SenseTime's 1/50 Cost: A Dimensionality Reduction Experiment for Embodied Intelligence

Recently, the joint research team of SenseTime SenseNova released a technical report on arXiv, systematically evaluating the actual performance of OpenAI's GPT-6-Astra on embodied intelligence tasks, covering 12 simulation benchmarks and 8 real robot platforms. The test results themselves are well worth discussing: after relaxing the limit of interaction steps (increasing it to 10 times the standard budget) and enabling In-Context Learning, GPT-6-Astra has approached or even hit the saturation line on multiple simulation benchmarks, and the success rate of RoboDojo tasks has reached 81.9% — this figure corresponds to the performance under the 10x step budget, while the success rate of the same benchmark under the standard budget is 27.6%. However, on real robot platforms, long-horizon tasks, cross-task generalization, and fine control of flexible objects are still obvious weak points.

What is more worth analyzing is another concurrent move made by SenseTime: it open-sourced the SenseNova-RoboRSI execution framework and the supporting 27B embodied model, claiming to reduce the task completion cost to 1/50 of the direct call of GPT-6-Astra. According to SenseTime's disclosure, this data comes from its own tests and is yet to be retested by third parties; it measures the cost reduction of specific tasks such as stacking bowls and putting pens into pen holders under extended budget conditions, rather than a universal cost benchmark.

The cost of GPT-6-Astra in the extended budget test can be quantified: stacking bowls takes an average of 131 minutes, with an inference cost of about 19 US dollars; putting pens into pen holders takes an average of 263 minutes, with an inference cost of about 55 US dollars. 1/50 means that for the same tasks, SenseTime's open-source solution can reduce the cost to less than 1 US dollar. However, the applicable scope of this conversion needs to be clarified: it measures the cost reduction of specific tasks under the extended budget conditions set by SenseTime, which is not equivalent to a universal cost benchmark across scenarios and task types.

This is not a simple academic open-source move. It took place two months after SenseTime just released its first IFRS profitable interim report since its listing, at a time when the capital market still has fierce divergences over its valuation logic. SenseTime is using a very pragmatic approach to answer a question it has avoided for a long time: when its main business has not yet fully achieved operating profitability, how to reposition itself in a broader technological narrative.

The strategic shift from "selling models" to "building ecosystems"

SenseTime's financial structure in the first half of 2026 is the premise for understanding this move.

Revenue in the first half of the year reached 2.91 billion yuan, a year-on-year increase of 23.4%; IFRS net profit was 617 million yuan, the first positive figure since its listing. However, the composition of this profit requires prior explanation: the main factor driving IFRS profitability is not the full turnaround of operating profit, but the "net other gains" of 2.127 billion yuan, which includes 521 million yuan of gains from the sale of subsidiaries and 1.249 billion yuan of unrealized gains from financial assets measured at fair value. After excluding these items, the adjusted Non-IFRS loss still stood at 386 million yuan, and the main business has not yet crossed the operating break-even line. In other words, the signaling significance of the "first profit" is greater than the substantial operational qualitative change.

At the same time, SenseTime's revenue structure has completed the transition from visual AI to generative AI: generative AI revenue reached 2.327 billion yuan, accounting for 79.9% of the group's total revenue; traditional visual AI revenue was 497 million yuan, with its proportion shrinking to about 17%. Token service volume is expanding rapidly, and the large computing facility business is promoting the large-scale commercial application of domestic heterogeneous computing power. However, the heavy-asset logic of the Token factory means that there is always a gap that needs continuous verification between the growth of scale and the improvement of profits.

At this juncture, embodied intelligence has been pushed to the forefront.

The logic is not complicated. The competitive landscape of generative AI is already highly concentrated, and SenseTime is facing the full suppression of rivals such as Alibaba, ByteDance, and DeepSeek in terms of model capability and pricing power. The business of Token factory is essentially a hybrid of computing power leasing and model services. The scale can be expanded, but the flexibility of profit margin is constrained by upstream computing power costs and downstream customer bargaining power. As for embodied intelligence — or "AI for the physical world" — it is still a track where standards have not been solidified and the landscape has not converged. Intervening at this stage and establishing the influence of developer ecosystem and technical standards in an open-source way is a means to leverage a long-term ecological niche at a very low cost. Strategically, this is an attempt to exchange open source for ecological niche, and its success depends on whether the community adopts it in practice.

The entry point SenseTime chose is not to manufacture robot bodies, nor to participate in the general embodied large model benchmarking competition, but to "make embodied tasks cheaper". The design idea of the RoboRSI framework is very clear: package each simulation benchmark and its dependencies into an independent Docker image, expose environment interaction and native scoring interfaces through a unified HTTP API, so that researchers only need one configuration file to run end-to-end evaluation, without manually building each simulator. The positioning of the 27B model is to "internalize" the verified experience into the model weights to reduce the number of trial and error. According to SenseTime's disclosure, the success rate of RoboRSI-27B on RoboCasa365 increased from 26.0% to 46.5%, while the Token usage decreased by 73.3%; this result comes from SenseTime's own evaluation, and RoboCasa365 is a simulation benchmark, whose migration capability to the real physical environment still needs to be verified.

