Reference Guide

OpenAI

ChatGPT, GPT and o-series models, Codex, DALL·E, Whisper, custom GPTs, and the API platform — what each one does and which to reach for.

← Back to Reference Hub

Best for: General-purpose chat, drafting, analysis, image generation, voice conversation, and as a frontend to OpenAI's full model lineup.

  • Switches between GPT-5.5 (default flagship, launched Apr 23, 2026), GPT-5.5 Pro, GPT-5.5 Thinking, and image tools from inside one conversation
  • Built-in web search, file uploads, code interpreter, image generation, advanced voice mode, and Deep Research
  • Memory across conversations and Projects (folders that hold instructions, files, and chats together)
  • Browser automation and multi-step research via ChatGPT Agent (formerly Operator) on Plus and Pro
  • Custom GPTs are usable on every paid tier; building requires Plus or higher

Limitations: Free tier is ad-supported in the US since Feb 2026 and capped on messages, image gen, and Deep Research. Video generation was discontinued April 26, 2026 with the Sora shutdown — no replacement product has launched. Memory and custom GPTs are off in some regions and on Free.

Consumer ChatFree Tier

Best for: Picking the right subscription tier — the gap between Free and Pro is large and most decision-makers get this wrong.

  • Free ($0/mo) — GPT-5.3 access, ad-supported in the US, capped messages and uploads, 5 lightweight Deep Research queries/month
  • Go ($8/mo) — Originally India-only, now global; ad-supported entry tier, higher caps than Free
  • Plus ($20/mo) — GPT-5.5 and GPT-5.5 Thinking (3,000 msgs/wk), DALL·E and GPT Image, advanced voice, Codex (cloud), 25 Deep Research queries/mo, build custom GPTs
  • Pro $100/mo (launched Apr 9, 2026) — 5x Plus caps, GPT-5.5 Pro access, 10x Codex usage through May 31, 2026; the sweet spot for Codex-heavy individual developers
  • Pro $200/mo — Near-unlimited usage, GPT-5.5 Pro at highest priority, 250 Deep Research queries/mo
  • Business ($20/seat/mo annual, or $25–$30/seat/mo monthly) — Renamed from the prior team tier on Aug 29, 2025; price-dropped Apr 2, 2026. Plus features plus shared workspace, team-built GPTs, admin console
  • Enterprise (custom) — No usage caps, 2x faster, SSO/SAML, SCIM, no training on your data, custom retention

Limitations: Four overlapping individual/team tiers (Plus, Pro $100, Pro $200, Business) make the picker confusing. Most heavy individual users land on Plus; Pro $100 is the new sweet spot for Codex-heavy work; Pro $200 is for Deep Research power users.

Pricing

Best for: Standard chat, drafting, RAG, function calling, multimodal input. Use these when you don't need extended reasoning.

  • GPT-5.5 (default flagship, launched Apr 23, 2026; API Apr 24): $5 input / $30 output per 1M tokens — 2x the rate of the predecessor flagship it replaced. 272K context window
  • GPT-5.5 Pro (top-tier variant, available on Pro/Business/Enterprise): $30 input / $180 output per 1M — for the hardest tasks
  • GPT-5: $0.625 input / $5.00 output — the budget flagship, near-parity on most tasks
  • GPT-4o: $2.50 input / $10 output — legacy multimodal model, still served for compatibility
  • All accept text, image, and audio input; GPT-5.5 has the strongest tool-use behavior and instruction following. 1M+ context supported on the API.

Limitations: GPT-4o still appears in older code paths and ChatGPT toggles — no longer the right default for new builds. GPT-5.5 doubled input pricing vs. the predecessor flagship, so cost-sensitive workloads should benchmark whether GPT-5 covers the use case at one-eighth the input cost before defaulting to the new flagship.

Flagship Models

Best for: Hard reasoning, math, code generation, scientific analysis, and any task where you'd let a smart human think for several minutes before answering. The reasoning workhorse since the o-series picker retirement.

  • GPT-5.5 Thinking is the extended-reasoning sibling of GPT-5.5, available in the ChatGPT model picker on Plus and above and through the API
  • Absorbs the use case the o-series picker entries previously served — math, code generation, scientific reasoning, complex multi-step tasks
  • Reasoning models bill internal "thinking" tokens as output — expect 5–10x more output tokens than the visible answer
  • o3-pro reportedly remains available API-only for the very hardest problems (verify against OpenAI's current model lineup before relying on it as of audit)
  • API pricing follows the GPT-5.5 family ($5 / $30 per 1M tokens) plus the reasoning-token output multiplier

Limitations: Higher latency than non-thinking GPT-5.5 (often 30s–5min). Reasoning tokens make cost prediction harder — budget for output volume to be much higher than what you see returned. The full o-series picker (o3, o4-mini, o1) and GPT-4.1 were retired from ChatGPT between Feb 13, 2026 (o4-mini) and April 2026; legacy code paths citing those models on the ChatGPT surface no longer apply.

Reasoning

Best for: Autonomous coding tasks — multi-file edits, test running, PR drafting, long-horizon refactors.

  • Cloud Codex: ChatGPT-hosted coding agent included with Plus, Pro, Business, Edu, and Enterprise — runs jobs in OpenAI's sandboxed cloud and pushes results to a branch
  • Codex CLI: Open-source Rust-based local terminal agent with read/write/run access to a chosen directory — OpenAI's answer to Claude Code
  • codex-mini-latest API: $1.50 input / $6.00 output per 1M, 75% prompt cache discount, served through the Responses API
  • Plus is enough for most individual developers; the Pro $100 tier's 10x Codex promo through May 31, 2026 is the best value while it lasts
  • Cloud and CLI share the same model and prompt patterns — switch based on whether you want sandboxed or local execution

Limitations: Cloud Codex draws from your plan's Codex limits, which are not always transparent in the UI. CLI is newer than competing tools and the orchestration patterns are still settling.

