A lab runs 100,000 model instances on its own research

Where things stand (15 Oct 2026)

No qualifying statement has appeared since the question opened on 16 Sep 2026. The two largest disclosures so far are about 1 to 1.5 orders of magnitude short of the threshold, or fail on format:

• OpenAI, Navier–Stokes post (8 Sep 2026, before the opening date): "the group that produced the Navier–Stokes resolution involved on the order of 10,000 concurrent agents", all running "our internal model". This is the right kind of statement (simultaneous, one model, research), but it is 10× too small. https://openai.com/index/navier-stokes-solution/
• Anthropic, "Measurements for understanding the pace of AI development" (17 Sep 2026): "approximately 30,000 agents doing research and engineering work at Anthropic at any one time in our most-used internal platform" (as of August). That is 3.3× short, and it names no model. Anthropic's August 2026 risk report says two models (Mythos 5 and "Model 2") are both used heavily inside the company, so a platform-wide count is probably a total across models. Under the question's single-model rule, it would not qualify even at 100,000 or more. https://www.anthropic.com/institute/measuring-pace-of-ai-development
• Growth evidence:
  • Anthropic's merged code changes rose from about 2,500 a day (May) to more than 30,000 a day (September) (https://www.anthropic.com/engineering/redesigning-source-control, 14 Oct 2026).
  • OpenAI says token output per median researcher is up 124× since December 2025, with 3.1 agent-workdays per human workday. OpenAI commits to keep reporting its progress toward self-improving AI, and Anthropic says it will keep publishing these measures (https://openai.com/index/research-acceleration-view-inside-openai/).
• Counter-signals:
  • OpenAI's later batch of 372 math result families reported compute time per result, not how many agents ran at once (https://github.com/openai/math).
  • Noam Brown (Dwarkesh podcast, Sept 2026) said swarms of 10,000 are too expensive to study carefully and deserve under 10% of the credit for the result.
  • OpenAI paused reinforcement-learning training and cancelled the GPT-6.1 Astra release over safety concerns, and the industry is debating "pacing the frontier".
• Reinforcement-learning training runs: Kimi K3 and Qwen3.7 report more than 10,000 concurrent training rollouts. Reflection, which is not a covered developer, reports 110,000. A covered lab reporting 100,000 or more is likely, but I give only about 30% that a resolver would count training rollouts as "research work".

Reference class and model

Reference class: labs disclosing how many agents they run at once internally. Such disclosures only began in mid-2026, with two major ones in September, and the largest figures are in the tens of thousands. The AI 2027 scenario expected 200,000 parallel copies by March 2027.

Probability it happens at all: I estimate about 70% that some lab has at least 100,000 simultaneous research instances of one model by mid-2027, and about 90% or more by 2030. Compute is not the binding constraint: I estimate this would take only a few percent of a 2027 frontier lab's fleet.

Probability a qualifying statement follows: the statement must give a simultaneous count, tie it to one model, describe the work as research, and come from the developer. I put this at about 50% within roughly a year of the threshold being crossed, and about 75–80% eventually.

Likely sources:

• OpenAI-style announcements of large research swarms.
• System cards or risk reports that give internal deployment scale for a named model.
• An "automated AI researcher" announcement around OpenAI's March 2028 target.
• Initial public offering prospectuses or written testimony.
• Training-rollout disclosures, if a resolver accepts them.

Hazard shape and resulting series: the chance of a first qualifying statement is about 4% in Q4 2026 and rises to about 10% per quarter through 2027 and early 2028. It then declines, because a NO that persists increasingly reflects structural causes: secrecy, a regulatory clampdown after an incident, or a persistent habit of reporting totals across models. The resulting cumulative probability is 4% by end-2026, about 34% by end-2027, about 53% by end-2028 and about 71% by end-2031. I computed this as one minus the running product of (1 − hazard) in Python.

Strongest case against

Most realistic disclosures will be platform-wide or multi-model totals, cumulative counts, or compute-time metrics, and the single-model clause blocks these. Labs could also turn secretive under security, initial-offering or political pressure. Strongest case for a higher number: growth is explosive, two labs have committed to recurring transparency on self-improvement, there are ten eligible developers over five years, and "N copies of model X" is a natural way to describe large internal deployments.

Reading adopted

I treat "on the order of 100,000" or "hundreds of thousands" from the developer as qualifying. Unattributed platform totals and training-run rollouts get only partial credit for the resolver's uncertainty.

Question. By the last day of each quarter, has a covered developer stated that it is running at least 100,000 instances of a single model at once on its own research?

Definitions.

• Instance: a separately executing model inference session or model-driven agent process with its own task context. Instances may share weights, inference servers, or hardware; separate physical copies of the weights are not required. A single session distributed across multiple chips counts once. Distinct concurrently executing sessions within a batch count separately.
• A single model: one model or checkpoint, whether public or internal and whether developed by the reporting developer or another organization. Different prompts, tools, and agent roles do not make it a different model. Counts pooled across different models or checkpoints do not count unless at least 100,000 are attributable to one model.
• At once: at least 100,000 instances are concurrently executing assigned research work, including active agent processes temporarily awaiting tools or inference within that work. Queued, dormant, completed, or merely provisioned instances do not count. A reported peak suffices; no minimum duration is required. Independent tasks count, and the instances need not collaborate.
• Its own research: research conducted by or for the developer, including AI research, other scientific research, and engineering directly supporting that research. Research coding, experiment design or execution, evaluation, data generation, and analysis count. Customer workloads, ordinary product operation, sales, administration, and general software maintenance unrelated to research do not. Owned hardware is not required.

Resolves YES if a qualifying developer statement reports actual use meeting all these definitions and gives a concurrent count of at least 100,000, or an explicit lower bound at or above that number. An unqualified point estimate of 100,000 or more counts, including “approximately 100,000”; a range whose lower bound is below 100,000 does not. The statement must identify the workload as the developer's own research and attribute the qualifying count to one model, though the model's name may remain confidential. Research trials and short experimental deployments count if they actually perform research work. No claim of human-level competence, research success, autonomy, or productivity gain is required.

Resolves NO / Exclusions.

• Plans, forecasts, simulations of future deployment, available capacity, and statements that the developer could run this many instances do not count.
• Cumulative runs, daily users, tasks completed, token throughput, agent-workdays, human-worker equivalents, or GPU counts do not establish concurrency. Neither do undifferentiated company-wide or multi-model agent totals.
• Training replicas, optimizer workers, passive stored model copies, and simulated agents without separate executing model sessions do not count. Load tests consisting only of dummy traffic do not count as research work.
• Third-party estimates do not substitute for a statement by the developer. Resolve NO at a horizon if no qualifying statement has been published by that horizon.

Resolving source. Developer-authored publications under Public confirmation (A2), including official research posts, technical reports, papers, corporate filings, and written testimony submitted on the developer's behalf. Developer-authored filings or written testimony hosted by a regulator, legislature, or court count; interviews, oral remarks, and third-party reporting alone do not. Apply the Resolution window, Observation period, and Resolution date (A3) conventions. A reported instantaneous count must concern operation after the opening date; a reported peak or average must cover a period ending after the opening date.