Earlier this month, OpenAI introduced GPT-5.3-Codex, its most capable agentic coding model, which achieved record scores on the SWE-bench Pro and Terminal-Bench AI benchmarks. This new GPT-5.3-Codex model was initially made available only on Codex, OpenAI"s AI coding platform.
Now, OpenAI has made the GPT-5.3-Codex model available for other apps and services through its API platform. Microsoft has also announced that GPT-5.3-Codex is now available through Microsoft Foundry. Consequently, developers can now access the GPT-5.3-Codex API via OpenAI or Microsoft.
Thanks to optimizations by the OpenAI team, the GPT-5.3-Codex model experiences 25% faster execution time when compared to older models. Additionally, the model can now run for longer periods, making it more suitable for research, tool use, and complex, multi‑step execution. For the first time, GPT-5.3-Codex now supports mid-task steerability, allowing developers to redirect the model as it works without losing context. Finally, this new model also performs significantly better in computer-use capabilities when compared to older GPT-Codex models.
In terms of pricing, there is no change when compared to GPT-5.2-Codex. The model will cost $1.75 per million input tokens and $14 per million output tokens. Cached input will cost $0.175 per million tokens.
OpenAI also announced two new audio models: GPT-Realtime-1.5 and GPT-Audio-1.5. According to OpenAI, GPT-Realtime-1.5 delivers a 5% improvement on the Big Bench Audio benchmark, which measures the reasoning ability of the audio model. This model also delivers a 10% improvement on alphanumeric transcription and a 7% improvement on instruction following in internal evaluations.
The OpenAI team claims that this new, improved model will deliver smoother and more conversational audio output with improved pacing and prosody. Furthermore, the API now supports structured, tool‑driven interactions within real‑time audio flows.
Voice workflows just got stronger with gpt-realtime-1.5 in the Realtime API.
— OpenAI Developers (@OpenAIDevs) February 23, 2026
The model offers more reliable instruction following, tool calling, and multilingual accuracy.
Demo with @charlierguo pic.twitter.com/gGV57Wv91V
Both new audio models are now available on Microsoft Foundry as well. GPT-Realtime-1.5 is priced at $4.00 per text input ($0.04 cached) and $16.00 per text output, $32.00 per audio input ($0.40 cached) and $64.00 per audio output, and $4.00 per image input ($0.04 cached) and $16.00 per image output. GPT-Audio-1.5 costs $2.50 per text input and $10.00 per text output, $32.00 per audio input and $64.00 per audio output, and $2.50 per image input and $10.00 per image output.