Introducing Codex
OpenAI launched Codex, a dedicated LLM serving framework optimized for code understanding, generation, and execution—featuring real-time sandboxed execution, syntax-aware inference, and low-latency API access; it is neither a general-purpose LLM nor a variant of GPT-4.
Positioning and Architecture
Codex is an independently released, code-specialized serving framework from OpenAI—not a variant or fine-tuned version of the GPT series, nor does it release model weights. Its core integrates a closed-source base model (internal codename code-davinci-002 and successors) pre-trained on massive code corpora, coupled with a custom tokenizer, syntax-aware decoder, and sandboxed execution engine.
Key Capabilities
- Full context-aware code completion across 12 mainstream languages including Python, JavaScript, TypeScript, Go, and Rust; average token generation latency <120ms (P95, AWS us-east-1)
- Built-in configurable sandbox environment (based on gVisor isolation) enabling safe execution of generated code snippets, returning stdout/stderr and runtime resource metrics (CPU time ≤ 3s, memory ≤ 512MB)
- Two distinct API endpoints:
codex-complete(generation-only) andcodex-execute(joint generation + execution) - Default inference parameters (
temperature=0.2,top_p=0.95) ensure high output determinism; achieves 72.4% pass@1 on HumanEval (vs. 67.0% for GPT-4-1106-preview)
Application Boundaries
Codex explicitly excludes natural language dialogue, multimodal input, or long-document summarization. Its tokenizer vocabulary covers only code symbols and common programming terms (~50K tokens), excluding high-frequency general-domain words. SLA guarantees 99.95% uptime—but only for enterprise-tier API subscriptions (minimum $1000/month)