OpenAI rolls out ultra-fast Codex-Spark for real-time coding

The biggest bottleneck in AI coding isn't always model intelligence—it's waiting for the tokens to stream in. Staring at an editor while an LLM slowly rewrites a function breaks developer flow.

OpenAI just launched a research preview of GPT-5.3-Codex-Spark, a lightweight version of its flagship coding model built specifically for real-time collaboration. Clocking in at over 1,000 tokens per second, it delivers a near-instant interactive experience.

The release marks the first product milestone of OpenAI's multi-year, $10 billion-plus partnership with hardware maker Cerebras announced last month.

Giant chips and a rewritten network stack

To hit four-digit token speeds, Codex-Spark runs on Cerebras' Wafer Scale Engine 3 (WSE-3), a massive wafer-scale megachip packing 4 trillion transistors engineered for ultra-low-latency inference.

Cerebras, which raised $1 billion at a $23 billion valuation last week ahead of a planned IPO, provides the low-latency tier while standard GPUs continue handling long-running, bulk workloads.

OpenAI didn't just swap the chips. It also completely revamped the client-server pipeline to cut down lag before the first token even renders.

By implementing persistent WebSocket connections and streamlining the Responses API, OpenAI slashed client-server roundtrip overhead by 80%, per-token overhead by 30%, and time-to-first-token by 50%.

At launch, Codex-Spark supports a 128k context window and is text-only. In benchmarks like SWE-Bench Pro and Terminal-Bench 2.0, OpenAI claims Spark achieves solid agentic scores in a fraction of the time required by full-sized models.

Its behavior is deliberately lightweight: Spark focuses on surgical edits and skips running tests automatically unless explicitly instructed.

ChatGPT Pro users can test it immediately across the Codex desktop app, CLI, and VS Code extension under separate rate limits that don't eat into standard caps. A limited group of design partners gets API access today.

Why it matters

While AI labs have spent the past year obsessing over deep-reasoning models that run autonomously for hours, daily software engineering is mostly an endless loop of micro-adjustments. Waiting even five seconds for an autocomplete suggestion ruins the tool.

For developers, Codex-Spark is worth trying right away if you live in VS Code. At 1,000 tokens per second, UI refactoring and logic tweaks feel like local linters rather than remote server requests.

For the industry, OpenAI validating Cerebras with real production traffic is a major blow to the GPU monoculture. GPUs remain the cost king for throughput, but Cerebras just proved that dedicated inference hardware owns ultra-low latency.

The clear risk is scaling. Wafer-scale hardware is notoriously capital-intensive and scarce. OpenAI is already warning users about potential queues during peak hours, and rolling this out beyond Pro subscribers will be a massive datacenter hurdle.

Still, pairing a lightning-fast interactive front-end with heavy background sub-agents looks like the future of agentic workflows.