GPT-5.4, released March 5, 2026, is OpenAI's current flagship model — unifying its general-purpose and coding model lines into a single frontier offering with native computer use, configurable reasoning effort, and a 1M-token context window at competitive pricing.
Architecture Deep Dive → GPT-5.4 Architecture Breakdown — unified model pipeline, parameterized reasoning effort, context compaction design, native computer use implementation, and comparison with Claude Opus 4.6.
Why It's in Adopt
GPT-5.4 represents a decisive step forward for teams in the OpenAI ecosystem. Key improvements over the GPT-5.2 line:
- Unified Codex + GPT: GPT-5.4 incorporates the specialized coding capabilities of GPT-5.3-Codex into the general model — one API, one billing line, top-tier coding.
- Native computer use: The first general-purpose OpenAI model with native GUI automation — agents can operate browsers, desktop apps, and complex multi-application workflows.
- Configurable reasoning: Set effort to
none / low / medium / high / xhigh— pay only for the thinking the task needs. - 33% fewer factual errors vs. GPT-5.2.
- Tool search: GPT-5.4 finds and uses the right tool from large tool ecosystems more reliably than its predecessors.
Performance
| Benchmark | Score |
|---|---|
| SWE-bench Verified | 74.9% (resolving real GitHub issues) |
| GPQA Diamond | 92.8% (expert-level science) |
| AIME 2026 | Leading scores |
On SWE-bench, GPT-5.4 is narrowly behind Grok 4.2 (75%) but ahead of Claude Opus 4.6 on this benchmark.
Context Window & Output
- 1M token context (1.05M: 922K input, 128K output)
- Prompts over 272K tokens incur a 2× surcharge (plan accordingly for large codebase ingestion)
- GPT-5.4 mini and nano (released March 17, 2026) bring the same architecture to fast, cheap sub-tasks
Pricing
| Tier | Input | Output |
|---|---|---|
| Standard API | $2.50/M | $15/M |
| Pro (enhanced reasoning) | $30/M | $180/M |
| Batch / Flex | ~$1.25/M | ~$7.50/M |
GPT-5.4-Cyber (April 14, 2026)
OpenAI launched GPT-5.4-Cyber, a fine-tuned variant optimized for defensive cybersecurity. Key differences from the base model:
- Lower refusal boundary for legitimate security work — binary reverse engineering, vulnerability analysis, and malware triage that the base model would refuse
- Access-controlled via the Trusted Access for Cyber (TAC) program — vetted security researchers and teams only, not generally available
- Evaluated by CAISI and UK AISI for cyber capability and safeguard adequacy before broader release
GPT-5.4-Cyber is not a separate model for radar purposes — it uses the same architecture and pricing as GPT-5.4, with fine-tuning that relaxes safety boundaries specifically for authenticated defensive use cases.
How to Access
- ChatGPT — GPT-5.4 Thinking (Plus, Team, Pro, Enterprise)
- OpenAI API —
gpt-5.4andgpt-5.4-promodel IDs - GPT-5.4-Cyber — TAC program only; verify identity at chatgpt.com/cyber or through an OpenAI representative
- Codex — the dedicated agentic coding platform uses GPT-5.4 as its base
- OpenAI Agents SDK — recommended base model for production agents
Key Characteristics
| Property | Value |
|---|---|
| License | Proprietary |
| Context window | 1,050,000 tokens |
| Pricing | $2.50/M input, $15/M output |
| Provider | OpenAI |
| API model ID | gpt-5.4 |
| Release date | March 5, 2026 |
Further Reading
- Introducing GPT-5.4 (OpenAI)
- GPT-5.4 mini and nano (OpenAI)
- OpenAI Pricing
- GPT-5.4 Architecture Breakdown — full deep dive into the unified Codex + GPT architecture, reasoning levels, and computer use