Product
SketchPro Partners with OpenAI for GPT-5.5 Early Access
SketchPro has been testing GPT-5.5 in early access through its ongoing collaboration with OpenAI. Early results show exciting improvements in handling complex construction documentation and BIM workflows.
SketchPro has an ongoing research partnership with OpenAI that gives us early access to frontier models for testing and feedback. GPT-5.5 is the latest example. Through this collaboration, we’re helping shape how the latest model intelligence can be applied to real architectural workflows before broader availability.
Over the past few weeks, we’ve been evaluating GPT-5.5 on real Revit documentation workflows. We’re excited with its performance in helping architects automate complex documentation tasks & to bring it to our customers today.
BIM Fluency
We evaluate frontier models on a Revit/BIM fluency benchmark. The benchmark is built from internal evaluation models and prompt patterns derived from common AEC documentation tasks.
The benchmark is split into two tiers. Common use cases cover the tasks architects run dozens of times a day: placing tags, renumbering, basic view setup, standard dimensioning. Expert use cases cover the harder problems that typically require a senior architect or engineer– multi-view coordination, conditional logic across families, complex annotation cleanup, and edits that depend on reading a drawing's spatial context.
GPT-5.5 leads over Opus 4.7 and Gemini 3 Flash on both tiers, and the margin grows with task difficulty.
These scores measure raw model capability in isolation. In production, SketchPro orchestrates frontier models with the domain intelligence needed for professional quality documentation, which substantially improves performance on expert tasks.
Multimodal Reasoning
A separate benchmark measures multimodal reasoning: the model receives a typical floor plan drawing with tags and annotations and must identify and resolve annotation conflicts from a plain-language instruction.
In one representative case, three tags overlapped at an elevator core– a common congestion point where annotations and dimensions compete for the same space. GPT-5.5 correctly identified the collision and repositioned each tag against its host wall. Earlier models either missed the cluster entirely or moved unrelated tags elsewhere on the plan.
No model resolved every case in our evaluation suite, GPT-5.5 included. Unresolved cases are being shared with OpenAI as input to the next iteration.
GPT-5.5 is beind deployed within SketchPro to all of our clients today.
A note on security and privacy. Early access to GPT-5.5 runs through the same infrastructure SketchPro uses for all frontier models: secure enterprise API endpoints from OpenAI, Anthropic, and Google, not consumer or public products. Data sent through these endpoints is never used to train AI models, and all requests are encrypted and isolated.
