OpenAI arms developers with cost-slashing batch API and 500x file boost

OpenAI is making a hard pivot toward the enterprise with a suite of features that read less like a gentle upgrade and more like a checklist ripped from a CIO’s wishlist. The message is clear: the company that once stumbled through enterprise sales is now building the guardrails and cost controls that big business demands. For the first time, customers can lock down communications between Azure and OpenAI using Private Link, minimizing exposure to the public internet. That’s alongside native multi-factor authentication—no bolt-on required—which joins an existing security stack that already ticks the boxes for SOC 2 Type II, HIPAA, and AES-256 encryption. It’s not flashy, but it’s the kind of blocking-and-tackling that actually closes deals with banks and hospitals.

Where things get genuinely interesting for developers is the new Projects feature. Rather than one flat namespace where every API key can run wild, teams can now scope roles and keys to specific initiatives, restrict which models are available, and set hard spending limits. Project owners can even create service account API keys that aren’t tied to a single human—a small change that eliminates a massive headache when automating CI/CD pipelines. It’s the sort of organizational scaffolding that sounds dull until you’ve blown through a budget because someone left a GPT-4 key in a test script.

The Assistants API is getting the kind of overhaul that suggests OpenAI is listening to its loudest complainers. The headline number is a 500x jump in file capacity: the new ‘file_search’ tool can ingest up to 10,000 files per assistant, up from the paltry 20-file limit that made the previous version feel like a toy. The tool also supports parallel queries through multi-threaded searches, plus improved reranking and query rewriting. A new ‘vector_store’ object lets files be added once and reused across assistants and threads—goodbye, redundant embeddings. Streaming support, one of the most-requested features, arrives for real-time conversational responses, and developers finally get a ‘tool_choice’ parameter to force specific tools during a run. Fine-tuned GPT-3.5 Turbo models are now supported in the API as well.

The cost story might be the most pragmatic part of this release. For sustained workloads on GPT-4 or GPT-4 Turbo, committed throughput can knock 10 to 50 percent off the bill depending on the size of the commitment. A new Batch API lets developers run non-urgent jobs asynchronously at half the usual price, with results returned within 24 hours—a fit for model evaluation, offline classification, or synthetic data generation. Taken together, these moves don’t just make OpenAI more palatable for enterprises; they make it harder for competitors to argue that the platform is too expensive or too loose to trust with serious work.