OpenAI launches GPT Image 2.5, introducing Sunburst and Flare tiers with significant Elo score improvements over the previous GPT Image 2 generation.
On September 1, 2026, OpenAI officially expanded its generative media lineup with the release of GPT Image 2.5. Positioned strategically between the widely used GPT Image 2 and the upcoming high-fidelity 2.6 iteration, this release marks a significant milestone in the evolution of diffusion-based transformer architectures.
The release is bifurcated into two distinct tiers: Sunburst and Flare. These tiers are designed to cater to different developer needs, ranging from high-throughput application requirements to ultra-high-fidelity creative workflows. For AI engineers, this represents a more granular approach to scaling image generation costs against quality requirements.
GPT Image 2.5 utilizes an advanced multimodal transformer backbone that optimizes the relationship between text embeddings and latent diffusion processes. While OpenAI has not disclosed the exact parameter count, the model demonstrates enhanced spatial reasoning and semantic alignment compared to its predecessors.
The Sunburst tier focuses on rapid inference and high-volume generation, making it ideal for real-time applications. The Flare tier, conversely, prioritizes 'sharper details' and 'more precise editing' capabilities, allowing for finer control over local texture and lighting consistency. This architectural split allows developers to optimize their latency-to-quality ratio effectively.
The most compelling evidence for the leap in quality comes from the Artificial Analysis Elo leaderboard. GPT Image 2.5 Sunburst achieved an Elo score of 1229, ranking it 2nd overall in the industry. The Flare tier followed closely with an Elo of 1210, securing the 4th position.
To put this in perspective, the previous iteration, GPT Image 2 (released in April 2026), held an Elo of 1190, which placed it 5th. This jump indicates a substantial improvement in human preference and visual coherence. The evaluation methodology relies on blind head-to-head comparisons where human evaluators and automated vision-language models rank image quality based on prompt adherence and aesthetic value.
OpenAI has maintained a consistent pricing structure for the 2.5 generation, ensuring predictable costs for scaling production environments. Both the Sunburst and Flare tiers are priced at a premium level, reflecting the increased computational requirements for the improved detail density.
Developers should note that while the pricing is higher than the legacy GPT Image 2 models, the efficiency gains in prompt adherence may reduce the number of iterations required to achieve a desired output, potentially lowering the total cost of ownership (TCO) for creative pipelines.
The dual-tier nature of GPT Image 2.5 makes it highly versatile. The Sunburst tier is best suited for high-frequency applications such as dynamic UI asset generation, social media content automation, and real-time game asset prototyping.
The Flare tier is the preferred choice for professional creative workflows, including high-end concept art, advertising photography, and detailed architectural visualization. Its ability to handle precise editing makes it a powerful tool for agents performing iterative design tasks.
Access to GPT Image 2.5 is available immediately via the OpenAI API. Developers can integrate the model using the standard OpenAI Python SDK or direct REST API calls. Ensure your API key has the necessary permissions for the new 2.5 model endpoints.
For those looking to test the limits of the model, we recommend starting with the Sunburst tier to benchmark latency, then transitioning to Flare for high-fidelity requirements. Documentation regarding specific parameter tuning for the Flare tier can be found in the official OpenAI developer portal.
API Pricing — Input: $210.72 per 1k images / Output: $210.72 per 1k images / Context: Pricing applies to both Sunburst and Flare tiers per 1,000 images generated.