benchmarks
GPT-4.1 is OpenAI’s text-and-image model for coding, instruction following and long-document processing, with a context window exceeding one million tokens. OpenAI retired it from ChatGPT on February 13, 2026; that announcement explicitly left API access unchanged.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| GPQA DiamondPhD-level science questions | 66.9%14 Apr 2025 | View |
| SWE-bench VerifiedResolving real GitHub issues | 48.5%8 Feb 2026 | View |
| OTIS Mock AIME 2024-25Competition mathematics | 38.3%14 Apr 2025 | View |
| FrontierMath | 5.5%14 Apr 2025 | View |
| SimpleQA VerifiedShort factual questions (hallucination resistance) | 31.1%31 Aug 2026 | View |
| MATH Level 5 | 83%14 Apr 2025 | View |
| LMArena TextHuman preference Elo from blind chat votes (not a percentage) | 1383 Elo22 Oct 2026 | View |
| Aider PolyglotCode editing across six programming languages | 52.4%14 Apr 2025 | View |
Weighted mean of 7 verified benchmark results: GPQA Diamond 66.9%, SWE-bench Verified 48.5%, OTIS Mock AIME 2024-25 38.3%, FrontierMath 5.5%, SimpleQA Verified 31.1%, MATH Level 5 83%, Aider Polyglot 52.4%.
Capabilities
- Processes text and image inputs and produces text outputs.
- Handles large document collections and codebases within a 1,047,576-token context window.
- Generates and edits code, including repository-level changes.
- Follows detailed instructions and formatting requirements.
- Supports function calling and structured outputs for application integration.
What improved
- Higher vendor-reported SWE-bench Verified performance than GPT-4o.
- Better adherence to detailed instructions than GPT-4o in OpenAI’s evaluations.
- Expands the context window from GPT-4o’s 128,000 tokens to more than one million tokens.
- Improves long-context comprehension and image understanding in vendor evaluations.
- Lower standard API input and output token prices than GPT-4o.
Known flaws & limitations
- Its documented knowledge cutoff is June 2024; later facts require external sources.
- Long-context performance can decline as information retrieval and reasoning demands increase.
- Literal instruction following can make ambiguous or conflicting requirements more consequential.
- It does not support native audio input or audio generation.
- Output is limited to 32,768 tokens despite the much larger input context window.
Risks, safety & security issues
- Untrusted documents or tool results can contain instructions that attempt to redirect application behavior.
- Generated answers can contain factual errors; structured output does not establish factual correctness.
- Incorrect code or tool calls can affect external systems when applications grant execution privileges without validation.
Good fit for
- Coding assistants that propose repository changes for testing and review.
- Analysis and synthesis of long documents with source-grounded verification.
- Structured extraction and classification from text and images.
- Tool-using workflows with constrained permissions and validated outputs.
Poor fit for
- Native real-time voice or audio-generation applications.
- Answering current-events questions without external information sources.
- Unsupervised high-stakes decisions requiring guaranteed factual accuracy.
- Unrestricted execution of generated code or consequential tool calls.
References
Profile last reviewed 10/27/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.