OpenAI Deep Research: API-Driven Reasoning, Deprecations, and the GPT-5.4 Pivot
OpenAI's programmatic deep research capabilities have undergone a rapid structural transition. After releasing dedicated deep research models in the API—namely o3-deep-research and o4-mini-deep-research—OpenAI announced their deprecation on April 22, 2026, with a final service termination date set for October 23, 2026. In their place, OpenAI has positioned the newly launched GPT-5.4 and GPT-5.4 Pro as its flagship reasoning, computer-use, and deep web research engines.
This transition marks a shift from specialized, single-purpose search-and-synthesis models to general-purpose agentic models with native web browsing, tool search, and desktop control capabilities.1
The Rise and Deprecation of o3 and o4-mini Deep Research
Initially launched in the API to allow developers to programmatically trigger deep web research and synthesis, the dedicated deep research models were highly capable but carried high costs and rate limit complexities.2 For instance, o3-deep-research was priced at $10/$40 per million input/output tokens, while its lightweight counterpart, o4-mini-deep-research, offered a more affordable $2/$8 per million token structure.
Developer Simon Willison documented his experience with the mini variant:
"o4-mini-deep-research is significantly cheaper than o3-deep-research - $2/$8 per million for input/output compared to $10/$40 - so I tried that one... Total cost: 77 + 30.4076 + 3 = 110.4076 cents, $1.10." — Exploring OpenAI's deep research API model o4-mini-deep-research
With the deprecation announcement of these models, developers expressed frustration regarding the pricing and complexity of migrating to the recommended successor, gpt-5.4-pro:
"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." — Linda51, OpenAI Developer Community
The GPT-5.4 and GPT-5.4 Pro Frontier
Launched in mid-2026, GPT-5.4 and its premium variant GPT-5.4 Pro represent OpenAI's most advanced models for professional services and agentic workflows. They feature a 1M token context window, native computer-use capabilities via the computer tool, and a novel "tool search" feature that dynamically appends tool definitions to the prompt to conserve context and reduce latency.
For deep research workflows, GPT-5.4 Pro excels at agentic web search, achieving a record-breaking 89.3% on the BrowseComp benchmark (an increase over GPT-5.2's 65.8% and GPT-5.3's 77.3%):
"Today, we’re releasing GPT‑5.4 in ChatGPT (as GPT‑5.4 Thinking), the API, and Codex. It’s our most capable and efficient frontier model for professional work. We’re also releasing GPT‑5.4 Pro in ChatGPT and the API, for people who want maximum performance on complex tasks." — Introducing GPT-5.4
This consolidated architecture allows developers to build deep research agents that not only browse the web and synthesize reports but also interact with local software, click coordinate-based elements, and verify their outputs in real time.
-
An instance of Automating deep research requires extended reasoning compute, not instant data retrieval. — OpenAI's retirement of dedicated, single-turn research APIs in favor of multi-modal models like GPT-5.4 reflects a shift toward complex, multi-touch reasoning execution. ↩︎
-
An instance of Long-horizon agentic loops collapse under the unpredictable cost of reasoning tokens. — Early deep-research APIs are being retired due to the severe pricing, cost overhead, and architectural difficulty of maintaining specialized reasoning loops. ↩︎