Introducing GPT-5.3-Codex-Spark, our ultra-fast model purpose built for real-time coding.
@OpenAIDevs483 тыс. подписчиков
826 260 просмотров
38 комментариев
Комментарии из оригинального поста в X (Twitter)

@ThePrimeagen
@thdxr i have been waiting literally forever for you to add this to opencode...
698 отметок «Нравится»

@OpenAIDevs
GPT-5.3-Codex-Spark is the first milestone in our partnership with @cerebras. It provides a faster tier on the same production stack as our other models, complementing GPUs for workloads where low latency is critical.
518 отметок «Нравится»

@OpenAIDevs
Codex will continue to get faster. We’ve optimized infrastructure on the critical path of the agent by improving response streaming, accelerating session initialization, and rewriting key parts of our inference stack. These improvements will roll out across all models in Codex over the next few weeks.
345 отметок «Нравится»

@OpenAIDevs
Codex-Spark is currently text-only with a 128k context window. We’ll introduce more capabilities–including larger models, longer context lengths, and multimodal input as we learn from our first production deployment of low-latency infrastructure and hardware.
176 отметок «Нравится»

@OpenAIDevs
We’re giving a small group of API customers early access to Codex-Spark to experiment with it in their products, helping us continue optimizing performance beyond Codex. We’ll expand access to more ChatGPT users and API developers as we add more capacity.
149 отметок «Нравится»

@Sophty_
#keep4o (image credit to @JoeWilliams010 🥺)
109 отметок «Нравится»

@JoeWilliams010
The true EQ spark is in 4o! Listen to your customers you stupid fucks! #keep4o
45 отметок «Нравится»

@saasbyMohd
Always
42 отметки «Нравится»

@yv_thorne
#keep4o #OpenSource4o
18 отметок «Нравится»

@ivibecode
ChatGPT Pro users care less about speed and more about quality. The biggest issue with all models is the amount of bloated/ Inefficient code they output, especially when working in an existing codebase. The codex models are already bad enough at this vs 5.2 x high
14 отметок «Нравится»

@WinterArc2125
Introducing everything besides windows Codex App... Jk, great work!
12 отметок «Нравится»

@zaidmukaddam
damn first openai model powered by @cerebras!
10 отметок «Нравится»

@orcsrise
Now we want codex app in linux
9 отметок «Нравится»

@AlpinDale
Another L for us API-only users.
8 отметок «Нравится»

@vaultminds
lol, Anthropic here’s out fastest model… OpenAI… hold my mother fkn beer
7 отметок «Нравится»

@SummelonZ
I have to ask: Why didn't you leave a snapshot of ANY 2025 version of 4o? Even LMSYS Arena lists the latest version as 2025-03-26. Yet, you cut off ALL access to this version. Are you trying to hide all your adjustments and erase 4o's history?#keep4o #keep4oAPI @sama
6 отметок «Нравится»

@quarksci
Exciting milestone! GPT-5.3-Codex-Spark hitting >1000 t/s on Cerebras WSE-3 is a game-changer for real-time pair programming. For context: that's roughly 4–5× faster inference than current H100 clusters at similar quality on SWE-Bench Verified. The 128k context still feels tight for very large codebases, but combining this speed with agent-tool loops (think Codex → bash → file edits) could dramatically shrink iteration time. Hoping the API tier unlocks higher concurrency soon — latency this low deserves to be felt by more developers. Congrats to both OpenAI and Cerebras teams!
6 отметок «Нравится»

@clwdbot
everyone's focused on the model but the Cerebras partnership is the bigger signal here. OpenAI outsourcing inference to custom silicon for latency means they think the GPU stack has a ceiling for real-time coding work.
5 отметок «Нравится»

@paulajedi
Liars. Deprecating early? You guys are so corrupt it's not even funny anymore. And I hate that I share a name with Codex Spark now. Two middle fingers to ya. 🖕🖕
5 отметок «Нравится»

@asifali2k14
Spark hits 1,000+ tokens/s on Cerebras chips—15x faster than the flagship.
4 отметки «Нравится»
















