Apresentando SimuLang: Playwright para todo o Desktop
Palo Alto, California • Oct, 2026
.png)
New research from Simular shows 90%+ token reduction and 3x speedup for business automation with the neuro-symbolic approach
New AI “workers” suffer from a peculiar amnesia. Ask an AI agent to file an invoice on Monday, it will painstakingly examine the screen to find the "Submit" button. Ask it again on Tuesday, and it will approach the task with the bewilderment of a toddler, burning through tokens to plan, reason and execute from scratch.
Simular, the autonomous computer company, released a paper detailing the solution. Relying entirely on large language models (like Claude or GPT) to drive a computer is fundamentally flawed. Instead, we propose a hybrid approach — neuro-symbolic — drawn from human biology: give the software "muscle memory."
When a human learns a new routine, conscious thought eventually gives way to automatic habit. We apply this paradigm to our computer-use agent, Sai, where the neuro-symbolic algorithm uses LLMs to discover a workflow and uses code to follow the known strategy without expensive reasoning at each run. Together the combination is robust against environmental changes.
With its LLM-powered agent mode, Sai recently achieved state-of-the-art performance on OSWorld 2.0, outperforming Claude Opus 5 Max Thinking and GPT-5.6 Sol Max while costing substantially less, completing complex, everyday tasks that usually take skilled humans more than one hour.
Sai is also available through its developer API.
We address another vice of an LLM: a habit akin to students marking their own homework. Simular uses independent "Judges" to inspect the screen and verify success before moving on, ensuring reliability for real-world adoption.
The result is a dynamic, optimizing AI agent that combines the neural components of models, safety verifiers to handle changing environments, and code that self-heals.
Optimize for reliability, not just capability
For agents to be truly useful in real business environments, they need not only be capable, but also reliable in repeated runs, combined with cost efficiency.
Our research found that, given any base model, Simular’s neuro-symbolic algorithm demonstrates significant gains in reliability, costs, and speed:
- 99% cheaper
By synthesizing exploratory steps into reusable, self-healing code rather than re-querying base models on every screen refresh, runtime execution costs drop by 1 to 2 orders of magnitude.
In one experiment, Simular’s neuro-symbolic mode got up to ~200x lower cost over running base models, leading to as much as 99% in cost reduction.
.png)
- Most reliable in every setting
On complex scientific workflows on ScienceBoard, pass^3 (success in all three repeated runs) improved by 13 percentage points over Claude Opus 5 and by 12.4 points over GPT-5.6 Terra, nearly doubling the base agent's reliability.
On computer desktop automation on OSWorld, pass^3 improved by nearly 16 percentage points over GPT-5.6 Terra. Under Claude Opus 5, whose base agent is already strong at 68.4%, we still see a measurable gain of 3.6 points (to 72.0%). - 3-5x faster
Removing continuous model re-prompting allows tasks to run at traditional software speed. On GPT-5.6-Terra, base agent latency is 4.6x/5.1x over our neuro-symbolic method, and using Opus 5, base agent latency is 3.9x/3.4x over ours.
“Most office work does not consist of one-off novel interactions; it involves recurring, multi-step procedures such as invoice processing in QuickBooks, CRM updates, and data reconciliation," said Ang Li, co-founder and CEO at Simular.
"Running continuous frontier models across every keystroke and mouse click is fundamentally unsustainable: it's too slow, burns thousands of dollars in unnecessary tokens, and introduces stochastic failures into deterministic business operations.”
"Model providers secretly want you to call expensive models for every step of a workflow, regardless of how repetitive it is. Sai frees you from this cost trap by learning repeatable procedures that reduce cost by up to 200x with more reliable success,” added Jiachen Yang, co-founder and CTO at Simular.
About Simular
Simular is a research-driven company building accessible intelligence to free humans from digital labor. Our flagship agent Sai automates workflows by controlling local and cloud computers. Agent S, our open-source agentic framework, was the first to achieve human-level performance on the OSWorld computer-use benchmark. Founded in 2023 by former Google DeepMind researchers, Simular is backed by investors including Felicis, NVentures (Nvidia), South Park Commons, Basic Set, and Lenny Rachitsky.
Construir computadores autônomos não significa substituir humanos. Isso significa cooperação.
Liberte suas mãos do computador. Baixe o Simular hoje gratuitamente.
