OpenAI's Code Red: stop overpaying for one AI model
Yuri VonchitzkiUpdated 4 min readAt the start of December 2025, I saw something I had not seen since 2023: OpenAI was on the defensive. Reports confirmed that Sam Altman issued an internal "Code Red" at OpenAI. The order was to focus fully on ChatGPT improvements, because Google's Gemini 3 beat OpenAI on several benchmarks and Gemini grew from 450 million to 650 million monthly users.
OpenAI's Code Red is a green light for your budget. When the big AI vendors fight for customers, prices drop. If your automations depend on one vendor's model, you probably overpay. Audit your API usage, then move high-volume tasks such as classification and formatting to cheaper models like DeepSeek V3.2 or Gemini Flash.
Why the safe bet became a liability
For two years, many of you built your full AI stack on OpenAI. It was the safe bet. In December 2025, that safe bet became a liability.
My clients started to ask me a direct question. Why do they pay for OpenAI's "thinking tokens" when Gemini 3 can process their 50-page SOP documents for half the price? Thinking tokens are the tokens a reasoning model generates internally before it gives its answer, and you pay for them.
The numbers from December 2025 show the size of the gap:
| Model | What it offered in December 2025 |
|---|---|
| Google Gemini 3 Pro | A 1M token context window and input pricing of $2 per million tokens (prompts up to 200k tokens) |
| DeepSeek V3.2 | Input pricing of $0.28 per million tokens (cache miss), and DeepSeek says it performs at GPT-5 level |
A model at $0.28 per million input tokens makes background intelligence close to free. That price suits high-volume data-cleaning agents that run 24/7. For the bigger picture on falling prices, read the cost of AI reasoning just dropped.
From brand loyalty to model agnosticism
The market is moving from brand loyalty to model agnosticism. Model agnosticism means you treat the language model as a replaceable part of your automation, not as a fixed dependency. You choose the model for each task, based on cost and fit.
If you still hard-code gpt-4o into your automations, you overpay. You also lose the ability to profit when the next price drop comes. For simple, repetitive work, the commodity models have caught up.
How to cut your AI model costs: a 4-step audit
1. List every automation that calls an AI model. Open your automation platform and find each scenario that sends data to an AI API. Write down the model name and the task.
2. Mark the high-volume tasks. Classification and formatting are the usual candidates. These tasks run often, and they do not need a premium reasoning model.
3. Move those tasks to a cheaper model. Switch the high-volume tasks to DeepSeek V3.2 or Gemini Flash. Keep the premium models for the work that needs them, for example long documents or complex reasoning.
4. Stop hard-coding the model name. Store the model choice as a setting in one place. When the next model drops its price, you change one value instead of every scenario.
To see what the savings mean for your P&L, run the numbers in the ROI calculator. If you want a structure that makes models easy to replace, the Automation Architecture Guide shows how to name the pattern before you build.
The bottom line
OpenAI's Code Red is a sign of fierce competition, and fierce competition lowers prices. Audit your API usage now. Stop paying a premium for a brand name when the commodity models do the same job.
Build a calmer business,Yuri
Frequently asked questions
Why did OpenAI declare a Code Red in December 2025?
In early December 2025, reports said that Sam Altman issued an internal Code Red at OpenAI. The goal was to focus the company fully on ChatGPT improvements, because Google's Gemini 3 beat OpenAI on several benchmarks and Gemini grew from 450 million to 650 million monthly users. For businesses, the signal is that OpenAI is no longer the only safe default and that competition between model vendors pushes prices down.
Is it still smart to build all business automations on OpenAI models?
Building every automation on one AI vendor became a liability by December 2025. Clients asked why they paid for OpenAI's thinking tokens when Gemini 3 could process a 50-page SOP document for about half the price. Yuri Vonchitzki of Ninjabot advises a model-agnostic setup: pick the model per task and keep the model name out of hard-coded automations, so a cheaper model can replace it.
Which AI model should I use for high-volume tasks like classification and formatting?
For high-volume, simple tasks such as classification and formatting, Ninjabot recommends cheaper commodity models like DeepSeek V3.2 or Gemini Flash instead of a premium reasoning model. In December 2025, DeepSeek V3.2 cost $0.28 per million input tokens, and DeepSeek said it performed at GPT-5 level. Paying premium prices for these tasks only pays for a brand name.
What is model agnosticism in AI automation?
Model agnosticism means you design AI automations so that the language model is a replaceable part, not a fixed dependency. You choose a model for each task based on cost and fit, not on brand loyalty. When a competitor releases a cheaper or better model, you switch that one step instead of rebuilding the automation.