The Pulse
Subconscious Raises $5.1M, Claims 80% Lower Agent Costs
Subconscious has raised $5.1 million to commercialize an inference platform for long-running AI agents. The Cambridge, Massachusetts, startup says dynamic context compression can cut costs by up to 80% on workloads exceeding 200,000 tokens.
AI.info Team ·
Subconscious announces $5.1 million in funding on September 22, 2026, as the Cambridge, Massachusetts, startup launches an inference platform built for AI agents that run across thousands or millions of tokens.
The company says its runtime can reduce costs by up to 80% for agents processing more than 200,000 tokens, while completing tasks twice as fast and extending the effective context window of existing models beyond 5 million tokens. MassVentures leads the pre-seed and seed rounds, with participation from Foothill Ventures, Underscore VC, E14 Fund, Oakseed Ventures and the Agent Fund, among others.
Subconscious describes the product as an inference layer rather than another agent framework. Its system combines dynamic context compression with caching designed to preserve useful information as an agent works through long chains of tool calls. The company says engineering teams can use the service through a managed cloud deployment or run it on their own GPUs.
Subconscious targets the cost of long agent traces
Long-running agents create a difficult infrastructure problem: every new step can add tool results, instructions and intermediate work to the model’s context. Standard inference systems repeatedly process much of that history, increasing latency and token costs as the task continues.
Subconscious says its runtime removes less useful material during inference while retaining information needed for later steps. The company presents the technique as a way to reduce the amount of context that the model must process without requiring changes to the underlying hardware, model or application.
For workloads beyond 200,000 tokens, Subconscious claims its platform can complete tasks twice as fast, extend effective context beyond 5 million tokens, improve results on selected coding and workflow benchmarks by between 1% and 10%, and reduce costs by as much as 80%. Those figures are company-reported performance claims rather than an independent evaluation.
Benchmarks show lower cost on long coding tasks
On the TriE benchmark, which Subconscious uses to measure system performance, the company says its runtime completed tasks twice as fast as SGLang and supported 2.3 times as many concurrent requests. The startup argues that higher concurrency allows the same GPU cluster to handle more agent work.
Subconscious also reports results from DeepSWE, a benchmark for long coding tasks. GLM 5.2 served through its runtime solved 46% of problems at an average cost of $2.79, compared with a 44% success rate and $3.92 average cost for the same model on standard inference infrastructure.
The comparison combines accuracy and cost, but the announcement does not provide a full independent methodology, hardware breakdown or third-party reproduction. Subconscious attributes the difference to its compression and caching system.
A 20-person engineering team cut monthly AI spending
The company says a 20-person engineering team moved from Claude to GLM 5.2 hosted on Subconscious in July. Over the following two months, the team’s monthly AI spending fell from $40,000 to $6,000, while engineers reported faster token throughput and no rate-limit problems, according to the announcement.
One agent trace ran for 4,571 turns and made 9,556 tool calls. Subconscious says the trace recorded 449 million tokens on its system, compared with 2.6 billion tokens that a conventional runtime would have billed, an 82% reduction. The company says the engineering team reported no loss of model capability despite the compression.
Subconscious does not identify the engineering team in the funding announcement. The example therefore provides a customer-reported case study rather than a named customer reference.
Open models sit at the center of the launch
The platform is designed around open models, including GLM 5.2, while supporting popular agent applications such as Claude Code, Codex, Pi, Copilot and OpenCode. Subconscious says teams can connect those tools through its command-line interface in about 30 seconds.
Jack O’Brien, co-founder and CEO of Subconscious, frames the product around a shift in the economics of agent deployment. “Open models finally got good enough this summer that their quality vs closed source models stopped being a compromise,” he says. “Meanwhile every engineering team I talk to has put a ceiling on what it spends per developer.”
O’Brien adds that Subconscious is running open models in a way designed for coding agents, allowing teams to reduce spending while running agents for longer periods. The company’s pitch depends on both parts working together: cheaper model access alone would not solve the repeated-context costs that grow during long tasks.
MassVentures backs an inference-focused strategy
Stacy Swider, vice president of investments at MassVentures, says the fund views Subconscious as a team focused on one of the harder infrastructure problems in AI agents. “We could not be more excited to be a part of their growth as they scale up to support thousands of companies and billions or even trillions of AI agents,” Swider says.
The funding gives Subconscious capital to expand a product that began with research from MIT into model and runtime design. The company previously introduced its TIM family of inference models and TIMRUN runtime, which it says are built to manage long-horizon reasoning, tool use and context reclamation.
Subconscious says the platform is available now through its hosted service, with self-managed deployments for companies that want to run the system on their own infrastructure or keep workloads fully air-gapped. Its immediate commercial target is coding agents, where long traces and repeated tool calls make inference costs especially visible.