GPT-5.2 Is Being Retired: What Builders Get When They Switch to GPT-5.4

GPT-5.2 — including GPT-5.2 Thinking — was retired from ChatGPT on June 12, 2026; existing GPT-5.2 conversations there now continue automatically on GPT-5.5. On the API side, OpenAI’s deprecations page separately lists gpt-5.2-codex shutting down July 23, 2026, and gpt-5.2-chat-latest / gpt-5.3-chat-latest shutting down August 10, 2026 — with gpt-5.6-sol as the listed replacement for both.

If your production code still calls a GPT-5.2 model, the practical move is the same regardless of which exact snapshot you’re on: update the model string before its shutdown date arrives. This piece looks at why gpt-5.4 — released March 5, 2026 — is worth choosing for that update even though it isn’t OpenAI’s own default mapping: three capabilities it has that GPT-5.2 did not, and that directly affect how builders structure agent pipelines.

The One-Line Migration

Change the model string:

# Before
response = client.responses.create(
    model="gpt-5.2-thinking",
    ...
)

# After
response = client.responses.create(
    model="gpt-5.4",
    ...
)

That is the minimum viable migration. GPT-5.4 is backward-compatible with the Responses API surface that GPT-5.2 used. Existing prompts, tool definitions, and structured output schemas work without modification.

If you are still on the older Chat Completions API format, GPT-5.4 also supports that path, but OpenAI has signaled that new capabilities (Tool Search in particular) require the Responses API. The migration to client.responses.create() is worth doing in the same pass.

What GPT-5.4 Adds

GPT-5.4 was released on March 5, 2026, as OpenAI’s first general-purpose model with three capabilities that did not exist in GPT-5.2:

• Native computer-use — the model can interact with a desktop through screenshots, control the mouse and keyboard, and generate Playwright scripts for browser automation
• 1M-token context — standard context is 272K tokens; API and Codex users can configure up to 1M tokens
• Tool Search — a new mechanism for handling large tool inventories without loading every definition into the prompt

The benchmarks are also better: GPT-5.4 scores 75.0% on OSWorld-Verified desktop-automation tasks, edging past the 72.4% human-expert baseline — up from GPT-5.2’s 47.3% on the same benchmark. But for most migration decisions, the three structural capabilities above matter more than raw scores.

Native Computer-Use

GPT-5.4 is the first general-purpose OpenAI model that can operate a desktop. Previous computer-use capability was confined to the specialized Computer-Using Agent (CUA) model that powered Operator, OpenAI’s dedicated browser-automation product. With GPT-5.4, you can build agents that:

• Take a screenshot and reason about what is on screen
• Issue mouse and keyboard events directly
• Write and execute Playwright code for browser automation
• Operate across applications (not just the browser)

The model accepts screenshots as image inputs and outputs structured actions: {"type": "mouse_click", "x": 412, "y": 208} or {"type": "key", "key": "ctrl+c"}. It can also generate and run Playwright scripts autonomously when the browser automation path is preferred over raw input events.

This is relevant even for builders who do not intend to ship a computer-use product: computer-use is a plausible fallback for tools that lack an API — the model can click through a legacy web interface or fill in a form with no programmatic endpoint. GPT-5.4 makes this pattern available at standard API pricing without a specialized model swap.

1M-Token Context

The standard context window is 272K tokens. For most tasks, that is more than enough.

The 1M-token option is designed for long-horizon agents that need to plan, execute, and verify across a much larger scope: entire codebases, regulatory document review, multi-session conversation replay, or reasoning across many retrieved documents simultaneously.

The pricing structure reflects the cost — current rates from OpenAI’s pricing page:

Both input and output rates increase past the 272K threshold — input doubles, output rises 1.5x — and the higher rate applies to the full session, not just the tokens past 272K. The 1M window is not a free upgrade. For most production workloads, staying within 272K at standard pricing is the right call. The extended context makes economic sense when the task genuinely cannot be decomposed — or when the cost of multiple round-trips and retrieval errors exceeds the token price delta.

Tool Search

Tool Search is the most novel addition for builders who work with MCP or large tool inventories.

