OpenAI GPT-5.4 Release Notes: Key Features, Benchmarks and 1M Token Context

On March 5, 2026, OpenAI released GPT-5.4 — simultaneously in ChatGPT, API, and Codex.

This is not just another incremental update: the model for the first time combines the GPT-5.3-Codex coding pipeline

with general reasoning, gains native computer use, and a context window of up to 1M tokens.

In short: if you are building agentic workflows or coding tools —

this is a release worth paying attention to today.

📚 Table of Contents

• 📌 What was released and when
• 📌 3 main changes for developers
• 📌 Quick comparison with competitors
• 📌 What to do right now
• 📌 Want to go deeper?

🗓️ What was released and when

OpenAI officially announced GPT-5.4

on March 5, 2026. The model is immediately available across three surfaces:

• ChatGPT — as GPT-5.4 Thinking for Plus, Team, and Pro users (replaces GPT-5.2 Thinking). GPT-5.2 Thinking remains in Legacy Models until June 5, 2026
• API — endpoints gpt-5.4 and gpt-5.4-pro are available now
• Codex — becomes the default model, replacing GPT-5.3-Codex

GPT-5.4 Pro is available via API and for ChatGPT Pro ($200/month) and Enterprise plans.

Free users gain access to GPT-5.4 through query auto-rotation, according to

VentureBeat.

⚙️ 3 main changes

1. No longer need to choose between GPT-5.x and Codex

Before the GPT-5.4 release, the standard architecture for an agentic pipeline with mixed tasks

looked like this: GPT-5.2 for planning and reasoning steps, GPT-5.3-Codex for generation

and code execution. Each switch between models meant a separate API call, separate context management,

different behavior in edge cases, and different fine-tuning parameters.

For long agent trajectories, this accumulated into significant overhead in terms of latency and

code complexity.

GPT-5.4 eliminates this need. According to

OpenAI,

this is the first mainline reasoning model that incorporates frontier coding capabilities

of GPT-5.3-Codex into unified weights — a result of merging training stacks, not routing logic.

In practice, this means:

• SWE-Bench Pro: 57.7% vs 56.8% in GPT-5.3-Codex — GPT-5.4 reproduces

  the coding performance of the Codex model with lower latency and additional reasoning capabilities,

  according to gaga.art

• GDPval: 83.0% — a new OpenAI metric, 44 professions from 9 industries,

  1320 tasks from domain specialists with 14+ years of experience. GPT-5.4 surpasses

  GPT-5.2 (70.9%) and matches or outperforms a human domain specialist in 83%

  of comparisons, according to

  The Decoder

• Practically for developers: if your pipeline used two endpoints,

  now it's enough to change the model ID to gpt-5.4 — in most cases

  this is a swap without logic changes. GPT-5.4 becomes the default model in Codex, replacing

  GPT-5.3-Codex automatically

Separately, it's worth noting a new feature in ChatGPT Thinking: the model now shows a reasoning plan

before execution and allows to correct the direction mid-response —

no need to start the query from scratch if the model went in the wrong direction. Available

on chatgpt.com and Android, iOS — coming soon, according to

DataCamp.

2. Native computer use: mechanics and real figures

GPT-5.4 is OpenAI's first general model with built-in computer use. It's important to understand

the architecture: it's not a single mechanism, but two parallel approaches that the model combines

depending on the task:

• Code-based automation — the model writes code using Playwright or similar

  libraries to control browsers and desktop applications. Suitable for deterministic,

  repeatable workflows: forms, navigation, data extraction

• Screenshot-based control — the model receives a screenshot of the current screen state

  and issues mouse/keyboard commands. Suitable for tasks where the UI structure is unpredictable

  or changes between sessions

Behavior is steered via developer messages and custom confirmation policies:

developers can configure which actions require user confirmation, and which

are executed autonomously — an important mechanism for production deployments with varying levels

of risk, according to

OpenAI.

Key benchmarks:

• OSWorld-Verified: 75.0% — above the average human score (72.4%).

  For comparison: GPT-5.2 on the same benchmark showed only 47.3% — meaning an increase

  of more than 1.5×, according to

  VentureBeat

• BrowseComp: 82.7% (base) / 89.3% (Pro) —

  measures the agent's ability to find hard-to-reach information on the internet through

  persistent browsing. GPT-5.2 showed 65.8% — an increase of 17% absolute points

To demonstrate its capabilities, OpenAI released an experimental Codex skill

Playwright (Interactive): the model can visually debug web and Electron

applications in real-time — and even test the application during its creation.

