OpenAI introduces Images 2.6 across ChatGPT, Codex, and the API, adding GIF animation support and improved editing consistency. The update boosts generation speed by 50% and offers developers two distinct models: Flare for low-latency tasks and Sunburst for high-fidelity visuals.
Inference Latency
推理延迟
Time required from input request to generating the first or all tokens
Curated Intelligence
Typesafe AI's decision model Jev was adopted by 35% of paid teams on Vercel within two weeks, setting a platform record. By limiting output options for fast, low-cost data classification, its success validates the complementary role of specialized lightweight models alongside frontier LLMs.
Krea released Krea 3, an image editing model emphasizing precise local edits that change only requested areas. The company claims it is 4x faster than competitor Nano Banana Pro with lower per-image costs, offering API access and private deployment options.
HeyGen's new text-to-speech model, HeyGen Voice, is now available on OpenRouter after topping the Artificial Analysis TTS leaderboard in blind tests. The service offers standard and low-latency Turbo modes, with API users receiving a 50% discount ($15 per million characters) through October 31.
PolyAI CTO Shawn Wen stated at HumanX that voice AI has not yet reached its 'ChatGPT moment' despite the advent of full-duplex models. He identified fast reasoning for natural conversation as the next hurdle and emphasized the need for transparency and trust in enterprise deployments.
OpenAI introduced an 'Ultrafast' service tier for the existing GPT-6.1 Sol model in its Responses API, enabling lower latency without changing the model name. While this mode significantly reduces delay, it costs 6 times more per token than the standard tier. OpenAI claims performance approaches GPT-6 Astra in complex coding and professional tasks, targeting users with high willingness to pay for speed.
Hugging Face and Cerebras have partnered to integrate the Gemma 4 model into a low-latency inference platform, creating an open, modular architecture for real-time voice interaction. The solution employs a cascaded speech processing approach to significantly reduce response times, addressing the critical latency bottleneck in current voice AI user experiences.
OpenAI introduced WebSocket support in the Responses API, reducing end-to-end latency for agentic workflows like Codex by 40%. As model inference speeds approach 1,000 tokens/s, API service overhead became the bottleneck; WebSockets mitigate this by minimizing HTTP handshakes and round-trips during tool execution loops. This update significantly enhances the developer experience for building real-time, high-frequency interactive agents.
OpenAI has introduced GPT-5.3-Codex-Spark, its first real-time coding model, now available in research preview for ChatGPT Pro users. The model emphasizes speed and context capacity, claiming a 15x generation boost and support for a 128k context window.
OpenAI has partnered with chipmaker Cerebras to add 750MW of high-speed AI compute, aiming to reduce inference latency and accelerate ChatGPT's real-time workloads. This move represents a strategic infrastructure collaboration focused on performance optimization.