- GPT‑5.1 is available in Glean Agents: Glean’s model‑agnostic platform now supports GPT‑5.1, which is snappier, more steerable, and better at instruction following than GPT‑5, leading to more grounded tool use and higher‑quality outputs.
- Stronger instruction following and research performance: With clear system prompts, GPT‑5.1 adheres to tool‑use policies and completion criteria, iterating until coverage is sufficient. In deep research evaluations, it was more performant (+11.9%), more comprehensive (+14.8%), and followed instructions better (+5.4%) than GPT‑5; it also improved Fast Mode compliance for quick, tool‑grounded lookups.
- Faster, token‑efficient reasoning for enterprise: GPT‑5.1 uses fewer reasoning tokens on average (916 vs. 2,018 for GPT‑5) via dynamic thinking—shorter for easy tasks, longer for hard ones—delivering faster responses and more reliable instruction adherence that enables greater agent autonomy; it’s available in Glean today.
Glean is a model‑agnostic work AI platform that supports the latest and most advanced open-source and proprietary models, including support today for OpenAI’s newly launched GPT‑5.1 in Glean Agents. We’ve received early access to the model and have been evaluating it for chat and deep research, finding the model to be better at instruction following and more efficient with token usage than GPT-5.
Steerable model makes gains in instruction following
GPT-5.1’s high level of steerability enables it to more easily adopt a target behavior or style when guided by a system prompt. If we spell out tool‑use policies and completion criteria in the system prompt the model follows them. In practice, that means we can direct GPT-5.1 to keep calling tools until coverage is sufficient, and only synthesize once it has enough grounded signal. Steerability is tightly linked to instruction following; if the model can be steered, it can more effectively follow instructions around constraints, steps, tools, and formatting.
Absent steering, GPT-5.1 tends to respond directly more than GPT-5, drawing on its internal knowledge or asking clarifying questions rather than invoking external tools. We found that for GPT-5.1 on Assistant, the model chose to respond directly, even when tools were available, 30.6% of the time, compared with 19.9% on GPT-5.
GPT-5.1 performed well on deep research, a structured, multi‑step workflow that uses company and web knowledge to produce a comprehensive, sourced brief. GPT‑5.1 on deep research evaluation was overall more performant by 11.9%, more comprehensive by 14.8%, and more closely followed instructions by 5.4% compared to GPT‑5.
We saw that GPT-5.1 follows the deep research system prompts for comprehensiveness, iterating and adapting its plan until it has fully answered the query. For insightfulness, it provides second-order analysis rather than simply regurgitating snippets of data it found. GPT‑5.1 picks up these nuances in the system instructions and adheres to them.
Taking our learnings from deep research, we applied GPT-5.1 to Fast Mode in Glean Assistant—a latency and token‑efficient mode for straightforward, information retrieval queries. Fast Mode has explicit instructions to make use of tools, and we found that GPT-5.1 with no reasoning set was actually better at following instructions than GPT-5 on minimal reasoning.
The combination of steerability and instruction following lets us shape GPT‑5.1’s behavior to the task at hand. For deep, multi‑step work, we can require exhaustive, cited synthesis. For quick lookups, Fast Mode in Assistant stays efficient and grounded—without skipping tools when they’re needed.
Faster, efficient token usage
GPT‑5.1 is also a more efficient thinker. It avoids unnecessary reasoning and keeps internal thinking short on straightforward tasks. This is a behavior aligned with dynamic thinking—a new capability of GPT-5.1 where thinking is shorter for easier questions and longer for harder ones at the same reasoning effort. Dynamic thinking comes through in the user experience—one of our immediate takeaways from GPT-5.1 is that it’s just a snappier, more efficient model. When we evaluated GPT-5.1 on enterprise use cases, we found that it used significantly fewer reasoning tokens on average: 916, compared to 2,018 for GPT-5.
Speed and instruction following in the enterprise
GPT‑5.1 delivers noticeably faster responses by trimming unnecessary reasoning steps, resulting in a faster user experience. It also shows improvements in instruction following, so the model stays aligned with what you ask it to do. That matters for agents at work: only when agents reliably follow instructions does it make sense to grant them greater autonomy.
GPT 5.1 is available in Glean Agents today. If you’re looking to get started with Glean or want to learn more, sign up for a free demo.
Authors: Nikhil Mandava, Rohan Dhoopar, Kunal Patil, Seema Jethani, Julie Mills
