GPT-4 Is Released

OpenAI released GPT-4 on March 14, 2023, making it available to ChatGPT Plus subscribers ($20/month) and via the API to developers on a waitlist. Microsoft had already integrated GPT-4 into Bing Chat (launched February 7, 2023) before the formal public release — an early signal that enterprise integration was ahead of consumer access. OpenAI published a GPT-4 Technical Report the following day, disclosing extensive benchmark results but deliberately omitting parameter count, training compute, dataset composition, and architecture specifics, citing “competitive and safety reasons.” This represented a notable break from the research transparency of GPT-2 (2019) and GPT-3 (2020), both of which had technical papers with full methodology. On standard academic benchmarks, GPT-4 scored in the 90th percentile on the bar exam (GPT-3.5 had scored the 10th percentile), in the 93rd percentile on the SAT reading/writing section, and 88th percentile on SAT math. On MMLU (57 academic subjects, 5-shot), GPT-4 scored 86.4% versus GPT-3.5’s 70%. The model launched with 8,192 token context (a 32K context version was available to select API partners), compared to GPT-3.5’s 4,096 tokens.

GPT-4 was announced as multimodal — able to accept images as input — but image inputs were not available at launch for general users. The visual capability debuted publicly through a collaboration with Be My Eyes (a visual assistance app for blind users) as a limited preview, and was not broadly available in ChatGPT until GPT-4V (GPT-4 with Vision) launched in September 2023. The launch API supported text-only inputs with GPT-4 and GPT-4-32k model IDs. Anthropic launched Claude 1 the same day (March 14, 2023), positioning as the first major frontier model release directly competing with OpenAI on timing. On March 23, 2023, OpenAI announced ChatGPT Plugins — a system for connecting GPT-4 to third-party APIs (web browsing, code interpreters, database tools) that anticipated the later “tool use” and “function calling” paradigm standardized in the June 2023 API update.

The release accelerated a new software architecture for applications. Developers using the GPT-4 API could delegate reasoning, extraction, summarization, code generation, and structured data extraction to the model rather than writing deterministic code for each task. LangChain (first released October 2022) grew rapidly as a framework for chaining multiple GPT-4 API calls with retrieval from vector databases (Pinecone, Chroma, Weaviate) into “retrieval-augmented generation” (RAG) pipelines. The architecture introduced new engineering dependencies: latency (200–1,000ms per API call vs. microseconds for local computation), nondeterminism (the same input could produce different outputs at temperature > 0), rate limits, token-based pricing ($0.03/1K input tokens and $0.06/1K output tokens for GPT-4 8K at launch), and behavioral drift across model updates that could silently change application outputs. These tradeoffs drove new practices around prompt versioning, output validation, and regression testing for AI-integrated applications.