Why It Matters
This content serves as a crucial look at the 'upper limit' of current agentic capabilities. By documenting extreme-budget...
So I've been using gpt-5.6 for awhile...
Theo - t3․gg
The Signal
In an extended early-access trial, the speaker used GPT 5.6 to complete high-stakes infrastructure, mobile, and agent-based builds, arguing the model excels at long-horizon task persistence and browser-based autonomy. While the results demonstrate significant capability, the speaker emphasizes this showcase features atypical, extreme compute usage rather than a general-purpose product review.
The Case
Operational Autonomy and Persistence
- The speaker documents 20-plus hour autonomous runs without manual intervention, noting the model reliably maintains intent and context across longer horizons than previous models.
- In a notable test of agentic behavior, a 71.2 billion-token goal run for a file-syncing project led the model to autonomously register for a PlanetScale cloud-database account.
- The model successfully performed complex system recovery, autonomously entering grub shells and fixing broken boot partitions via remote KVM access, a task the speaker previously viewed as nightmare fuel.
Coding and Infrastructure Refactoring
- GPT 5.6 enabled two full native rewrites of the T3 Code mobile app—switching from React Native to Swift AppKit and SwiftUI respectively—with each rewrite completed end-to-end in just 2–4 hours.
- The Lakebed cloud platform moved from a monolithic JavaScript structure to a robust TypeScript architecture, with the model managing CI pipelines, deployment refresh workflows, and centralized OAuth integration.
- An experimental Rust rewrite of the Hermes Agent achieved a usable state with a 15MB RAM footprint, though it remains far from complete feature parity with the baseline system.
The 1 Minute Signal Take
The speaker’s experience suggests GPT 5.6 is highly effective at recursive architectural work and long-running autonomous operations that would previously stall shorter-context models. However, because these successes stem from massive, non-standard inference budgets and hand-picked projects, the model's actual performance in constrained or production-critical environments remains an open question.
Pro Analysis
Full analysis always available on Pro.
Time saved:
Tags
Written by: 1 Minute Signal Editorial Team