GPT-5.4 nano on streaming video
GPT-5.4 nano · OpenAI · Multimodal · Closed API
GPT-5.4 nano is the smallest, cheapest tier of OpenAI's GPT-5.4 family, released in March 2026 for tasks where speed and cost matter most: classification, data extraction, ranking, and subagent work. It keeps the family's 400K-token context window while pricing at a fraction of the larger tiers. On Overshoot it is available as a passthrough model under the api id gpt-5.4-nano, reachable through the same stream publishing and OpenAI-compatible chat-completions request shape used across the API.
GPT-5.4 nano is available through the Overshoot API as a passthrough model (gpt-5.4-nano, status: ready, verified against the live model catalog on 2026-07-14) - requests are forwarded to the upstream provider rather than running on Overshoot's hosted infrastructure.
- Developer
- OpenAI
- Parameters
- Undisclosed
- Context window
- 400K tokens
- Max output
- 128K tokens
- License
- Proprietary
- Released
- Mar 2026
- Knowledge cutoff
- Aug 2025
- Inputs
- Text, images, video frames
- Overshoot availability
- API passthroughas of 2026-07-14
What GPT-5.4 nano is good at
GPT-5.4 nano targets high-volume, latency-sensitive work: classifying what a frame shows, extracting structured fields from a document or screenshot, ranking candidates, and running as a cheap subagent inside a larger pipeline. OpenAI positions it as a significant upgrade over GPT-5 nano at the same job, with pricing of $0.20 per million input tokens and $1.25 per million output tokens.
Despite its size, it inherits the full 400K-token context window of the GPT-5.4 family and supports tool calling and reasoning tokens, so an application can carry long frame history or conversation state in a single thread while still paying nano-tier rates per call.
The tradeoff is reasoning depth. GPT-5.4 nano is tuned for responsiveness rather than complex multi-step analysis, so it is the tier to pick when queries are well-scoped and repeated, not when a single answer needs careful deliberation.
- Frame classification and structured extraction at very low per-call cost
- Ranking and routing decisions inside agentic pipelines
- High-volume visual checks where throughput matters more than depth
Running GPT-5.4 nano on Overshoot
Publish a camera or screen share over WebRTC through LiveKit to create a Stream, then send a chat-completions request whose image_url or video_url is an ovs:// reference: the latest frame, an exact timestamp, or a bounded recent segment drawn from the 600-second retention window.
GPT-5.4 nano reaches the API as a passthrough model, so Overshoot manages the stream lifecycle and frame selection, then forwards the resolved request to OpenAI through the OpenAI-compatible endpoint and streams the response back over SSE. The request and response shape match Overshoot-hosted models exactly, so swapping the model id is the only change needed to trial it.
Because nano-tier calls are cheap, it pairs naturally with a polling pattern: query the latest frame on an interval for a lightweight status check, and escalate to a larger model only when the nano-tier answer flags something worth deeper analysis.
GPT-5.4 nano versus GPT-5.4 mini
GPT-5.4 mini carries more reasoning capacity and costs more per token; nano trades that depth for the lowest price and fastest responses in the family. For live video, that makes nano the right default for continuous monitoring, simple visual question answering, and extraction jobs that run on every sampled frame.
Both tiers share the 400K context window and identical request format on Overshoot, so a pipeline can mix them freely: nano for the high-frequency loop, mini or full GPT-5.4 for the occasional hard query, all against the same stream and thread structure.
Latency expectations on the passthrough path
Overshoot's roughly 200ms response figure applies to models on its hosted fast path. GPT-5.4 nano is a passthrough model, so end-to-end latency depends on OpenAI's serving stack plus the forwarding hop, though nano is the fastest-responding tier OpenAI offers in the GPT-5.4 line.
For applications that need the tightest loop on live video, an Overshoot-hosted open-weight model is the lower-latency choice, with GPT-5.4 nano as a strong passthrough option when OpenAI's instruction following and tool calling are worth the extra round trip.
Frequently asked questions
Is GPT-5.4 nano available on Overshoot?
Yes. GPT-5.4 nano is in the live Overshoot catalog as a passthrough model with ready status under the api id gpt-5.4-nano. Requests are forwarded to OpenAI through the OpenAI-compatible API, and you can confirm current status any time via GET /v1beta/models.
Can GPT-5.4 nano analyze live video?
Yes. Through Overshoot, GPT-5.4 nano answers questions about a live WebRTC stream using ovs:// frame references inside a standard chat-completions request. Overshoot resolves the reference to actual frames, forwards the request to OpenAI, and streams the answer back over SSE.
How is GPT-5.4 nano different from GPT-5.4 mini?
Nano is the smallest and cheapest GPT-5.4 tier, tuned for classification, extraction, ranking, and subagent tasks at $0.20 per million input tokens. Mini carries more reasoning depth at a higher price. Both share the 400K context window and the same request format on Overshoot.
How fast is GPT-5.4 nano on Overshoot?
GPT-5.4 nano is a passthrough model, so latency depends on OpenAI's serving stack rather than Overshoot's roughly 200ms hosted fast path. It is still the quickest tier in the GPT-5.4 family, which makes it a practical choice for frequent, lightweight queries against a stream.