Introducing GyroThink 1.5 Pro, our frontier coding model

A substantial coding task is rarely one prompt followed by one answer. A feature touches a data model, an API, and an interface. A bug crosses a boundary between services. A refactor must preserve assumptions that are spread across a repository. The useful unit of work is the complete change, including the evidence that it behaves as intended.

GyroThink 1.5 Pro is designed for that kind of work. Use it to inspect an unfamiliar project, reason through an implementation, work with development tools, and review the result. Gyroscape keeps those activities in the same workspace so that the code, terminal output, and conversation remain close to the task.

GyroThink 1.5 Pro is fine-tuned from the Kimi K3 base model for Gyroscape’s coding and cybersecurity workflows. It works with the project context, development tools, and review surfaces available in the workspace.

GyroThink 1.5 Pro supports a 1 million-token context window and up to 131,072 output tokens per response. The context window holds the working conversation, code, and tool results; the output limit describes the maximum generated response. Actual output is also governed by your remaining allowance and request limits.

A large context window is most useful when the context is relevant. Start with the goal, identify the files and constraints that matter, and let the model inspect the rest as needed. For a migration, that might mean the old contract, its consumers, the proposed replacement, and the tests that define compatibility.

Native tool use lets GyroThink 1.5 Pro participate in Gyroscape’s development workflow. It can work with the tools made available to the task, including file inspection, terminal commands, and change review. Its text and image inputs also let screenshots and interface references inform implementation.

For a frontend task, share the relevant screen and describe the behavior you want to change. For a backend task, provide the failing request, expected response, and important constraints. Ask for the change to be verified against those requirements, then inspect the diff and test results before shipping.

Gyroscape also includes Cyber for repository security reviews. Use it to inspect code, investigate potential vulnerabilities, and turn findings into reviewable remediation work. GyroThink 1.5 Pro can support those tasks with the files, terminal, and testing tools available in your project.

Start with a defined scope, such as an authentication flow or a recent change. Ask for supporting evidence, review the proposed fix, and verify the affected behavior. The finding, patch, and test results stay attached to the same task.

The coding comparisons use published results for GPT-5.6 Sol, Kimi K3, GLM-5.3, and Claude Opus 5. They cover DeepSWE v1.1, Terminal-Bench 2.1 and 3.0, FrontierSWE v2, ProgramBench, and SWE-Marathon v1.1.

We also include current published comparisons for GPT-6 Astra and Claude Fable 5.1. Those results retain their exact benchmark versions and source attribution. Terminal-Bench 4.0 cannot be substituted for earlier versions, and evaluation setups can differ between publishers. The tables preserve that context so you can see what each score describes.

Benchmarks describe particular tasks under particular conditions. The agent scaffold, tool access, reasoning settings, and provider safeguards all matter. For your own evaluation, choose representative tasks from your codebase and assess correctness, the quality of the change, and the work required to review it.

Get the latest Gyroscape desktop app, open a project, and select GyroThink 1.5 Pro when it appears in your account’s model picker. Describe a concrete outcome and the checks that would establish success. You can begin with a contained bug fix or a well-defined feature and expand the task as you learn how the model works with your project.

The model page collects its capabilities, usage policy, benchmark methodology, and the latest published comparison snapshot in one place. It is a useful reference when deciding which model to use for your next task.