Coding Agent

Best for: Marketing assets, illustrations, product mockups, and in-chat image creation.

  • GPT Image 1.5 (latest, Apr 2026): token-based pricing, roughly $0.03 (low) to $0.19 (high quality) per image
  • GPT Image 1: $0.009–$0.20 per image depending on quality and resolution
  • GPT Image 1 Mini: $0.005–$0.052 per image — cheapest tier, fine for thumbnails and bulk jobs
  • DALL·E 3: $0.04 standard / $0.08 HD at 1024×1024; $0.08–$0.12 for portrait/landscape
  • Inside ChatGPT, all paid tiers can generate images; Free has a daily limit. Image editing (inpainting/style transfer) is available on the API and ChatGPT Plus+

Limitations: Style consistency across a series is weaker than dedicated image tools (Midjourney, Ideogram). Text rendering inside images has improved but still misses on complex copy. C2PA provenance is on by default and visible in metadata.

Image

Best for: Speech-to-text in 50+ languages — meeting transcripts, podcast captions, voice-driven applications, and as the input stage for downstream LLM processing.

  • Whisper (open source): model weights are freely available; ~4M monthly Hugging Face downloads make it the de facto open-source ASR baseline
  • Whisper API: $0.006/min ($0.36/hr) on the OpenAI platform
  • GPT-4o Transcribe: $0.006/min, supports diarization (speaker separation)
  • GPT-4o Mini Transcribe: $0.003/min ($0.18/hr) — cost-sensitive option, slightly lower accuracy
  • Self-hosting Whisper breaks even around 500+ hours/month vs API; below that, the API is cheaper than running your own GPU

Limitations: Real-time streaming requires the Realtime API, not Whisper-the-batch-endpoint. Accuracy on heavy accents and overlapping speakers still trails specialist providers (AssemblyAI, Deepgram) for some use cases.

AudioOpen Source

Best for: Packaging instructions, files, and tool access into a reusable assistant that lives inside ChatGPT. Best for internal team workflows and lightweight customer-facing helpers.

  • Build with natural-language config: instructions, knowledge files, conversation starters, and Actions (OpenAPI-spec tools)
  • Available to use on every paid tier; building requires Plus, Pro, Business, or Enterprise
  • The GPT Store publishes public GPTs with categories, trending, and staff picks
  • Revenue sharing for US builders is engagement-based (~$0.03/conversation average), with a 25 conversations/week minimum to qualify
  • Business and Enterprise can build private workspace GPTs only their org can use

Limitations: Most public GPT Store builders earn $0; the formula remains undisclosed and payouts are unpredictable. Workspace Agents (see below) are the successor for enterprise use cases.

Custom Agents

Best for: Building production AI applications — embedding OpenAI models into your own product, with separate billing from ChatGPT subscriptions.

  • Responses API: OpenAI's recommended default since Mar 2025; better caching (40–80% hit rates), supports Deep Research, MCP, computer use, and stateful Conversations
  • Chat Completions API: Still supported, not deprecated — use it for simple stateless calls and migrations from older code
  • Realtime API: WebSocket voice and text; audio at $32/1M input, $0.40/1M cached input, $64/1M output tokens
  • Batch API: Flat 50% discount on all models for async jobs that complete within 24 hours
  • Prompt caching: Automatic on Responses, Chat, Realtime, and Batch — up to 90% reduction on cached input tokens

Limitations: Pricing is per-token, not per-subscription — a runaway loop can rack up real money fast. The Responses/Chat/Assistants split has caused real migration churn; pin yourself to Responses for new work.

Developer Platform

Best for: Tasks that require navigating websites, filling forms, gathering and synthesizing information, and operating SaaS dashboards on your behalf.

  • ChatGPT Agent: Operator merged into ChatGPT — combines browser automation, Deep Research synthesis, and ChatGPT in one agent mode (Plus and Pro)
  • AgentKit (Oct 2025): developer toolkit with Agent Builder (multi-agent flows), Connector Registry (data/tools), and ChatKit (embeddable agent UI)
  • Computer Use: tool exposed through the Responses API for building custom browser/desktop automation agents
  • Workspace Agents (Apr 2026): enterprise successor to custom GPTs — native Slack, Google Drive, Salesforce integration; free through May 6, 2026 then credit-based
  • Frontier (Feb 2026): enterprise platform for managing AI "coworkers" across an org

Limitations: Agent mode in ChatGPT is rate-limited and queues during peak hours. Computer Use can fail on heavily JS-driven sites. Workspace Agents pricing post-May 6 is not yet announced — budget conservatively if rolling out enterprise-wide.

Autonomous Agents

Migration Notice: Assistants API → Responses API

The deadline

OpenAI deprecated the Assistants API on Aug 26, 2025, with a sunset date of Aug 26, 2026 — about four months from now. Existing Assistants integrations will stop working after that date.

Where to migrate

The replacement is the Responses API plus the Conversations API for stateful threads. OpenAI publishes an official migration guide; the move is mostly mechanical for standard tool-use patterns.

Why it's worth doing now

  • Better prompt caching (40–80% hit rates vs Assistants' weaker caching)
  • Native support for Deep Research, MCP servers, and computer use as built-in tools
  • Stateful conversation_id handling without juggling thread/run/message objects
  • 3% better SWE-bench scores from OpenAI's own benchmarks

Action items

  • Do not start new builds on the Assistants API — use Responses from day one
  • Inventory existing Assistants usage and plan migration before July 2026 to leave buffer
  • Azure OpenAI follows the same deprecation timeline — same deadlines apply there