{"models":[{"provider":"openai","name":"chat-latest","architecture":{"type":"moe","parameters":{"total":900,"active":{"min":90,"max":300}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-3.5-turbo","architecture":{"type":"dense","parameters":{"min":20,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."}],"sources":[],"deployment":{"tps":20.0,"ttft":0.67}},{"provider":"openai","name":"gpt-3.5-turbo-0125","architecture":{"type":"dense","parameters":{"min":20,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-3.5-turbo-1106","architecture":{"type":"dense","parameters":{"min":20,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-3.5-turbo-16k","architecture":{"type":"dense","parameters":{"min":20,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."}],"sources":[],"deployment":{"tps":7.5,"ttft":1.07}},{"provider":"openai","name":"gpt-3.5-turbo-instruct","architecture":{"type":"dense","parameters":{"min":20,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-3.5-turbo-instruct-0914","architecture":{"type":"dense","parameters":{"min":20,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-4","architecture":{"type":"moe","parameters":{"total":1760,"active":{"min":176,"max":528}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":38.0,"ttft":0.92}},{"provider":"openai","name":"gpt-4-0613","architecture":{"type":"moe","parameters":{"total":1760,"active":{"min":176,"max":528}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-4-turbo","architecture":{"type":"moe","parameters":{"total":880,"active":{"min":88,"max":264}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":7.5,"ttft":1.17}},{"provider":"openai","name":"gpt-4-turbo-2024-04-09","architecture":{"type":"moe","parameters":{"total":880,"active":{"min":88,"max":264}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":7.5,"ttft":1.17}},{"provider":"openai","name":"gpt-4.1","architecture":{"type":"moe","parameters":{"total":352,"active":{"min":35,"max":106}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":45.0,"ttft":0.95}},{"provider":"openai","name":"gpt-4.1-2025-04-14","architecture":{"type":"moe","parameters":{"total":352,"active":{"min":35,"max":106}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":45.0,"ttft":0.95}},{"provider":"openai","name":"gpt-4.1-mini","architecture":{"type":"dense","parameters":{"min":40,"max":112}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":49.0,"ttft":0.95}},{"provider":"openai","name":"gpt-4.1-mini-2025-04-14","architecture":{"type":"dense","parameters":{"min":40,"max":112}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":49.0,"ttft":0.95}},{"provider":"openai","name":"gpt-4.1-nano","architecture":{"type":"dense","parameters":{"min":10,"max":37}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":69.0,"ttft":0.97}},{"provider":"openai","name":"gpt-4.1-nano-2025-04-14","architecture":{"type":"dense","parameters":{"min":10,"max":37}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":69.0,"ttft":0.98}},{"provider":"openai","name":"gpt-4o","architecture":{"type":"moe","parameters":{"total":440,"active":{"min":44,"max":132}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":41.0,"ttft":0.65}},{"provider":"openai","name":"gpt-4o-2024-05-13","architecture":{"type":"moe","parameters":{"total":440,"active":{"min":44,"max":132}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":70.0,"ttft":1.33}},{"provider":"openai","name":"gpt-4o-2024-08-06","architecture":{"type":"moe","parameters":{"total":440,"active":{"min":44,"max":132}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":10.0,"ttft":2.21}},{"provider":"openai","name":"gpt-4o-2024-11-20","architecture":{"type":"moe","parameters":{"total":440,"active":{"min":44,"max":132}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":34.0,"ttft":0.86}},{"provider":"openai","name":"gpt-4o-mini","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":58.0,"ttft":0.53}},{"provider":"openai","name":"gpt-4o-mini-2024-07-18","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":43.0,"ttft":0.65}},{"provider":"openai","name":"gpt-4o-mini-search-preview","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-4o-mini-search-preview-2025-03-11","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-4o-mini-transcribe","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-4o-mini-transcribe-2025-03-20","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-4o-mini-transcribe-2025-12-15","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-4o-mini-tts","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-4o-mini-tts-2025-03-20","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-4o-mini-tts-2025-12-15","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-4o-search-preview","architecture":{"type":"moe","parameters":{"total":440,"active":{"min":44,"max":132}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-4o-search-preview-2025-03-11","architecture":{"type":"moe","parameters":{"total":440,"active":{"min":44,"max":132}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-5","architecture":{"type":"moe","parameters":{"total":300,"active":{"min":30,"max":90}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":59.0,"ttft":3.02}},{"provider":"openai","name":"gpt-5-2025-08-07