Tech • AI • Robotics • Game
OpenAI's GPT-6.2 Preview Is Insane! Beats Opus 5.5 & Fable 5.5?
9.4/10TL;DR
Reports from early testers suggest OpenAI may be quietly routing some GPT-6.1 Soul coding sessions to a newer checkpoint widely referred to as GPT-6.2 Soul, with users claiming major gains in coding, visual generation and tool use.
KEY POINTS
Stealth-update speculation
Users testing GPT-6.1 Soul inside Codex at higher reasoning settings reported noticeably stronger performance than in earlier sessions, prompting speculation that some requests are being served by an unreleased GPT-6.2 Soul checkpoint. The theory centers on unexpected jumps in quality rather than any official model announcement. Reports describe the change as especially visible in coding-heavy workflows.
Knowledge probes offered weak evidence
One informal test asked the model, without web search, who hosted the Tony Awards on June 7, 2026, with the expected answer circulated as Pink. But that question was quickly treated as inconclusive because some users said GPT-6.1 Soul already returned the answer through the API. Additional probe questions focused on newer topics including Fable 5, Tibbo the recent guy, and Andrej Karpathy's move in May 2026, on the assumption that a fresher checkpoint might know information beyond the older model's cutoff.
Routing and settings claims
Testers said the most consistent way to trigger the stronger behavior was to select GPT-6.1 Soul in the model catalog and run it at High or Extra High thinking levels. Extra High was described as a practical sweet spot, while Ultra was seen as usable but less efficient. None of those routing claims has been independently confirmed.
Coding performance drew the strongest praise
In one benchmark-style trial, the model was asked to build a Minecraft-like clone and reportedly produced infinite terrain generation, textured environments, caves, ravines, ore types, block placement and block-breaking animations. Testers said the result felt closer to the original sandbox game's core systems than earlier outputs from the same model family. The emphasis was on breadth of implemented features, not merely a rough prototype.
Visual generation appeared sharper
Side-by-side comparisons of a prompt for a UFO hovering over an ocean were cited as evidence that the newer checkpoint handles waves, atmospheric lighting and scene composition better than earlier GPT-6.1 Soul runs. Testers described the improvement as obvious in head-to-head outputs. The claim was that the model's visual reasoning had advanced along with its coding ability.
Long-form coded visuals were a standout
A more ambitious test generated a cinematic sequence tracing human progress from prehistoric fire to orbital solar arrays and ecological restoration by 2035. The sequence included a forward-looking narrative in which AI accelerates its own training improvements by 2027, experiments multiply, and energy storage helps phase out fossil infrastructure by 2030. The result was presented as evidence that the model can script and render polished, concept-driven visual stories through code alone.
SVG, computer control and 3D tasks impressed testers
Testers highlighted clean SVG generations of three legendary Pokémon, arguing they surpassed competing outputs in clarity and style. Another demo had the model use direct computer-control capabilities to create a detailed painting in a paint application without writing code, a result described as unusually precise. A separate 3D example produced a black hole simulation with gravitational lensing and a glowing accretion disc, with no external assets or pre-rendered video.
Game clones and Blender comparisons fueled rivalry with Opus
A prompt to build an Elden Ring-style prototype reportedly yielded parry mechanics, blocking, combo attacks, stamina, save points and a boss encounter from a single request. In another comparison, the model was said to perform at or slightly above Opus 5.5 in Blender, including a detailed 3D house recreation with strong geometry and textures. Across several examples, testers argued the new checkpoint matched or exceeded Opus 5.5 while using tokens more efficiently.
CONCLUSION
The emerging picture is of a model checkpoint that may be significantly stronger than GPT-6.1 Soul, particularly in agentic coding, graphics generation and software interaction. Without official confirmation, the claims remain anecdotal, but the breadth of reported improvements has intensified expectations around a possible GPT-6.2 release.
Ask a question