OpenAI pulls GPT-5.3-Codex-Spark next week — new model to replace it
The fast, Pro-only Codex variant goes dark within days; a better model is coming. Plus Ant's free 124B open multimodal, and RTK's token-savings claims tested.
Copy markdown
Spark goes offline within days
OpenAI Codex lead Tibo said on Sept 11 that GPT-5.3-Codex-Spark — the low-latency, Pro-only coding model — will be taken offline next week, citing declining usage and a stronger model on the way. If you route Codex jobs to Spark for speed, line up a replacement (gpt-5.3-codex stays) before it vanishes.
Keep a doomed model with your own API key
OpenAI's standing workaround for anyone who needs a model past its retirement is to run Codex with your own API key instead of ChatGPT sign-in. Worth wiring up now if Spark sits anywhere in your pipeline.
Ant open-sources Ling-3.0-Flash-VL
Ant Group added vision — image and video input — to its 124B open-weights MoE, which activates just 5.5B params per token and scores 38 on the Artificial Analysis Intelligence Index. It's free to download: a self-hostable multimodal reasoner sitting on the intelligence-vs-size efficiency frontier.
RTK's token savings don't shrink the bill
RTK wraps your CLI commands to cut the tokens fed to coding agents like Claude Code, advertising 60–90% fewer tokens. Independent benchmarks from Quesma found that token cut didn't translate into a lower dollar bill — measure on your own workload before trusting the headline number.