Subconscious raises $5.1M to cut the cost of long-running AI agents
The Cambridge startup says its inference system can reduce costs by up to 80%; its funding backs a push to sell the runtime to engineering teams.
By Ryan Merket · Published
Primary source: Subconscious
Why it matters
Subconscious is trying to monetize inference efficiency as teams increase agent use: the bet is that reducing token and GPU costs can make open-model agents cheaper to run without replacing customers' tools or hardware.
Subconscious raised $5.1 million across pre-seed and seed rounds to build and sell an inference platform for long-running AI agents, putting its bet on cheaper, faster execution rather than a new foundation model. The company announced the funding on September 22nd, 2026, and says the service is available to engineering teams through cloud and on-premises deployments.
Co-founders Jack O'Brien and Hongyin Luo say Subconscious grew out of MIT research into inference optimization. Luo's academic work includes a PhD in MIT's electrical engineering and computer science department, and a 2025 paper he co-authored with O'Brien and others described techniques for extending reasoning beyond conventional context limits. O'Brien's stated thesis is that open-weight models have become good enough for coding work, while the cost and speed of running them over long agent sessions remain barriers. In his announcement post, he said Subconscious had been operating for more than a year and a half before the funding and managed-service launch.
MassVentures led both rounds, according to O'Brien's post and the company announcement. The announcement names Foothill Ventures, Underscore VC, E14 Fund, Oakseed Ventures and Agent Fund among the participants. O'Brien's post also lists Companyon Ventures and Taihill Venture. The company did not provide a valuation or a breakdown of the two rounds in the announcement.
Subconscious says its runtime combines dynamic context compression with caching to reduce the amount of context an agent must repeatedly process. The pitch targets a particular cost driver: software agents can make many model calls over extended tasks, accumulating large token volumes. Subconscious says that for workloads above 200,000 tokens, its platform can make tasks twice as fast, expand models' effective context to more than five million tokens and cut costs by as much as 80%. Those are company-reported performance claims, not independently verified results.
The company's comparisons give a more specific, though still vendor-produced, view of the claim. On its TriE systems benchmark, Subconscious says its runtime completed tasks twice as fast as SGLang and handled 2.3 times as many concurrent requests. On DeepSWE, a long-coding-task benchmark, it reports that GLM 5.2 hosted on Subconscious solved 46% of problems at an average cost of $2.79, versus 44% and $3.92 for the same model on what the company calls standard inference infrastructure. The announcement does not describe the full test setup or provide independent replication, so the results should be read as Subconscious's own benchmark claims.
The strongest commercial evidence in the release is also a customer account provided by the company. Subconscious says a 20-person engineering team switched in July from Claude to GLM 5.2 hosted on its platform, reducing monthly AI spending from $40,000 to $6,000 over the following two months. One engineer's reported agent trace ran for 4,571 turns and made 9,556 tool calls; Subconscious says its runtime recorded 449 million tokens, compared with 2.6 billion tokens that a conventional runtime would have billed. The company attributes the reduction to context compression and says the customer reported no loss in model capability. It does not identify the customer or detail how the cost comparison was calculated.
That case points to the buyer Subconscious is pursuing: teams already spending heavily on coding agents that want to run more work on open models without rewriting their applications or buying more hardware. The service also supports deployment on a customer's own GPUs, according to the announcement. That option gives Subconscious a route to organizations that want to keep data and compute inside their own environment, while tying its economics to making existing GPU capacity handle more requests.
O'Brien argues that open models' improving quality makes this timing opportune, while engineering teams are imposing spending limits even as developers use agents more heavily. The company's product announcement names Claude Code, Codex, Pi, Copilot and OpenCode as tools that can connect through its CLI. Subconscious is therefore selling an infrastructure layer intended to work across existing agent products, rather than asking customers to adopt a new agent interface.
The $5.1 million gives O'Brien and Luo capital to turn their research into a managed inference business and to support the cloud and customer-operated deployment paths. Whether compression can deliver the advertised savings across varied customer workloads will determine how far the platform can expand beyond the early coding-agent use case. The announcement offers one unnamed team's reported spending reduction; it does not yet establish how broadly those savings generalize.