GPT-4.1 is here, and it was built for developers

OpenAI just dropped a surprise new set of models: GPT-4.1, 4.1 Mini, and 4.1 Nano. Interestingly, these aren't rolling out on the main ChatGPT website but are specifically launching via the API, signaling a strong focus on developers and integrating AI into other applications.

This launch seems like a direct response to competitors like Google's Gemini and Anthropic's Claude, particularly targeting areas where OpenAI might have been lagging, like coding performance and tool usage.

Key Topics & Details:

• The Models:
  • Three new models introduced: GPT-4.1, 4.1 Mini, and 4.1 Nano.
  • These are API-only releases, meaning you can't just go use them on the standard ChatGPT site. You need to access them through code or tools built on the API (like T3 Chat, Cursor, Windsurf etc.).
  • The presenter feels this API-only strategy is intriguing and highlights a shift towards empowering developers directly.
  • GPT-4.1 was apparently tested in stealth via OpenRouter before the official announcement (as 'Alpha' and 'Optimus Alpha').
• Performance & Benchmarks:
  • GPT-4.1 shows significant improvement on coding benchmarks like SWE Bench, closing the gap with competitors like Claude 3.5.
  • It also performed well on the Scale multi-challenge (checking various capabilities) and video context understanding benchmarks.
  • Presenter is waiting for independent benchmark results (like from Artificial Analysis) for a clearer picture but is initially impressed.
  • Compared to GPT-4.0, 4.1 is positioned as roughly equivalent in general capability but much better specifically at coding and instruction following.
• Pricing:
  • GPT-4.1 is cheaper than GPT-4.0 ($2/million input tokens, $8/million output tokens vs $3/$15 for 4.0 Turbo).
  • GPT-4.1 Mini is more expensive than the older 4.0 Mini ($0.40/$1.60 vs $0.15/$0.60 per million tokens).
  • GPT-4.1 Nano is the cheapest model OpenAI has released, priced identically to Google's Gemini 2.0 Flash ($0.10/$0.40 per million tokens).
  • There's a new, increased prompt caching discount (75%) for repeated context, making large context tasks cheaper if the initial prompt data is reused.
• GPT-4.1 Nano Concerns:
  • The presenter is quite confused and skeptical about the value proposition of GPT-4.1 Nano.
  • While priced like Gemini 2.0 Flash (a favorite model for its price/performance), OpenAI's own charts suggest Nano is dumber than the older GPT-4.0 Mini, which itself is considered significantly less capable than Flash.
  • Latency and throughput comparisons via OpenRouter also don't show a clear advantage for Nano over Flash.
  • Presenter feels Nano might not have a strong use case currently, potentially only useful if extremely low latency is the absolute priority over capability, but even there it doesn't seem to dominate.
• Context Window:
  • Massive increase to a 1 million token context window for GPT-4.1 (up from 128k). This matches Google's Gemini.
  • This allows feeding huge amounts of information (like entire codebases - estimated 8 copies of the React codebase) into the model at once.
  • This fundamentally changes what's possible, reducing the need for complex RAG (Retrieval-Augmented Generation) setups in some cases.
  • OpenAI claims strong performance on 'needle in the haystack' tests, meaning the model can still find specific information even within that massive context. They even open-sourced a new benchmark (MRCR) to test this more rigorously.
  • Presenter notes this large context is part of why it's API-only, as pasting millions of tokens into a chat UI isn't practical.
• Coding Improvements:
  • This is a major focus. GPT-4.1 is touted as a huge leap in coding ability, even potentially surpassing the specialized Claude Opus (O3) Mini in some benchmark aspects.
  • It's specifically trained to be better at following code 'diff' formats (showing only changed lines), which saves costs and latency compared to regenerating entire files.
  • Real-world tests from partners like Windsurf show significant improvements in code generation acceptance rates and efficiency compared to GPT-4.0.
  • Presenter notes that while O3 Mini might be good for deep reasoning on hard problems, 4.1 seems much better for day-to-day coding tasks and UI generation.
• Tool Calling & Instruction Following:
  • This is seen as another critical improvement, directly challenging Claude's dominance in this area.
  • Tool calling (letting the AI use external functions, like checking weather or analyzing files in an IDE) is essential for complex applications.
  • GPT-4.1 is significantly better at following instructions, including:
    • Format Following: Adhering to specified output formats (XML, YAML, etc.).
    • Negative Instructions: Understanding instructions like "don't do X".
    • Order Instructions: Following sequential steps correctly.
    • Context Requirements: Ensuring specific information is included (e.g., "always include protein count").
    • Ranking: Ordering lists correctly.
    • Overconfidence: Better at saying "I don't know" when appropriate.
  • Presenter emphasizes how crucial tool calling is and how Claude's strength here justified its premium price; 4.1 is OpenAI fighting back hard on this front.
  • Interestingly, the presenter notes that 'reasoning' models (like O3 Mini) sometimes struggle with tool calls, potentially overusing them, making non-reasoning models like 4.1 potentially better for tool-heavy workflows.
• Other Points:
  • GPT-4.5 Preview is being deprecated (users have until July 2025 to migrate), likely to free up resources for the new models. Presenter doesn't think many used it due to high cost.
  • The presenter views this launch as OpenAI acknowledging developer needs and responding strongly to competition, even releasing a major model API-first.