OpenAI GPT-5.4
by OpenAI
OpenAI's million-token frontier model with native computer use and a five-level reasoning dial
GPT-5.4 is OpenAI's frontier model, released March 2026, unifying the Codex and GPT lines into one system. It pairs a 1,050,000-token context window with native computer use, a five-level reasoning-effort dial and text-plus-image input, aimed at teams running long-context document work, agentic coding and desktop automation through one API.
GPT-5.4 is OpenAI's frontier model, released on 5 March 2026 under the single snapshot gpt-5.4-2026-03-05, and it unifies what were previously the separate Codex and GPT model lines into one system covering coding, document understanding, tool use and instruction following. It takes text and image input and returns text, with a 1,050,000-token context window and up to 128,000 output tokens, and a knowledge cutoff of 31 August 2025. Its defining control is a five-level reasoning-effort dial — none, low, medium, high and xhigh — letting one model serve both latency-sensitive traffic and long-horizon agentic work without switching endpoints. The API surface covers streaming, structured outputs, function calling, file search, file uploads, image input, web search and prompt caching, and the hosted tool set adds a code interpreter and computer use. On the OSWorld-Verified desktop-agent benchmark it scored 75 percent, against 47.3 percent for GPT-5.2 and an average human score of 72.4 percent, and OpenAI reported a 33 percent reduction in factual errors relative to GPT-5.2. List pricing is $2.50 per million input tokens, $0.25 per million cached input tokens and $15.00 per million output tokens, but any request whose input exceeds 272,000 tokens is billed at double input and 1.5 times output for the whole session and falls into a separate, lower rate-limit tier — so the headline million-token window carries a real premium. Rate limits scale from 500 requests per minute at Tier 1 to 15,000 at Tier 5. Independent benchmarking by Artificial Analysis places GPT-5.4 at xhigh reasoning effort at an intelligence index of 51, ranking 19th of 185 models, with output at 141.2 tokens per second. OpenAI has since released GPT-5.5 as its successor, though GPT-5.4 does not appear on OpenAI's published deprecation schedule.
Engineering leaders consolidating separate coding and general-purpose model endpoints onto one API, especially where agents need to drive a desktop or browser rather than only return text.
One model and one reasoning dial covers cheap high-volume calls and long-horizon agentic work, removing the routing layer teams built to switch between Codex-class and GPT-class models.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Usage-based
- Target Market
- CTOs, CIOs, Enterprise Developers, Data Scientists, Platform Engineers
- Deployment
- API-based, Cloud-only
- Founded
- 2015
- Headquarters
- San Francisco, United States
- Team Size
- 500+
Key Features
- ✓1,050,000-token context window
Holds entire codebases, contract sets or filing archives in one request, with up to 128,000 output tokens per completion.
- ✓Five-level reasoning effort control
none, low, medium, high and xhigh let one endpoint serve latency-sensitive traffic and deep agentic reasoning without switching models.
- ✓Native computer use
Built-in desktop control scoring 75 percent on OSWorld-Verified, against 47.3 percent for GPT-5.2 and 72.4 percent average human.
- ✓Unified Codex and GPT lines
One model covers coding and general reasoning, removing the routing layer teams maintained between separate specialised endpoints.
- ✓Prompt caching at ten percent of input price
Cached input costs $0.25 per million against $2.50 standard, materially cutting cost on repeated system prompts and long shared context.
- ✓Hosted tool suite
Web search, file search, code interpreter, function calling and structured outputs are available server-side without building separate tool infrastructure.
- ✓Multimodal input
Accepts text and images in the same request, so screenshots, scanned documents and diagrams are analysed alongside prose.
Capabilities
Use Cases
- •Whole-repository code review
Load an entire codebase into context so the model reasons across module boundaries instead of one diff at a time.
- •Desktop-agent automation
Drive legacy applications with no API through native computer use, at roughly average human performance on OSWorld-Verified tasks.
- •Contract and filing analysis at scale
Process a full deal room or regulatory filing set in a single call rather than chunking and reassembling results.
- •Tiered inference on one endpoint
Serve cheap classification at effort none and escalate the same request path to xhigh for genuinely hard cases.
- •Multimodal support triage
Combine a customer's screenshot, error log and account history in one request to diagnose an issue without human handoff.
