GPT-5.6 Checkpoints: A Real Vision Upgrade
Unreleased GPT-5.6 checkpoints show a marked jump in vision—long the GPT-5.x weak spot. Sharper SVG output has testers calling the model ahead of Gemini.
Although OpenAI has not officially announced GPT-5.6, unreleased checkpoints circulating among developers suggest a significant upgrade to its vision capabilities. Visual output, particularly Scalable Vector Graphics (SVG) generation, appears cleaner and more structurally coherent than in previous GPT-5.x models, addressing a historical weakness of the line.
As of early June 2026, GPT-5.5, released on April 23, remains OpenAI's official flagship model, with no documentation or API listings yet available for a successor. Instead, evidence of a follow-up model relies on internal codenames appearing in Codex backend logs and a week of community testing, which consistently show that the latest builds generate and interpret images noticeably better.
Three GPT-5.6 Checkpoints Observed in One Week
The pace of the updates is notable. Over the course of a single week, developers tracking Codex backend logs observed three distinct codenamed checkpoints appear and disappear. The earliest variant, 'joule-alpha,' was initially categorized as a 'Mythos-class' base model before giving way to 'kepler-alpha,' which was subsequently tested alongside 'kindle-alpha.'
Testers monitoring the updates, including @chetaslua, identified 'kindle-alpha' as a likely release candidate, noting its proficiency in handling image references during moderate reasoning tasks—an early, hands-on read on its vision capabilities specifically. OpenAI has not confirmed these findings, as the codenames represent backend identifiers for A/B routing tests rather than official announcements.
The rapid deployment cadence itself is a significant signal. The testing of three distinct checkpoints within a single week suggests to the developer community that the model is in its final tuning stages, especially as kindle-alpha's visual performance appears to improve with each successive build.
Refined SVG Generation Demonstrates Vision Upgrades
While the rapid updates suggest development speed, SVG generation offers evidence of improved vision capabilities. SVG generation has become a key method for evaluating vision capabilities, as SVGs consist of code describing shapes rather than pixels; a clean output indicates the model is reasoning about coordinates and structure rather than texture. Successfully generating complex SVGs therefore serves as a demonstration of spatial reasoning.
The test cases used by the community are demanding, including detailed renderings of an Xbox controller, a BMW M4, and a pelican riding a bicycle. Testers, including @TeksEdge, published side-by-side comparisons with Gemini 3.1 Pro using identical prompts, with general consensus favoring the GPT-5.6 checkpoints for depth and detail. Although these comparisons rely on selected screenshots rather than standardized benchmark evaluations, the consistent results across different testers point to a genuine improvement rather than an isolated success.
Addressing a Historical Weakness in the Model Line
This consistent performance indicates that the update is more significant than a routine, minor release. Throughout the GPT-5.x cycle, vision capabilities—including image comprehension, spatial reasoning, and visual output—have lagged behind the model's strengths in coding and text processing. A checkpoint that narrows this performance gap is therefore more notable than incremental coding improvements. Within the Codex logs, internal details also point to a larger context window of approximately 1.5 million tokens alongside coding and agentic upgrades, but the primary development remains the improvement in vision.
Important caveats remain, as no official benchmark scores from evaluations like MMMU or OmniDocBench are available, and user screenshots do not constitute a formal evaluation. Nonetheless, an earlier sighting of the 'Iris' build pointed to similar progress, and prediction markets currently estimate the probability of an official release before June 30 at 80 to 89 percent. The consensus among testers remains consistent: the vision capabilities of the GPT model series, historically a weak aspect, have shown measurable improvement.