","architecture":{"type":"moe","parameters":{"total":300,"active":{"min":30,"max":90}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":59.0,"ttft":3.02}},{"provider":"openai","name":"gpt-5-chat-latest","architecture":{"type":"moe","parameters":{"total":300,"active":{"min":30,"max":90}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-5-codex","architecture":{"type":"moe","parameters":{"total":300,"active":{"min":30,"max":90}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-5-mini","architecture":{"type":"dense","parameters":{"min":25,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":76.5,"ttft":4.57}},{"provider":"openai","name":"gpt-5-mini-2025-08-07","architecture":{"type":"dense","parameters":{"min":25,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":76.5,"ttft":4.57}},{"provider":"openai","name":"gpt-5-nano","architecture":{"type":"dense","parameters":{"min":5,"max":18.5}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":84.5,"ttft":1.57}},{"provider":"openai","name":"gpt-5-nano-2025-08-07","architecture":{"type":"dense","parameters":{"min":5,"max":18.5}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":84.5,"ttft":1.57}},{"provider":"openai","name":"gpt-5-pro","architecture":{"type":"moe","parameters":{"total":3600,"active":{"min":360,"max":1080}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":2.0,"ttft":17.77}},{"provider":"openai","name":"gpt-5-pro-2025-10-06","architecture":{"type":"moe","parameters":{"total":3600,"active":{"min":360,"max":1080}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":2.0,"ttft":17.77}},{"provider":"openai","name":"gpt-5-search-api","architecture":{"type":"moe","parameters":{"total":300,"active":{"min":30,"max":90}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-5-search-api-2025-10-14","architecture":{"type":"moe","parameters":{"total":300,"active":{"min":30,"max":90}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-5.1","architecture":{"type":"moe","parameters":{"total":300,"active":{"min":30,"max":90}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":37.8,"ttft":1.15}},{"provider":"openai","name":"gpt-5.1-2025-11-13","architecture":{"type":"moe","parameters":{"total":300,"active":{"min":30,"max":90}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":37.8,"ttft":1.15}},{"provider":"openai","name":"gpt-5.1-chat-latest","architecture":{"type":"moe","parameters":{"total":300,"active":{"min":30,"max":90}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-5.1-codex","architecture":{"type":"moe","parameters":{"total":300,"active":{"min":30,"max":90}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-5.1-codex-max","architecture":{"type":"moe","parameters":{"total":300,"active":{"min":30,"max":90}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-5.1-codex-mini","architecture":{"type":"dense","parameters":{"min":25,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-5.2","architecture":{"type":"moe","parameters":{"total":420,"active":{"min":42,"max":126}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":48.0,"ttft":3.45}},{"provider":"openai","name":"gpt-5.2-2025-12-11","architecture":{"type":"moe","parameters":{"total":420,"active":{"min":42,"max":126}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":48.0,"ttft":3.45}},{"provider":"openai","name":"gpt-5.2-chat-latest","architecture":{"type":"moe","parameters":{"total":420,"active":{"min":42,"max":126}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-5.2-codex","architecture":{"type":"moe","parameters":{"total":420,"active":{"min":42,"max":126}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-5.2-pro","architecture":{"type":"moe","parameters":{"total":5040,"active":{"min":504,"max":1512}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":15.0,"ttft":11.3}},{"provider":"openai","name":"gpt-5.2-pro-2025-12-11","architecture":{"type":"moe","parameters":{"total":5040,"active":{"min":504,"max":1512}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":15.0,"ttft":11.3}},{"provider":"openai","name":"gpt-5.3-chat-latest","architecture":{"type":"moe","parameters":{"total":420,"active":{"min":42,"max":126}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-5.3-codex","architecture":{"type":"moe","parameters":{"total":420,"active":{"min":42,"max":126}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":32.2,"ttft":2.11}},{"provider":"openai","name":"gpt-5.4","architecture":{"type":"moe","parameters":{"total":450,"active":{"min":45,"max":150}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":58.0,"ttft":1.69}},{"provider":"openai","name":"gpt-5.4-2026-03-05","architecture":{"type":"moe","parameters":{"total":450,"active":{"min":45,"max":150}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":58.0,"ttft":1.69}},{"provider":"openai","name":"gpt-5.4-mini","architecture":{"type":"dense","parameters":{"min":56,"max":158}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":71.0,"ttft":0.8}},{"provider":"openai","name":"gpt-5.4-mini-2026-03-17","architecture":{"type":"dense","parameters":{"min":56,"max":158}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":71.0,"ttft":0.8}},{"provider":"openai","name":"gpt-5.4-nano","architecture":{"type":"dense","parameters":{"min":15,"max":58}