# GPT-5.4

GPT-5.4 is an AI model released by OpenAI on Mar 5 2026. At release it scored 48% on BullshitBench v2, 92.8% on GPQA Diamond and 52.1% on Humanity's Last Exam (with tools).

## Facts

| Field | Value |
| --- | --- |
| Model | GPT-5.4 |
| Developer | OpenAI |
| Release date | Thursday, Mar 5 2026 |
| Licensing | Proprietary |

## API pricing

All rates in USD per 1,000,000 tokens, pay-as-you-go.

| Tier | Input | Cached input | Output |
| --- | --- | --- | --- |
| Up to 272K input tokens | $2.50 | $0.25 | $15.00 |

Verified August 18, 2026 against the first-party source: https://developers.openai.com/api/docs/models/gpt-5.4

## Benchmark scores published at release

| Benchmark | Score | Source | What it measures |
| --- | --- | --- | --- |
| BullshitBench v2 | 48% | [BullshitBench](https://github.com/petergpt/bullshit-benchmark) | Given a confidently-worded but nonsensical prompt, does the AI spot that it makes no sense and push back — instead of playing along and inventing an answer? The score is how often it clearly called out the nonsense. Higher is better. |
| ProgramBench | 0% | [ProgramBench](https://programbench.com/), retrieved 2026-09-11 | The AI receives a working program and its documentation, then builds a replacement from scratch without the original source code, internet access or decompilation. The score is the percentage of 200 programs that pass every behavioral test. We record each model's best published mini-SWE-agent result, including higher reasoning efforts where available. Partial test-pass rates and almost-solved programs do not count toward this score. Equal scores share a rank here; the official board also uses partial progress to break ties. Higher is better. |
| Next.js Evals | 65% | [Next.js Evals](https://nextjs.org/evals) | Vercel's open eval of how well AI coding agents build and migrate real Next.js apps — measured as the share of tasks the agent completes successfully. Higher is better. |
| Terminal-Bench 2.0 | 75.1% | Lab | Can the AI work in a command-line terminal — running commands and finishing technical setup tasks the way a developer would? (Version 2.0 of the test.) Higher is better. |
| Expert-SWE (Internal) | 68.5% | Lab | OpenAI's private set of expert-level software-engineering problems. Higher is better. |
| Toolathlon | 54.6% | Lab | Tests how well the AI uses everyday real-world tools and apps to get things done. Higher is better. |
| BrowseComp | 82.7% | Lab | Can the AI browse the web and track down hard-to-find answers? Higher is better. |
| CyberGym | 79% | Lab | Tests the AI on cybersecurity challenges — finding and exploiting software weaknesses inside a safe sandbox. Higher is better. |
| Humanity's Last Exam (with tools) | 52.1% | [BenchLM](https://benchlm.ai), retrieved 2026-08-31 | Humanity's Last Exam — extremely hard expert questions across many subjects. “With tools” means the AI is allowed to search the web or run code while answering. Higher is better. |
| ARC-AGI-2 | 73.95% | [BenchLM](https://benchlm.ai), retrieved 2026-08-24 | Puzzle-style tests of abstract reasoning and pattern-finding — the kind of thing people find easy but AIs often struggle with. Higher is better. |
| FrontierMath (Tier 1–3) | 47.6% | Lab | Very hard, research-level math problems. Tiers 1–3 are the (still extremely difficult) lower tiers. Higher is better. |
| FrontierMath (Tier 4) | 27.1% | Lab | Very hard, research-level math problems. Tier 4 is the hardest — close to what professional research mathematicians tackle. Higher is better. |
| GPQA Diamond | 92.8% | Lab | Graduate-level science questions in biology, physics, and chemistry — hard enough that subject-matter PhDs score around 65%. Higher is better. |
| OSWorld-Verified | 75% | Lab | Can the AI actually operate a computer — clicking, typing, and using real apps — to finish tasks on its own? Higher is better. |
| GDPval (win/tie rate) | 83% | Lab | How often the AI's work matches or beats a human expert's on real knowledge-work tasks. Higher is better. |

## Questions and answers

### When was GPT-5.4 released?

GPT-5.4 was released by OpenAI on Thursday, Mar 5 2026.

### Who made GPT-5.4?

GPT-5.4 was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research.

### How much does GPT-5.4 cost?

GPT-5.4 costs $2.50 per million input tokens and $15.00 per million output tokens through the OpenAI API. Cached input is $0.25 per million tokens. Rates are pay-as-you-go API prices verified against OpenAI's published pricing on August 18, 2026.

### What benchmark scores did GPT-5.4 get?

GPT-5.4 reports 15 tracked benchmark scores — BullshitBench v2: 48%; ProgramBench: 0%; Next.js Evals: 65%; Terminal-Bench 2.0: 75.1%; Expert-SWE (Internal): 68.5%; Toolathlon: 54.6%; BrowseComp: 82.7%; CyberGym: 79%; Humanity's Last Exam (with tools): 52.1%; ARC-AGI-2: 73.95%; FrontierMath (Tier 1–3): 47.6%; FrontierMath (Tier 4): 27.1%; GPQA Diamond: 92.8%; OSWorld-Verified: 75%; GDPval (win/tie rate): 83%. Scores are the figures published at release by OpenAI.

### Is GPT-5.4 open source?

No. GPT-5.4 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms.

### What came before and after GPT-5.4?

OpenAI's previous tracked release was GPT-5.3-Codex-Spark on Feb 12 2026, 21 days earlier. It was followed by GPT-5.4-Pro on Mar 5 2026.


---

Canonical page: https://aireleasetracker.com/model/openai/gpt-5.4
Full dataset: https://aireleasetracker.com/llms-full.txt · JSON: https://aireleasetracker.com/models.json
Source: AI Release Tracker (https://aireleasetracker.com). Most benchmark scores come from lab launch material; gathered results identify the leaderboard that published them.