The short answer (September 2026): ChatGPT runs on the GPT-5.6 model family -- Sol for Plus and Pro users, Luna for Free and Go. Standard text chat has been unlimited on every plan since August 6, 2026. The cheapest plan worth paying for if you use ChatGPT professionally is Plus at $20/month, which unlocks Deep Research, Agent Mode, the Sol model, and an ad-free experience. The fastest way to improve output quality right now: fill in custom instructions so you stop re-establishing context every session, and move recurring work into Projects so memory persists. Use the Think slider for analytical questions and skip it for simple lookups.
Last verified: September 1, 2026, against official docs and pricing pages.
Which ChatGPT Plan Fits Your Workload
OpenAI currently offers five individual plan tiers and a multi-seat Business option (formerly called Team, renamed August 2025).
| Plan | Price | Standard chat model | Deep Research | Agent Mode | Ads |
|---|---|---|---|---|---|
| Free | $0 | GPT-5.6 Luna | 5 tasks/month | No | Yes (rolling out) |
| Go | $8/month | GPT-5.6 Luna | Not included | No | Yes |
| Plus | $20/month | GPT-5.6 Sol | 25 tasks/month | Yes | No |
| Pro $100 | $100/month | GPT-5.6 Sol | 250 tasks/month | Yes | No |
| Pro $200 | $200/month | GPT-5.6 Sol Pro | 250 tasks/month | Yes | No |
| Business (Standard seat) | $25/user/month (monthly) or $20/user/month (annual), min 2 seats | GPT-5.6 Sol | 25 tasks/user/month | Yes | No |
Go versus Plus is the live decision for most individuals. Go gives you higher Free-tier usage and unlimited text chat, but OpenAI deliberately left out Deep Research, Agent Mode, and the Sol model. It also still shows ads. For $12 more per month, Plus adds all of those features, upgrades you to Sol, and removes ads. If your use is genuinely occasional and low-stakes, Go is defensible. For professional work, it is not.
The two Pro tiers differ only in volume. Pro $100 gives 5x Plus usage limits; Pro $200 gives 20x plus GPT-5.6 Sol Pro reasoning and higher Codex task allowances. Most practitioners who are not running dozens of Deep Research reports or multi-hour agentic sessions weekly do not need Pro $200.
Business requires a minimum of two seats. Standard seats are $25/user/month (monthly) or $20/user/month (annual). Premium seats, which include more Codex usage, are $125/user/month (monthly) or $100/user/month (annual). Enterprise pricing is negotiated. Current seat options are at the OpenAI business pricing page.
The Models Inside ChatGPT Today
The GPT-5.6 family launched July 9, 2026 and replaced prior generations as the ChatGPT defaults. Three variants; one family name.
GPT-5.6 Sol is the flagship. OpenAI describes it as state-of-the-art across coding, knowledge work, cybersecurity, and scientific reasoning. Sol is the default for Plus and Pro standard chat. Sol Pro (Pro $200 only) allocates more compute for extended reasoning tasks.
GPT-5.6 Terra is positioned as delivering performance competitive with GPT-5.5 at roughly half the cost. Terra appears in ChatGPT Work and Codex agentic tasks on paid plans, not in the standard chat interface.
GPT-5.6 Luna is the fastest and most efficient variant. Free and Go users get Luna as their standard chat model. OpenAI expanded Luna access to free users in August 2026, at which point text chat became unlimited on Free and Go. Luna includes a Think button for structured reasoning on harder questions.
One operational implication: the Think button and Think slider (more granular on Plus/Pro) do not switch models. They tell the current model to spend more compute before answering. Knowing this separates "I should use a smarter model" from "I should ask this model to think harder," which are different choices.
Configure ChatGPT Before You Prompt It
Two settings change every subsequent conversation. Set them once and they pay dividends every session.
Custom instructions (Settings > Personalization) let you define your role, preferred output format, and standing constraints. ChatGPT applies them globally to every chat. A working template for a practitioner: "I'm a product manager at a B2B SaaS company. Use precise terminology. Keep answers under 250 words unless I ask for depth. Flag when you're uncertain rather than speculating." Without custom instructions, you spend the opening message of every conversation re-establishing context ChatGPT should already have. The full field list and length limits are in the custom instructions help article.
Projects are persistent workspaces that carry memory, uploaded reference files, and workspace-specific instructions across every chat inside them. OpenAI describes them as: "Group together chats, upload reference files, and add custom instructions so ChatGPT remembers what matters and stays on-topic." Create one project per ongoing workstream -- a client account, a research domain, a recurring report. Project memory is independent of global memory, so context from personal conversations does not surface in professional work. File size limits and usage details are in the Projects in ChatGPT help article.
Global memory stores details from your conversations automatically. If ChatGPT is working from stale or wrong assumptions, go to Settings > Personalization > Manage Memory, read what it has stored, and delete what is incorrect.