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":51.2,"ttft":1.28}},{"provider":"openai","name":"gpt-5.4-nano-2026-03-17","architecture":{"type":"dense","parameters":{"min":15,"max":58}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":51.2,"ttft":1.28}},{"provider":"openai","name":"gpt-5.4-pro","architecture":{"type":"moe","parameters":{"total":5400,"active":{"min":540,"max":1800}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":2.0,"ttft":5.57}},{"provider":"openai","name":"gpt-5.4-pro-2026-03-05","architecture":{"type":"moe","parameters":{"total":5400,"active":{"min":540,"max":1800}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":2.0,"ttft":5.57}},{"provider":"openai","name":"gpt-5.5","architecture":{"type":"moe","parameters":{"total":900,"active":{"min":90,"max":300}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":47.0,"ttft":2.16}},{"provider":"openai","name":"gpt-5.5-2026-04-23","architecture":{"type":"moe","parameters":{"total":900,"active":{"min":90,"max":300}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":47.0,"ttft":2.16}},{"provider":"openai","name":"gpt-5.5-pro","architecture":{"type":"moe","parameters":{"total":5400,"active":{"min":540,"max":1800}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":5.0,"ttft":3.45}},{"provider":"openai","name":"gpt-5.5-pro-2026-04-23","architecture":{"type":"moe","parameters":{"total":5400,"active":{"min":540,"max":1800}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":5.0,"ttft":3.45}},{"provider":"openai","name":"gpt-5.6-luna","architecture":{"type":"moe","parameters":{"total":180,"active":{"min":18,"max":60}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":55.0,"ttft":1.99}},{"provider":"openai","name":"gpt-5.6-sol","architecture":{"type":"moe","parameters":{"total":900,"active":{"min":90,"max":300}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":44.0,"ttft":2.98}},{"provider":"openai","name":"gpt-5.6-terra","architecture":{"type":"moe","parameters":{"total":450,"active":{"min":45,"max":150}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":52.0,"ttft":2.26}},{"provider":"openai","name":"gpt-6-astra","architecture":{"type":"moe","parameters":{"total":1500,"active":{"min":150,"max":500}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":35.0,"ttft":4.2}},{"provider":"openai","name":"gpt-6-luna","architecture":{"type":"moe","parameters":{"total":38,"active":{"min":3.8,"max":11}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":68.0,"ttft":2.17}},{"provider":"openai","name":"gpt-6-sol","architecture":{"type":"moe","parameters":{"total":375,"active":{"min":38,"max":112}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":60.0,"ttft":2.04}},{"provider":"openai","name":"gpt-6.1-sol","architecture":{"type":"moe","parameters":{"total":375,"active":{"min":38,"max":112}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":46.0,"ttft":3.45}},{"provider":"openai","name":"o1","architecture":{"type":"moe","parameters":{"total":440,"active":{"min":44,"max":132}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":6.0,"ttft":1.09}},{"provider":"openai","name":"o1-2024-12-17","architecture":{"type":"moe","parameters":{"total":440,"active":{"min":44,"max":132}}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":6.0,"ttft":1.09}},{"provider":"openai","name":"o3-mini","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":92.0,"ttft":3.06}},{"provider":"openai","name":"o3-mini-2025-01-31","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":92.0,"ttft":3.06}},{"provider":"openai","name":"o4-mini","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":92.0,"ttft":2.93}},{"provider":"openai","name":"o4-mini-2025-04-16","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":92.0,"ttft":2.93}},{"provider":"openai","name":"o4-mini-deep-research","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"o4-mini-deep-research-2025-06-26","architecture":{"type":"dense","parameters":{"min":8,"max":28}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."},{"code":"model-arch-multimodal","message":"The model architecture is multimodal, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-35-turbo","architecture":{"type":"dense","parameters":{"min":20,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."}],"sources":[],"deployment":{"tps":20.0,"ttft":0.67}},{"provider":"openai","name":"gpt-35-turbo-0125","architecture":{"type":"dense","parameters":{"min":20,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-35-turbo-1106","architecture":{"type":"dense","parameters":{"min":20,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-35-turbo-16k","architecture":{"type":"dense","parameters":{"min":20,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."}],"sources":[],"deployment":{"tps":7.5,"ttft":1.07}},{"provider":"openai","name":"gpt-35-turbo-instruct-0914","architecture":{"type":"dense","parameters":{"min":20,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}},{"provider":"openai","name":"gpt-35-turbo-instruct","architecture":{"type":"dense","parameters":{"min":20,"max":70}},"warnings":[{"code":"model-arch-not-released","message":"The model architecture has not been released, expect lower precision."}],"sources":[],"deployment":{"tps":null,"ttft":null}}]}