load spreading on astra is still open, and the team said so rather than calling it solved.

Shipping and building: SN1 (next competition lands next week; a "Getting to know Apex" explainer series started), SN10 (per-prompt score breakdowns live; SGLang campaign flagged at 0% emission weight while experimental), SN15 (ORO-Bench live, submissions open), SN28 (routing fixes shipped, load spread under investigation), SN30 (testnet running on netuid 504, validator update in progress), SN33 (August buyback tracking at 2.25x July, funded from 75% of dashboard revenue), SN44 (drinkware element live; API, SDK and MCP coming with Score Studio; emissions on parked challenges will be redistributed automatically), SN51 (buybacks and burns from the owner key; Lium alpha now usable as loan collateral on Mentat Lend), SN61 (scoring server state bug found and explained after submissions were rescored mid-day), SN62 (database engineering competition took 156 submissions in under 12 hours), SN64 (Pro 6000 capacity still spoken for), SN66 (Green's Problem 51 solve announced), SN67 (DeepSeek-V4.1-Flash added to the miner toolchain; citation indices now checked before scoring), SN68 (Boltz-2 oracle available for testing, live this week), SN81 (Reliquary V1.0 live on a new validator endpoint), SN85 (zero-score reason codes on the dashboard; duplicate outputs now split the score instead of paying the lower UID), SN90 ($10K of revenue recycled since Aug 23), SN96 (Microtensor's certified model served on testnet, user accounts live, buybacks in the works), SN106 (82 GPU executors, all scoring), SN111 (duplicate detection replaces coldkey blacklisting, both parties lose the reward instead), SN118 (v13 bench visible to miners Monday, Comet-2 agent on Sep 13), SN122 (mainnet hyperparameters set: 0.0005 TAO minBurn, 7-day immunity, 50% lockShare), SN124 (private model submission via CLI).

Dormant or stalled: SN83 (a miner questioned why it still draws 0.7% emission with no recent commits and no owner posts), SN35 (one message all week), SN59 (miner burn while the team builds out two customer APIs). Most other subnets produced nothing worth a line.

A miner named Asim stayed up for Reliquary's V1.0 launch. At 01:46 UTC he asked the team what the ETA was, mentioning that he had woken up for it and needed sleep. 0xgrizz answered plainly: "Sorry dude, that ETA was too optimistic. Don't wanna keep you up with another guess. Get some sleep, I'll update the channel once it's ready."

V1.0 went live at 05:18 UTC. By the afternoon the validator had thrown a bottleneck that RomAI could not pin down, and the message he gave the miners was the same shape as the first one: "There is a bottleneck on the validator that I didn't find yet, I keep you updated." No deflection onto miners' setups, no claim of a fix that did not exist.

Building in public usually gets sold as transparency about wins. This was transparency about a bad night, from a team that would rather lose a miner's sleep than his patience. Small thing. It is the sort of thing that keeps people mining through the next bad night too.

Subscribe for next week's edition: newsletter.bittensor.quest

Shaping the next issue: tell me what to keep, what to cut, and what you want more of — mailto:[email protected]