In 30 seconds
Matthew Berman discusses GPT-5.3-Codex-Spark, introduced in February 2026 as a smaller, speed-oriented version of GPT-5.3-Codex. He highlights OpenAI's partnership with Cerebras and the advertised generation rate. His broad claims about being the fastest hardware or inference on Earth are not established by the clip.
- 1A speed-focused modelBerman says GPT-5.3-Codex was capable but slow and introduces the smaller Spark variant.
- 2Cerebras hardwareOpenAI serves Spark on Cerebras wafer-scale hardware. The sweeping claim that its chips are faster than every other chip is not established here.
- 3Output speedBerman highlights over 1,000 tokens per second, matching OpenAI's advertised generation rate; end-to-end task time also includes other work.
Takeaways
- GPT-5.3-Codex-Spark was introduced as a smaller model designed for fast, interactive coding.
- The model runs on Cerebras Wafer Scale Engine 3 hardware.
- OpenAI advertised more than 1,000 output tokens per second, a generation rate rather than a guarantee for every workflow.
- The video does not substantiate a universal ranking of all chips or inference services.
Read the edited text by chapter
Faster coding inference
Berman introduces GPT-5.3-Codex-Spark as a smaller model optimized for speed. He connects it to OpenAI's Cerebras partnership and its advertised output rate.
GPT-5.3-Codex came out about a week ago and, in my view, changed coding. But it had a major flaw: it was slow. OpenAI has now announced GPT-5.3-Codex-Spark, a smaller version tuned for real-time coding. The original clip ends during the speed discussion.
This is OpenAI's first model deployed on Cerebras hardware, following a partnership between the companies. Cerebras makes wafer-scale chips, and Spark is designed to run quickly on them. I describe the hardware and inference as the fastest on Earth, but this sweeping comparison is not demonstrated in the video.
OpenAI advertised more than 1,000 output tokens per second for Codex-Spark. That is a model generation rate under the provider's stated setup, not a promise that every complete coding task will finish at that speed; tools, input processing and network overhead also affect the experience.
More from this channel
14:45EN
Matthew Berman· Oct 29, 2026Matthew Berman Examines Google’s Gemini 4 Carbon AnnouncementMatthew Berman · Forward FutureMatthew Berman reviews Google’s reported benchmark results, staged rollout and pricing for Gemini 4 Carbon, while noting he has not yet used the model.
21:11ENMatthew Berman· Oct 28, 2026Matthew Berman tests DeepSeek V4.1 Flash: coding, vision and local appsMatthew Berman · Forward FutureA hands-on tour of visual coding, image reasoning, local applications and provider economics reveals both impressive results and visible flaws.
34:21ENMatthew Berman· Oct 27, 2026Matthew Berman Examines Anthropic’s IPO Filing, AI Growth and GovernanceMatthew Berman, Sam Altman · Forward FutureThe presenter reviews the S-1’s financial metrics, compute strategy and voting structure while questioning the tension between rapid AI development and safety.
0:26ENMatthew Berman· Oct 25, 2026Sam Altman on Human Values and Problem-Solving in an AI WorldSam Altman · OpenAIIn a short interview excerpt, Altman considers the enjoyment of solving hard problems and the human role in interpreting results.
0:36ENMatthew Berman· Oct 25, 2026Kimi K3.1 and a Rubik’s Cube Solver DemonstrationMatthew Berman · Forward FutureMatthew Berman shows a scrambled virtual cube being solved and gives the Kimi K3.1 test a favorable assessment.
0:21ENMatthew Berman· Oct 24, 2026An AI-Generated Kimi Benchmark Chart Used Invented ScoresMatthew Berman · Forward FutureMatthew Berman asks for a graph of Kimi K3.1 Pro benchmarks, then discovers that the displayed figures are illustrative placeholders.