# Microsoft releases open browser agent post-trained on 64 B200s for six days > Source: > Published: 2026-07-27 23:29:13+00:00 Microsoft made the weights for Fara1.5, a family of browser-use models built on Qwen3.5, [publicly available under an MIT license](https://www.microsoft.com/en-us/research/articles/fara1-5-computer-use-agent/) on July 22nd. The release includes models with 4 billion, 9 billion and 27 billion parameters, each trained to operate websites by reading screenshots and producing mouse, keyboard and navigation actions. The largest model required 64 Nvidia B200 GPUs for six days of supervised fine-tuning, according to its [Hugging Face model card](https://huggingface.co/microsoft/Fara1.5-27B). That equals 9,216 GPU-hours, a relatively compact post-training run for a model that Microsoft positions against proprietary computer-use systems. A July 27th [thread from 0xSero (@0xSero)](https://x.com/0xSero/status/2081866296564515314?s=46) estimated the compute at roughly $40,000 to $60,000. Current [Lambda pricing](https://lambda.ai/pricing) lists eight-GPU B200 instances at $6.69 per GPU-hour, which would put 9,216 GPU-hours at about $61,655. That comparison is only a rental-price proxy. It does not represent Microsoft's internal infrastructure costs, and it excludes the expense of generating data, running teacher models, evaluating trajectories and paying the researchers and engineers who built the system. The useful number is the scale: Microsoft converted an existing 27-billion-parameter model into a specialized browser agent with less than a week of post-training compute. That gives developers a clearer reference point for building narrow computer-use systems without funding a foundation-model pretraining run. ### The training pipeline behind the six-day run Fara1.5 came from Microsoft Research AI Frontiers, the lab led by Ece Kamar. The [technical paper](https://arxiv.org/abs/2606.20785), first submitted on June 18th and revised on July 22nd, lists Ahmed Awadallah and 14 co-authors across the project. The researchers supervised fine-tuned Qwen3.5 models using data from FaraGen1.5, a pipeline that creates web tasks, records attempts to complete them and filters the resulting trajectories. Microsoft used both live websites and six synthetic environments covering services such as email, calendars, marketplaces and experiment-management tools. Those replicas address a basic training problem for computer-use agents. Researchers cannot safely collect unlimited live demonstrations involving account logins, messages, purchases and other actions that change real data. In a synthetic environment, Microsoft controls the interface, database and expected result, allowing automated verifiers to compare the state before and after an agent acts. Microsoft says the resulting training mixture contains roughly 2 million samples. The 27B model card describes approximately 1 billion text tokens and fewer than 1 billion training images. A solver powered by GPT-5.4 generated task trajectories, while separate verifiers graded correctness, efficiency and whether the agent stopped when it needed missing information, clarification or approval for an irreversible action. This teacher-student structure is central to the economics. A large proprietary model performs tasks and produces demonstrations, then a much smaller open-weight model learns the specialized behavior. Developers pay the larger model during data generation rather than every time the deployed agent clicks through a website. ### Microsoft's benchmark claims need context Microsoft reports that Fara1.5-27B reached 72.3% on Online-Mind2Web, a benchmark of tasks performed on live websites. Microsoft's research article lists 58.3% for OpenAI Operator, 57.3% for Google's Gemini 2.5 Computer Use and 64.7% for Yutori Navigator n1 in the same comparison table. The results remain Microsoft-reported evaluations rather than independent tests. Microsoft used Browserbase to stabilize sessions and averaged Fara results across three runs. Browser-agent scores can move with website changes, blocked sessions, evaluator design and the exact tools available to each system, limiting how much a single table proves about real deployments. Microsoft's own materials also contain a discrepancy. The research paper and model card report 72.3% for Fara1.5-27B on Online-Mind2Web. A companion May 21st Microsoft blog post says the model exceeded 90% on that benchmark. The 72.3% figure is the one repeated in the technical paper, research article, repository and model card. WebVoyager reporting varies slightly as well. The research article gives Fara1.5-27B a score of 88.6%, while the current model card lists 89.3%. Both sources place the 27B model above the smaller Fara variants, though the version difference makes false precision unhelpful. ### An agent developers can inspect and contain Fara1.5-27B has a 262,144-token context window and observes browser screenshots without reading the document object model or an accessibility tree. It predicts grounded coordinates for clicks and drags, and it can type, scroll, visit URLs, run searches, ask the user questions and end a task. Microsoft recommends running the model through its [Fara repository](https://github.com/microsoft/fara) or MagenticLite, which places the browser in a container and adds domain allow-lists, action logs, monitoring and a pause control. The model card explicitly warns against giving Fara unrestricted browser access on a machine containing sensitive files or credentials. The MIT license gives developers room to inspect and adapt the weights, while the 27B size keeps deployment within reach of organizations that can serve a model across a small GPU cluster. Microsoft says the model has been tested on A6000, A100, H100 and B200 hardware, with at least two GPUs recommended for BF16 inference. Fara1.5 still carries the failure modes that make computer-use agents risky: prompt injection from page content, compounding mistakes across long tasks, hallucinated page states and confusion caused by ambiguous interfaces. Microsoft excludes unsandboxed use and high-stakes legal, medical and financial tasks from the model's intended scope. The release establishes a practical recipe for teams pursuing browser automation: use a frontier model to generate and verify interaction data, specialize a smaller open model, then constrain it inside a purpose-built execution harness. The six-day training run is the visible expense. The synthetic environments, teacher-model calls and verification system are the harder assets to reproduce.