The problem it solves: as agents connect to more MCP servers, the combined token cost of loading every tool definition into every request becomes significant. A cluster of 36 MCP servers may carry tens of thousands of tokens of tool definitions before the first user message is processed.

OpenAI evaluated Tool Search against 250 tasks from Scale’s MCP Atlas benchmark with all 36 MCP servers enabled. The result: 47% fewer tokens consumed, with no accuracy loss compared to loading full definitions upfront.

How It Works

Instead of sending every tool definition in the tools array, you declare a list of available tools and add {"type": "tool_search"} to the array. The model receives lightweight descriptors (name, a short summary) for all tools, then fetches the full definition of a tool only when it decides to use it.

response = client.responses.create(
    model="gpt-5.4",
    tools=[
        {"type": "tool_search"},      # enable dynamic tool lookup
        {"type": "mcp", "server_label": "github", "server_url": "..."},
        {"type": "mcp", "server_label": "linear", "server_url": "..."},
        {"type": "mcp", "server_label": "slack", "server_url": "..."},
        # ... more MCP servers
    ],
    input="Triage the open GitHub issues from this sprint and post a summary to #eng-standup"
)

The API handles the tool lookup internally. When the model decides it needs the create_issue tool from the GitHub MCP server, it fetches the full schema for that tool and appends it to the context at that moment. Tools the model does not need never consume tokens.

Only gpt-5.4 and gpt-5.4-pro support tool_search, per OpenAI’s model documentation. It requires the Responses API — it is not available via Chat Completions.

When Tool Search Matters

Tool Search is high-value when:

• You connect four or more MCP servers simultaneously
• Your MCP servers expose large numbers of tools (tens or hundreds per server)
• Your agent queries are narrowly scoped relative to the full tool inventory (the agent only needs 2–3 tools per task, but 36 servers are registered)

Tool Search has minimal benefit when:

• You have a small, fixed tool set (< 10 tools, all definitions under 2K tokens combined)
• Your tasks genuinely use most available tools in each request
• You need deterministic, reproducible token counts for billing estimation

GPT-5.4 Pro

gpt-5.4-pro is also available — OpenAI describes it as built for people who want “maximum performance on complex tasks.” Pricing is higher: $30.00 / MTok input and $180.00 / MTok output as of this writing (check OpenAI’s pricing page for current rates, since Pro pricing has changed before). Use gpt-5.4 as the default unless you have a task where gpt-5.4 measurably underperforms and the Pro price delta is justified by the outcome difference.

Pricing Summary

Builder Action Checklist

If you are still calling any GPT-5.2 model from your own code:

• Search your codebase for gpt-5.2, gpt-5.2-codex, gpt-5.2-chat-latest, and gpt-5.3-chat-latest model strings
• Check OpenAI’s deprecations page for your exact snapshot’s shutdown date (gpt-5.2-codex: July 23, 2026; gpt-5.2-chat-latest / gpt-5.3-chat-latest: August 10, 2026)
• Replace with gpt-5.4 in all production call sites
• Run a smoke test against the Responses API if you are migrating from Chat Completions at the same time
• Update any hardcoded model references in environment variables or config files

Optional GPT-5.4 upgrades (do after migration is stable):

• If you have 4+ MCP servers registered: add {"type": "tool_search"} to your tools array and measure token reduction on your actual workload
• If you have tasks that would benefit from > 272K context: test against 1M token window, compare cost against decomposition alternatives
• If you have any legacy-UI automation tasks (apps without APIs): evaluate native computer-use as a replacement for your current brittle workarounds

Bottom Line

GPT-5.2 is already gone from ChatGPT as of June 12, 2026, and the API snapshots follow this summer — gpt-5.2-codex on July 23, gpt-5.2-chat-latest / gpt-5.3-chat-latest on August 10. Updating the model name takes five minutes. Leaving it until the shutdown date risks a production outage over a trivial change.

Once you have migrated, Tool Search is the feature most likely to reduce your monthly bill if you run MCP-heavy agents — 47% token reduction on MCP Atlas is not a marginal improvement. It requires no prompt changes, no architecture changes, and no additional API surface to learn. It is a one-line addition to your tools array that pays for itself on any workload where the model uses a small fraction of registered tools per request.