OpenAI gives GPT-4 a toolbelt with native function calling

OpenAI has given its most powerful models the ability to call your code. That’s the core of a sweeping update that introduces function calling to the Chat Completions API, letting gpt-4-0613 and gpt-3.5-turbo-0613 intelligently output JSON objects with arguments ready to plug into external tools and APIs. Instead of merely generating text, the models can now detect when a function needs to be triggered—say, converting a chat message like “What’s the weather in Boston?” into a structured call for get_current_weather—and respond with a matching signature. This isn’t just a parlor trick; it’s a direct pipeline from conversational intent to executable action, and it mirrors the logic behind ChatGPT Plugins without requiring developers to build their own plugin architecture.

Alongside the new capability, OpenAI is rolling out a 16k context version of gpt-3.5-turbo, quadrupling the standard 4k token window. That’s a meaningful shift for applications that need to process long documents or maintain extended conversational memory. The pricing story is equally aggressive: input tokens for gpt-3.5-turbo are now 25% cheaper, and the company’s embeddings model sees a 75% cost reduction. If you’re building search or retrieval-augmented generation systems, that last cut alone could reshape your infrastructure budget.

But model upgrades always come with a shadow. OpenAI acknowledged that while its evaluation metrics broadly improve, some specific tasks may regress with newer versions. The gpt-4-0613 model, for instance, shows significant gains in function calling but isn’t universally better at everything. That’s why the company continues to let developers pin specific model versions like gpt-4-0314—a practice that will be relevant given the newly extended deprecation timeline. After listening to community feedback, OpenAI pushed the sunset date for those older models from September 2023 to at least June 13, 2024. The message is clear: stability matters, and forced migrations are unpopular.

The company is also candid about safety research that remains unfinished. A proof-of-concept exploit demonstrated how untrusted data piped back from a tool’s output can manipulate the model into unintended behavior. OpenAI says it’s working on mitigations and advises developers to consume only trusted information in their function-calling implementations. It’s a rare moment of public caution from a company usually more bullish about its releases. For anyone stitching GPT into production systems, that note shouldn’t be skimmed.