How to Prompt for Specific Results
OpenAI's prompt engineering guide for ChatGPT identifies three components that drive quality: context (what you're working on), instructions (exactly what you want), and constraints (tone, length, format rules).
Omitting context is the most common failure. "Summarize this document" produces a generic summary. "Summarize this Series B pitch deck for a venture associate tracking healthcare SaaS companies, focusing on unit economics and competitive moat" produces a targeted one. The same model, the same document, a meaningfully different result.
Iteration outperforms prompt perfection. A rough first prompt plus a pointed follow-up -- "that's too long, cut by half, keep only the hard numbers" -- reaches a usable result faster than spending ten minutes on the initial message.
Format instructions work reliably when they are explicit. If you need a table, say "return this as a markdown table with these columns." If you need JSON, specify the schema. ChatGPT will follow explicit formatting instructions; it does not volunteer structure unless you ask for it.
File uploads outperform text dumps for long content. Upload the PDF and reference specific pages in your question rather than pasting thousands of words into the chat window.
When you turn on Think for a question, give it a question worth the extra latency -- ambiguous analysis, multi-step logic, a case where a confident wrong answer is worse than a slower right one. Turning Think on for simple factual queries adds wait time with no benefit.
Power Features: When Each One Is Worth Using
Deep Research (Plus: 25 tasks/month; Pro: 250 tasks/month) runs a multi-step autonomous research process and returns a cited report. It is the right tool for synthesizing technical documentation, academic literature, or competitive landscapes -- tasks where breadth and source tracing matter. Budget 5 to 30 minutes per task. You can now restrict Deep Research to specific trusted domains or connect it to MCP servers, which matters if you want it drawing from internal knowledge bases rather than the open web. Deep Research is a form of agentic AI: it plans its own research steps and executes them autonomously.
Agent Mode (Plus and above) gives ChatGPT the ability to take actions in your browser, confirmed by you before irreversible steps. Current use cases include form-filling, multi-step web navigation, and extracting data from web UIs. Think of it as an on-ramp to AI work automation -- supervised, with explicit confirmation steps, not a fully autonomous system.
Canvas opens a shared document workspace. You and ChatGPT work on the same document: it makes targeted edits you can accept or reject, and it can do full rewrites on request. Most useful for iterative long-form writing and code that needs refinement across several passes rather than a single-prompt answer.
Advanced Voice Mode (Plus and above, daily limits apply) uses native audio processing with emotional variation and real-time response. Useful for thinking out loud, mobile use, or accessibility. Daily usage is capped for Plus users; Pro users get higher allocations. It is not the right tool for tasks that require a copyable written output.
What ChatGPT Still Gets Wrong
ChatGPT has no live data by default. Web search must be enabled per-session, or you must use Deep Research (which always uses web access). For current market prices, live API documentation, or breaking news, enable web search explicitly before asking.
Long-running single conversations can lose coherence. For complex multi-session work, use Projects with uploaded reference files and start fresh chats within the project rather than extending one conversation indefinitely.
ChatGPT cannot execute code on your local machine, call internal APIs, or interact with systems outside its native tools without Agent Mode and integrations. For terminal-integrated coding tasks, tools like Claude Code handle local execution better than ChatGPT's browser-based interface.
Mathematical reasoning has improved significantly with Sol and reasoning models, but long arithmetic chains still produce errors. Verify outputs that feed into financial models or scientific calculations.
FAQ
What is the difference between the Think button and Deep Research?
Think tells the current model (Sol or Luna) to spend more compute reasoning through your question before answering -- it takes seconds, not minutes, and operates entirely within your chat window. Deep Research launches a separate autonomous process that browses the web, synthesizes sources, and returns a cited report; it takes 5 to 30 minutes per task.
Can I share a ChatGPT conversation with someone who does not have an account?
Yes. You can generate a shareable link from any conversation and anyone can view it without logging in. The recipient cannot continue the conversation from that shared link; they can only read it.
Does ChatGPT remember things between sessions by default?
Yes, if memory is enabled, which it is by default on all plans. You can review what it has stored in Settings > Personalization > Manage Memory, delete individual entries, or turn memory off entirely. Project memory persists independently within each project, separate from global memory.
Is API access included with a ChatGPT subscription?
No. ChatGPT at chatgpt.com and OpenAI API access at platform.openai.com are separate products with separate billing. As of July 2026, GPT-5.6 Sol is priced at $5 per million input tokens and $30 per million output tokens via the API, with a promotional rate of $4 input / $20 output through November 21, 2026. GPT-5.6 Luna is $0.20 input / $1.20 output.
How does ChatGPT compare to Claude and Gemini for daily work?
Each frontier model has different strengths in coding, instruction-following, and long-context tasks. For a current side-by-side breakdown, see the ChatGPT vs Claude vs Gemini guide.
Get AI Weekly free, three issues a week, read by 40,000+ practitioners.