According to

DataCamp,

this combination of code generation and visual feedback loop points to a direction where AI agents

will be able to iterate on frontend with minimal human involvement.

3. Tool Search: from static manifest to on-demand discovery

This is perhaps the most practically important change for developers building systems

with a large number of tools. Previously, passing tool definitions into the system prompt

was inefficient: all schemas were loaded into context with each call,

regardless of whether they were needed at a specific step.

GPT-5.4 solves this through a new architecture: the model receives only a lightweight

list of available tools, and loads full definitions on-demand

only when it decides to use a specific tool. According to

The Decoder,

large tool ecosystems previously added tens of thousands of unnecessary tokens

to each request.

Practical effect of Tool Search:

• −47% tokens on agentic tasks with a large number of tools,

  according to

  VentureBeat

• Scalability: tool search allows working with ecosystems

  containing tens of thousands of tools — for example, corporate

  MCP servers or large API catalogs, according to

  Apidog

• Cache hit rate: since the lightweight tool list is more stable between

  requests than the full manifest, caching works more efficiently — further reducing

  inference cost

• Limitations: available exclusively via Responses API, not via

  Chat Completions

Separately, it's worth noting the improvement in accuracy: on a set of de-identified prompts,

where users previously noted factual errors, GPT-5.4 shows

−33% erroneous statements and −18% responses with any

errors compared to GPT-5.2, according to

OpenAI.

For production systems where accuracy is critical (legal analysis, financial calculations),

this is a measurable improvement in reliability.

📊 Quick comparison with competitors

Current as of March 2026. Sources: Digital Applied, OpenAI, gaga.art.

💡 Full comparison with 11 parameters, inference cost analysis, and practical hierarchy model → GPT-5.4: Architectural breakdown for developers

✅ What to do right now

If you have an agentic workflow or coding pipeline

• Swap model ID to gpt-5.4 and run your evals.

  If you previously used GPT-5.3-Codex — GPT-5.4 reproduces its SWE-Bench Pro

  (57.7% vs 56.8%) with lower latency. If you used GPT-5.2 — expect

  improvements on coding tasks without reasoning degradation

• Consider migrating to Responses API if you use Chat

  Completions with a large number of tools. Responses API unlocks Tool Search

  (−47% tokens), CoT between turns, and native compaction — three features unavailable

  via Chat Completions

• Enable /fast mode in Codex for tasks where speed is

  critical: the same GPT-5.4, but up to 1.5× faster token velocity, according to

  target="_blank">VentureBeat

• For 1M context window in Codex configure

  model_context_window and model_auto_compact_token_limit

  in Codex settings. Important: requests beyond the standard 272K are priced

  at 2× the normal rate, according to

  gaga.art

If you are building computer use agents

• Use the updated computer tool in the API. In OpenAI's documentation

  there are recommendations for original and high image detail settings —

  they significantly improve localization and click accuracy

• Configure custom confirmation policies for actions with different risk levels:

  define which operations are performed autonomously, and which require confirmation from

  the user before execution

• Try Playwright (Interactive) in Codex for visual debugging

  web and Electron applications — an experimental skill, but already functional for real

  frontend tasks

If you have simple high-throughput tasks

• Do not migrate in a hurry — gpt-5-mini or gpt-5.3-chat-latest remain

  the better choice for cost/latency for classification, summarization, and template-filling.

  GPT-5.4 will be overkill and more expensive for these scenarios

• GPT-5.2 in the API has no announced deprecation date — so

  legacy systems can be left untouched for now

Key Dates

• June 5, 2026 — GPT-5.2 Thinking is disabled in ChatGPT

  (moves to Legacy Models now, full disablement in 3 months).

  If you use it in a product via the ChatGPT interface — migrate before this date

• August 26, 2026 — Assistants API sunset. If you are still using

  Assistants API — migration to Responses API is a priority task right now

🔬 Want to understand how it works?

This article is a brief overview of what was released. If you are interested in the engineering mechanics:

how the reasoning pipeline changed from GPT-5.0 to 5.4, why a consolidated model is

an architectural compromise, and how reasoning.effort affects cost

and latency — read the detailed breakdown:

👉

GPT-5.4 in 2026: from specialized models to consolidated architecture — what changed and why

14 min read · 5 sections · benchmarks · tables · FAQ

Sources:

OpenAI — Introducing GPT-5.4

TechCrunch — OpenAI launches GPT-5.4

VentureBeat — GPT-5.4 native computer use

Digital Applied — GPT-5.4 vs Claude vs Gemini

OpenAI Academy — GPT-5.4 Thinking and Pro