GPT-5.4

GPT-5.4 launched today — March 5, 2026. It is OpenAI’s most capable model, the first with native computer-use capabilities, and the new default across ChatGPT, the API, and Codex. It replaces GPT-5.2 Thinking and GPT-5.3-Codex in a single unified model.

What is new

GPT-5.4 is not an incremental update. It introduces three capabilities that no previous GPT model had:

Native computer use. GPT-5.4 can operate software through screenshots and keyboard/mouse commands — browsing websites, filling forms, navigating desktop environments. On OSWorld-Verified (desktop navigation), it scores 75.0%, surpassing human performance at 72.4%. This is not a plugin or tool — it is built into the model.

Tool search. Previous models loaded all tool definitions into context upfront. GPT-5.4 receives a lightweight index and loads only the definitions it needs at runtime. On Scale’s MCP Atlas benchmark with 36 MCP servers enabled, this reduced token usage by 47% with no accuracy loss. This matters for anyone building agents with large tool ecosystems.

1M token context. The context window extends to 1.05M tokens (up from 272K on GPT-5.2). Prompts exceeding 272K tokens are priced at 2x input / 1.5x output. This enables analysis of entire codebases, long document collections, or extended agent trajectories in a single request.

Native compaction. GPT-5.4 is the first model trained to support compaction — summarizing earlier context to extend effective agent trajectories without losing key information.

Benchmarks

OpenAI published extensive benchmarks comparing GPT-5.4 against GPT-5.3-Codex and GPT-5.2:

Professional work

Benchmark	GPT-5.4	GPT-5.3-Codex	GPT-5.2
GDPval (matches/exceeds professionals)	83.0%	70.9%	70.9%
Investment Banking Modeling	87.3%	79.3%	68.4%
OfficeQA	68.1%	65.1%	63.1%

GDPval tests agent performance across 44 occupations — sales presentations, accounting spreadsheets, urgent care schedules, manufacturing diagrams. GPT-5.4 matches or exceeds industry professionals in 83% of comparisons.

Coding

Benchmark	GPT-5.4	GPT-5.3-Codex	GPT-5.2
SWE-Bench Pro (Public)	57.7%	56.8%	55.6%
Terminal-Bench 2.0	75.1%	77.3%	62.2%

GPT-5.4 matches GPT-5.3-Codex on coding while being lower latency. Codex /fast mode delivers 1.5x faster token velocity.

Computer use and vision

Benchmark	GPT-5.4	GPT-5.2	Human
OSWorld-Verified	75.0%	47.3%	72.4%
WebArena-Verified	67.3%	65.4%	—
Online-Mind2Web	92.8%	—	—
MMMU-Pro (no tools)	81.2%	79.5%	—

The jump from 47.3% to 75.0% on OSWorld is the headline number — GPT-5.4 now outperforms humans on desktop navigation tasks.

Tool use

Benchmark	GPT-5.4	GPT-5.2
BrowseComp	82.7%	65.8%
MCP Atlas	67.2%	60.6%
Toolathlon	54.6%	45.7%

Abstract reasoning

Benchmark	GPT-5.4	GPT-5.2
ARC-AGI-1 (Verified)	93.7%	86.2%
ARC-AGI-2 (Verified)	73.3%	52.9%
Frontier Science Research	33.0%	25.2%
FrontierMath Tier 1–3	47.6%	40.7%
Humanity’s Last Exam (with tools)	52.1%	45.5%

ARC-AGI-2 going from 52.9% to 73.3% is a 20-point jump — the largest improvement in abstract reasoning between any two consecutive GPT releases.

Factuality

GPT-5.4 is OpenAI’s most factual model:

• Individual claims are 33% less likely to be false vs GPT-5.2
• Full responses are 18% less likely to contain any errors

Pricing

Model	Input	Cached Input	Output
GPT-5.4	$2.50/M	$0.25/M	$15.00/M
GPT-5.4 Pro	$30.00/M	—	$180.00/M
GPT-5.2	$1.75/M	$0.175/M	$14.00/M

GPT-5.4 is priced ~43% higher on input than GPT-5.2, but OpenAI claims greater token efficiency offsets this — the model uses fewer tokens to solve equivalent problems.

Batch and Flex pricing available at half rate. Priority processing at 2x rate. Context >272K tokens priced at 2x input / 1.5x output.

API access

from openai import OpenAI
client = OpenAI()

response = client.responses.create(
    model="gpt-5.4",
    input="Analyze this codebase for security vulnerabilities.",
    reasoning={"effort": "medium"},
    text={"verbosity": "medium"}
)

Reasoning effort levels: none (default), low, medium, high, xhigh Verbosity levels: low, medium (default), high

Computer use requires the updated computer tool in the Responses API. Tool search uses tool_search capability with deferred loading.

Variants

Variant	Best for	Context
gpt-5.4	General-purpose: reasoning, coding, agentic tasks	1.05M
gpt-5.4-pro	Maximum performance on complex problems	1.05M
gpt-5-mini	Cost-optimized reasoning and chat	—
gpt-5-nano	High-throughput classification and instruction following	—

Steerability

New in GPT-5.4 Thinking (ChatGPT): the model shows an upfront preamble of its plan before executing, so users can adjust course mid-response. This reduces back-and-forth — you see where it is headed and can redirect before it finishes.

Safety

OpenAI rates GPT-5.4 as High cyber capability under their Preparedness Framework — the first general-purpose model with this classification. Deployed with:

• Expanded cyber safety stack with monitoring systems
• Trusted access controls
• Asynchronous blocking for higher-risk requests on Zero Data Retention surfaces
• Chain-of-Thought controllability testing (model cannot easily hide its reasoning)

Timeline

GPT-5.2 Thinking remains available in ChatGPT for 3 months under Legacy Models. Retirement date: June 5, 2026.

Availability

• ChatGPT: Rolling out now to Plus, Team, Pro users as GPT-5.4 Thinking
• Enterprise/Edu: Available via admin early access settings
• API: Available now as gpt-5.4 and gpt-5.4-pro
• Codex: Available with experimental 1M context support
• OpenRouter: Available as openai/gpt-5.4

Open-source alternatives

GPT-5.4 is proprietary with no public weights. For self-hosted alternatives:

• Qwen3.5-397B-A17B (Apache 2.0) — 403B MoE, 17B active, open-weight frontier reasoning
• DeepSeek V3.2 (MIT) — 685B MoE, strong reasoning + chat at 1/40th the price
• GLM 5 (MIT) — 754B params, Z.ai’s open-weight flagship
• LiquidAI LFM2-24B-A2B — novel architecture, efficient inference on consumer hardware