The goal of this combination is not to be "the strongest", but to be "the most usable". In the process of embodied intelligence moving from the laboratory to real scenarios, cost, rather than benchmark score, is becoming the real bottleneck. Whether SenseTime's solution can truly solve this bottleneck depends on actual deployment and retesting.

The true nature of 1/50: the limiting conditions behind the figure

1/50 is a highly communicable figure, but it needs to be understood under specific test conditions.

SenseTime's comparison benchmark is the inference cost of GPT-6-Astra under the "extended budget" condition. A considerable part of GPT-6-Astra's failure modes is not that it "cannot do it", but that it "stops too early" — among the 9 tasks of RoboDojo, all 5 originally failed, but after only allowing the model to continue interacting and increasing the step limit by 10 times, all 5 became successful. This means that a large part of GPT-6-Astra's inference cost is consumed in repeated trial and error and exploration of super-long steps.

The cost reduction logic of RoboRSI essentially uses "pre-trained experience" to replace "trial and error during inference". The 27B model solidifies the task patterns verified by evaluation into the weights, so that the model does not need to start repeated interactions from scratch when facing new tasks. SenseTime's own expression is "building tested experience into model weights, making it require less trial and error".

This explains the true nature of 1/50: it measures the cost reduction of a specific type of embodied tasks (short-range operations such as stacking bowls and putting pens into holders) under specific evaluation conditions (extended budget, simulation benchmarks), rather than a universal cross-scenario cost benchmark. SenseTime also admits that cost "is still a problem". In terms of long-horizon tasks, cross-task migration and fine control on real robot platforms, the bottleneck has not disappeared because of open-sourcing a 27B model.

Open source as a strategic tool: what is SenseTime betting on

SenseTime's choice to open source at this juncture has its specific strategic considerations.

First, the ecological niche of embodied intelligence has not been locked. Different from generative AI — which already has clear leading models and pricing systems — the software stack, evaluation standards and execution frameworks of embodied intelligence are still in the early stage. Open-sourcing a reproducible evaluation framework and supporting model at this stage is to strive to become a "tool used by default". Once RoboRSI becomes a common entry for researchers to retest and compare embodied capabilities, SenseTime will obtain a kind of non-exclusive influence in this track. The key lies in whether the community adopts it in practice.

Second, open source is a low-cost means to hedge the valuation pressure of the main business. SenseTime's market value has shrunk by more than 80% from its peak after listing, and the market's patience for "whether the Token factory can make profits independently" is still being consumed. Embodied intelligence provides a new narrative fulcrum: a company that has accumulated perception and spatial understanding capabilities in the visual AI era has technical proximity in the early stage of physical AI. Whether this proximity can be transformed into commercial competitiveness still requires commercial verification, but at least it provides the capital market with a differentiated label that distinguishes it from "just another loss-making large model company".

Third, SenseTime's financial structure determines that it cannot repeat the heavy-asset investment of the Token factory in the embodied intelligence field. The R&D cost of the open-source framework and the 27B model is far lower than building a ten-thousand-card cluster to train a trillion-parameter embodied large model. This is a play that uses engineering efficiency and ecological strategy to replace capital density, which matches SenseTime's current resource constraints.

Conclusion: From "proving profitability" to "proving there is still a position"

SenseTime's story in 2026 is undergoing a center of gravity shift.

The IFRS profit in the first half of the year solved the problem of "will this company collapse financially" — to be precise, "will it collapse at the statement level", rather than "whether the main business already has self-sustaining capability". The Non-IFRS loss narrowed to 386 million yuan, Token service volume is expanding, and the large computing facility is promoting the large-scale commercial application of domestic computing power — these are progress at the survival level. However, the pricing of the capital market is never only about survival. It is about whether a company still has an irreplaceable position in the next stage of technological evolution.

The open source of RoboRSI is SenseTime's first official answer to this question. The quality of the answer does not depend on how much impact the figure 1/50 has in communication, but on three things: whether this framework can be truly adopted by the embodied intelligence research community, whether the generalization performance of the 27B model in actual robot tasks can withstand independent retesting, and whether SenseTime can find a path between the open-source ecosystem and commercial monetization that does not rely on the expansion of Token scale.

The window for embodied intelligence will not be open forever. The time SenseTime gains through open source needs to be cashed out with the depth of the ecosystem. Before that, 1/50 is still a promise that needs to be verified, not a conclusion that has been implemented.

This article is from the WeChat Official Account "Tech Business", author: Tech Business, published with authorization from 36Kr.