HybridAgents cuts an agent's tokens by up to 55%, but the key takeaway is that reasoning is now the bill

By El Ecosistema Startup · October 18, 2026.

A new preprint on arXiv, "Hybrid Reasoning Agents," introduces HybridAgents, a framework that lets an AI agent reason in a latent space, that is, in continuous internal representations without converting each step into text, and return to natural language only when it needs to act on the environment. According to the authors, it reduces tokens by between 35.5% and 54.7% and speeds up inference by between 1.8x and 2.6x.

The paper also reports an average performance improvement of 10.4% and cost savings of 2.7x to 3.7x. Cryptobriefing translates the latter into a compute reduction of roughly 63% to 73%. The evaluation covered 11 benchmarks (math, science, medicine, search and agentic reasoning) and three model scales. All figures come from the authors themselves, among them Jiaru Zou, of Stanford and Together AI.

The method rests on three pieces. HybridLink is a lightweight module that manages the switch between latent mode and text mode. A continuous working memory preserves context across tool calls. And a two-phase post-training, HybridSFT (supervised fine-tuning that stabilizes generation with tools) and HybridRL (reinforcement learning that teaches when it pays to think silently and when to act), completes the system. The code is on GitHub (togethercomputer/HybridAgent) and the models are on Hugging Face.

There are three important caveats. It is a preprint, with no guaranteed formal peer review. The tests used Qwen3-Base models with 4B, 8B and 14B parameters, so we don't know whether the effect holds for frontier models. And latent thoughts can be displayed as readable pseudo-text, but, as Cryptobriefing points out, that is not the same as reading the model's literal chain of thought.

The cost context explains why this matters. According to Gartner, inference cost per agentic workflow will rise more than fivefold through 2028, even as the price per token keeps falling (the firm forecasts a 95% drop by 2030). Gartner calls it the "inference paradox": better unit economics, higher total spending. According to the same analysis, a reasoning agent can cost up to 150 times more than a basic chatbot on a single task.

**Our take.** The thesis that interests us isn't the specific technique but what it reveals: agents' intermediate reasoning is AI's new hidden cost. For years we have celebrated falling per-token prices; agents eat up those savings because they think a lot, and they think in text that someone pays for and no one reads. Tackling that link matters more than gaining a point on a benchmark.

But there is a trade-off worth saying out loud. Reasoning in text is expensive, but it is auditable. If we move to thinking silently, we gain efficiency and lose some transparency, and the pseudo-text workaround is, for now, an approximation. In general, in regulated sectors such as medicine (one of the domains evaluated), that difference is not a detail: it will determine whether these agents can be deployed. It ties into an idea we advocate: the design question for agents is no longer just what they can do, but who can verify what they have done.

It's also wise to be wary of the "55% fewer" headline. Measuring success by tokens saved is measuring the symptom. What counts is the cost per completed outcome: saving tokens in a workflow that fails more often is a loss. With models of up to 14B, moreover, the test favors those who deploy on their own hardware, which fits with the democratization of open-weight: if these techniques scale, they lower the barrier to entry for useful agents without depending on a major provider.

In the short term, the practical effect is modest and in the hands of technical teams: separating reasoning tokens from final-answer tokens in tracking, routing each query to the cheapest model that can resolve it (the "inference tiering" Gartner recommends), and testing the hybrid approach on a bounded task before rewriting anything. In the long term, if compute per task falls steadily, agents stop being a thin-margin luxury and move closer to that scenario of abundance in which automating routine work is almost free. Today it is a lab promise; what's valuable is that it targets the right bottleneck.