The deprecations page says an alternative is 5.4-Pro which costs 3x the input and 4.5x the output. Additionally, how can we even have the same quality deep research with 5.4-Pro? I couldn’t find any code examples.
Are there any other alternatives?
I don’t know that Pro is required for this use, or would create much improvement for its 12x cost. It uses test-time compute, but on the API is not automatically a “use more tool calls” or “much higher reasoning Juice”, nor “write at length”. What is assumed is that multiple parallel token runs are analyzed for low perplexity. Use of Pro at $180/M with non-stop tool calling and repeated context could result in extreme billing.
The thing about Deep Research is its fine tuning or internal prompting has it persist at web exploration, and produce huge outputs.
You can offer an agentic AI model the web search tool, and prompt it up, but the feature of Deep Research vs OpenAI’s tool is that it doesn’t have text injected immediately after web tool return about what limited output to write, to damage applications, stop accumulating knowledge, and to override developer instructions with a search product as the only language. It has its own behavior to inspire full reports at length.
This is likely a “take away API, now use our do-all Codex product at $200” rent-seeking. Not that you could be competitive with ChatGPT before, anyway.
Also, being given the axe is Chat Completions web search models, that you could independently use to generate query web results - but then run those as tool calls with an agentic surface more free to write what you prompt.