Asana cuts AI browser agent costs 76x with GPT-6.1 Sol

Work management platform Asana has dramatically reduced the cost and increased the speed of its AI-powered browser automation agent by switching to a newer model, demonstrating a significant leap in the economic viability of deploying such agents at scale. According to a case study published by OpenAI, Asana achieved a 76-fold reduction in cost and a 5-fold increase in speed during tests by adopting GPT-6.1 Sol, a reasoning model within the GPT-6 Astra series, to power its agent.

How did Asana achieve such dramatic cost savings?

Asana's breakthrough came from a model switch within OpenAI's ecosystem. The company had been using an earlier model, GPT-4o, to power a browser automation agent that helps perform complex, multi-step tasks within web applications. In testing, Asana engineers replaced this with GPT-6.1 Sol, a reasoning-focused model from the newer GPT-6 Astra family. According to the OpenAI case study, this single change resulted in the agent completing tasks 76 times more cheaply and 5 times faster. The report attributes these gains to the newer model's improved efficiency and reasoning capabilities, which allow it to navigate tasks with fewer errors and steps.

What is GPT-6.1 Sol and why does it matter for automation?

GPT-6.1 Sol is a reasoning model within OpenAI's GPT-6 Astra series, which was announced earlier this year. The "Sol" designation indicates it is optimized for complex, multi-step problem-solving and planning—precisely the skills required for a software agent that must autonomously operate a web browser. For tasks like filling out forms, navigating between pages, and extracting data, an agent needs to understand context, recover from errors, and follow a logical sequence. The case study suggests GPT-6.1 Sol excels at this, making fewer mistaken clicks or navigation errors that waste time and computational resources. This efficiency translates directly into lower cost per task and faster completion.

What does this mean for businesses and developers using AI agents?

The Asana test signals a potential inflection point for the practical deployment of AI agents. Previously, the computational cost of running sophisticated browser automation could be prohibitive for widespread customer-facing use. A 76x cost reduction changes the calculus, making it feasible to offer more capable AI agents to a broader user base without unsustainable expenses. For developers, it underscores the importance of model selection and the rapid pace of improvement in reasoning efficiency. It also highlights a shift where the primary constraint for agent deployment may shift from pure cost to reliability, safety, and the design of robust workflows that can handle edge cases.

How can you experiment with AI browser agents in Grafy?

For Grafy users interested in exploring similar browser automation, the direct tool to experiment with is Operator, Grafy's browsing agent that performs web tasks based on instructions. You can task it with research, monitoring, or data collection to see how current models handle autonomous operation. Furthermore, you can use Grafy Chat to converse with and compare multiple cloud models side-by-side, including reasoning-focused ones, to test their planning capabilities for multi-step tasks. For a deeper technical understanding of model efficiency and selection, the blog post Lowest-cost mode: spend the cheapest model that can answer explores cost-aware routing, while Watch AI reasoning models think — then keep the answer delves into how reasoning models process complex problems.

What are the broader implications for AI-assisted work?

This case study is a concrete data point in the ongoing trend of AI capabilities increasing while costs fall, a dynamic that accelerates adoption. It suggests that AI agents capable of taking over routine digital clerical work—data entry, status updates, information retrieval—are becoming economically practical. This could lead to a new layer of automation within software itself, where AI agents act as bridges between different applications or handle the tedious parts of digital workflows. The focus for toolmakers will likely shift to building secure, controllable frameworks where these agents can operate safely, and for users, to learning how to effectively delegate and supervise such automated assistants.

Frequently asked questions

How exactly did Asana measure the 76x cost reduction?

According to the OpenAI case study, Asana compared the cost per task of running its browser automation agent using GPT-4o versus GPT-6.1 Sol. The newer model's improved efficiency and accuracy led to completing the same tasks for 1/76th of the previous compute cost, a metric directly tied to the tokens processed and API calls required.

Is GPT-6.1 Sol available for general use?

The model is available through OpenAI's API, as indicated by its inclusion in the case study. Access typically depends on a developer's API plan and availability tiers. Users should check OpenAI's current model offerings and documentation for the most up-to-date availability and pricing.

Can I build a similar agent without deep engineering resources?

Yes, platforms are emerging that simplify agent creation. In Grafy, Operator provides a ready-made browsing agent, while Bots allows you to build embeddable support agents. The underlying principle—using a capable reasoning model to plan and execute steps—is becoming more accessible as the models themselves become more efficient and easier to integrate.

The original case study can be found on the OpenAI website: https://openai.com/index/asana-browser-agent.

---

Source: OpenAI News — Asana cuts model costs 76x in browser tests with GPT-6.1 Sol