GPT-5.4 mini

GPT-5.4 mini is a faster, cheaper, and more efficient GPT-5.4 variant released by OpenAI on 17 March 2026. It brings the core capabilities of full GPT-5.4 to a model optimized for high-throughput workloads — production-scale chat, coding assistants, and agent workflows serving large user populations. Available in ChatGPT (Plus/Pro/Team/Enterprise), via the OpenAI API (`gpt-5.4-mini`), Azure OpenAI, and aggregators such as OpenRouter. Per OpenAI (llm-stats.com), it is the company's strongest mini model for coding, computer use, and sub-agent building — a significant quality jump over GPT-4o mini and GPT-5-mini.

Core specification: 400,000-token context window (max output 128,000), knowledge cutoff August 2025. Input accepts text and images (useful for OCR, diagrams, screenshots in agent workflows). Output: text, code, structured data, documents. The model does not expose a separate extended thinking mode (reasoning_effort=max) like GPT-5.4 Thinking / GPT-5.6 Sol — the design balances latency + cost against quality for mass workloads. Solid multi-step reasoning, reliable instruction following, and consistent performance across diverse tasks with improved cost efficiency vs earlier GPT-5-mini.

API pricing (as of 17 July 2026): $0.75 per 1M input tokens, $4.50 per 1M output tokens. Cache-hit pricing further reduces cost (30-day rolling average across most providers is 60-80% cheaper than list price after prompt caching — OpenRouter data). For comparison: full GPT-5.4 runs ~$3.00/$15.00 (5-6x more expensive), GPT-5.6 Sol with reasoning ~$5/$25. This positions GPT-5.4 mini as the best price-to-performance ratio in OpenAI's lineup for non-reasoning workloads — the default choice for agent orchestration, chat backends, and pipeline ETL/classification at scale.

Use cases and case studies. Codex CLI (OpenAI) uses GPT-5.4 mini as its backbone for lightweight agent command-line flows — confirmed e.g. in OpenAI's GPT-Red blog (case study 2: GPT-Red attack on a Codex CLI agent based on GPT-5.4 mini across 10 held-out data-exfiltration scenarios). Other typical uses: high-volume classification, data extraction from documents, agent routing/dispatch, multi-turn consumer-scale chat, RAG backend, background tasks in product pipelines. Modality support: text, image (understanding), documents (parsing), function calling, structured output, streaming, tool use, computer use.

Enterprise-grade security (OpenAI Trust Center): data-not-used-for-training-by-default (for API and enterprise offerings), encryption at-rest, encryption in-transit, SOC 2, ISO 27001, SSO for organizational environments, Enterprise Business Associate Agreements for HIPAA (via ChatGPT Enterprise / API). Data residency via Azure OpenAI (US, EU, selected regions). SLA and standard support in ChatGPT Business/Enterprise plans and via the OpenAI API. Preferred single-tenant unit via OpenRouter/CloudPrice/etc. as aggregator for multi-provider failover. Direct competition: Claude Sonnet 4.5, Gemini 3.5 Flash, DeepSeek V4, Grok Fast 3 — all in the same fast, paid mini-model category for production workloads.