Theo Browne shares his early-access experience with GPT-6.1 Sol, explaining why he considers it a stronger everyday option than the preceding Sol release. He emphasizes the economics of cached input in agent workloadsPrompt caching reuses compatible computation from previously processed prompt content to reduce repeated model-input work. and reports strong results in his own coding benchmarks. Those figures are provisional personal tests: he discusses incomplete runs, setup mistakes and excluded tasks rather than presenting a standardized independent comparison.
In real codebases, Theo Browne values the model's detailed investigation, bug reproductionBug reproduction is the process of reliably making a software problem happen under known conditions before trying to fix it. and thorough reviews. He reports fewer abrupt mistakes than with Astra and describes practical computer-use testsComputer use lets an AI agent interact with software through interface controls such as screens, buttons, keyboards and pointers., while making clear that improvements in consistency do not eliminate failures. Long unattended rewrites and fleet management still expose weaknesses in judgment and progress.
A fish-game demonstration illustrates the gap between attractive generated 3D graphics and usable software. Theo Browne criticizes unnecessary interface text, weak movement and poor gameplay despite appealing models. His preferred workflow pairs Opus for implementation with GPT-6.1 Sol for investigation and review, including a concrete review that caught regressions before a mergeCode review examines proposed source changes to assess their correctness, clarity and suitability before or around integration..
Watch on YouTube