Ideal For
Best For
- ✓Agentic coding and software-engineering workloads that previously needed a separate Codex-line endpoint alongside a general model
- ✓Computer-use and desktop-automation agents, where its 75 percent OSWorld-Verified score is at roughly average human level
- ✓Long-context document analysis across whole contract sets, codebases or filing archives that exceed a 200K window
- ✓Mixed-traffic products that need one endpoint to serve both sub-second responses and deep reasoning by changing an effort parameter
- ✓Multimodal workflows combining document images, screenshots and text in a single request
Not Ideal For
- ✗Cost-sensitive high-volume classification or extraction — at $2.50/$15.00 per million tokens it is ten times GPT-5 mini's input price for work that does not need frontier reasoning
- ✗Teams that actually intend to fill the million-token window: crossing 272,000 input tokens doubles input pricing, raises output pricing 1.5x for the whole session, and drops the request into a lower rate-limit tier
- ✗Latency-critical interactive applications at high reasoning effort — Artificial Analysis measured time-to-first-token at 118.4 seconds against a 2.77-second median for reasoning models in the same price band
- ✗Regulated security or legal workflows sensitive to over-refusal, where reviewers documented the model declining to analyse already-public court records on privacy grounds
- ✗Organisations that cannot absorb frequent model re-qualification — GPT-5.3 Instant shipped two days before GPT-5.4, and GPT-5.5 has already superseded it
Integrations
Deployment
Market Analysis
Pros
- ✓Artificial Analysis independently measures an intelligence index of 51, ranking 19th of 185 models, well above the median of 32
- ✓OSWorld-Verified score of 75 percent lifts desktop automation from demo-grade to roughly average human performance
- ✓Consolidating Codex and GPT into one model removes an entire routing layer from agent architectures
- ✓Prompt caching at $0.25 per million makes long shared system context ten times cheaper than uncached input
- ✓OpenAI publishes SOC 2 Type 2, ISO/IEC 27001, 27017, 27018, 27701 and 42001 certifications on its trust portal
Cons
- ✗The 1M context window is effectively a paid upgrade — past 272K input tokens the whole session is billed at 2x input and 1.5x output and drops to a lower rate-limit tier
- ✗Artificial Analysis measured time-to-first-token at 118.4 seconds at xhigh reasoning, against a 2.77-second median for reasoning models in the same price band
- ✗Documented over-refusal on legitimate professional queries, including declining to analyse court records the court had already published, until requests were reframed as teaching exercises
- ✗Release cadence creates real churn: GPT-5.3 Instant shipped 3 March and GPT-5.4 on 5 March, and GPT-5.5 has already superseded it
- ✗The Decoder criticised steep pricing increases across the 5.4 line, with the mini and nano variants reported at four times their GPT-5 equivalents
- ✗ZDNET found the Thinking variant hallucinated rarely but sometimes failed to follow prompts accurately
Pricing
Standard API (up to 272K input tokens)
From $2.50 per 1M input tokens
- ✓$2.50 per 1M input tokens
- ✓$0.25 per 1M cached input tokens
- ✓$15.00 per 1M output tokens
- ✓500 RPM at Tier 1 rising to 15,000 RPM at Tier 5
Long context (over 272K input tokens)
2x input, 1.5x output for the full session
- ✓Access to the full 1,050,000-token window
- ✓2x input and 1.5x output pricing applied to the whole session
- ✓Separate rate-limit tier of 200 to 8,000 RPM
- ✓Applies to standard, batch and flex processing
Metered purely per token with no seat licences: $2.50 per million input, $0.25 cached and $15.00 output. The critical detail buyers miss is the 272,000-token threshold — cross it and the entire session is rebilled at double input and 1.5 times output, and moves to a lower rate-limit tier, so the advertised million-token window is a premium product rather than an included one. Prompt caching is the main lever, cutting repeated context to a tenth of standard input price.
Security & Compliance
Connect
Sources
This page was written from 6 sources, 5 on domains other than developers.openai.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
TwelveLabs
Video intelligence API that makes every hour of enterprise footage searchable, analyzable and agent-ready.
Engram
A learned memory layer that makes AI actually know your organization — at up to 100x fewer tokens.
Mistral OCR 4
Structure-aware document AI that returns bounding boxes, typed blocks, and per-word confidence scores.
Hugging Face
The AI community building the future.
Mentioned In
AWS vs GCP vs Azure ML: The Real Costs Nobody Tells You
Enterprise AI analysis: AWS vs GCP vs Azure. Strategic insights, ROI considerations, and implementation guidance for technical and business leaders evaluatin...
March 15, 2026Vendor SelectionGPT-5.4 Mini and Nano Launch: How OpenAI Just Undercut Anthropic 5x (And Why Google Still Wins on Price)
GPT-5.4 Mini and Nano Launch. For enterprise decision-makers: strategic analysis, cost implications, and implementation guidance for AI investments and vendo...
March 22, 2026