ChatGPT crossed one billion weekly active users in August 2026, up from 900 million in February and 400 million a year before that. It is no longer a novelty you try out; it is a piece of infrastructure roughly one in eight people on Earth touch every week. The product has also changed more in the last five months than most people realize. As of September 2026 it runs on GPT-6 Astra and a three-model GPT-5.6 family, shows ads on its free tier, drives coding agents by voice, and builds working websites from a single prompt. This guide covers what ChatGPT actually is now, which model you should be on, where it still fails, and when a competitor serves you better.
The model stack is now the whole decision
“ChatGPT” is not one model. It is a router in front of several, and the one you land on decides most of what you experience. OpenAI retired GPT-4o from the product back in February 2026 and has since moved through the GPT-5.6 series to GPT-6 Astra, which began rolling out on September 4, 2026.
Here is the field as it stands:
- GPT-6 Astra is the new flagship. It carries a 1,050,000-token context window, up to 128,000 tokens of output, text and image input, and a knowledge cutoff of April 30, 2026. OpenAI calls it a generational jump for software engineering, science, and cybersecurity. It saturates FrontierMath Tier 4 at 97.6% and posts 99.9% on ARC-AGI-3. API pricing is $10 per million input tokens and $50 per million output. Access is rolling out to Plus, Pro, Business, and Enterprise.
- GPT-5.6 Sol was the flagship until Astra landed and remains a heavy-duty reasoning model. It unlocks
maxreasoning effort and anultramode that runs a subagent approach. On Terminal-Bench 2.1 it scores 88.8%, or 91.9% with ultra mode. API list price is $5 in, $30 out. - GPT-5.6 Terra is the balanced everyday model, roughly half Sol’s price at $2.50 in, $15 out, with performance competitive with last cycle’s GPT-5.5.
- GPT-5.6 Luna is the efficiency tier, $1 in, $6 out, tuned for high-volume, latency-sensitive work. It is the default model for Free and Go accounts.
The practical gap between these is large. Luna is fast and fine for everyday questions. Sol and Astra think before they answer, take longer, and pull meaningfully ahead on math, multi-step analysis, and code where being wrong has a cost. If you never touch the model picker, you are leaving most of the capability on the table on the paid tiers and running Luna on the free one. Pricing shifts often, so confirm current numbers on OpenAI’s pricing page before you commit a budget. The wider question of which tier to route which task to is its own discipline, one I covered in the AI model cost map.
Six tiers, and ads on two of them
The pricing structure has grown since the last version of this guide. In September 2026 the ladder runs:
| Tier | Price | What you get |
|---|---|---|
| Free | $0 | Unlimited text chats on GPT-5.6 Luna (announced August 6, 2026), with ads |
| Go | $8/mo | Higher limits than Free, still Luna, still ads |
| Plus | $20/mo | GPT-6 Astra and GPT-5.6 Sol, voice, image generation, agents, no ads |
| Pro | $200/mo | Maximum usage caps across everything (a $100 mid-Pro tier also exists) |
| Business | $25/user/mo | Shared workspace, admin controls, connectors (dropped from a higher price in April 2026) |
| Enterprise | Custom | Data guarantees, SSO, dedicated support |
The advertising change is the one to flag. Following OpenAI’s February 2026 ad rollout, both Free and Go now serve ads inside the interface. Plus at $20 is the first tier that removes them. That matters if you paste sensitive or client material into the free product: an ad-supported context is a different privacy posture than a paid one, and it is worth thinking about before your team standardizes on the free tier. I traced the strategic logic of that shift when ChatGPT ads hit $100 million in six weeks.
Where ChatGPT genuinely earns its keep
Cutting through the feature list, these are the areas where the product consistently delivers in September 2026.
Writing and communication. Still the mature core use case. It drafts, rewrites for tone, compresses long documents, and adapts a piece across formats faster than any person. Feed it three of your past emails and ask for a fourth in the same voice, and editing the output is now faster than starting from a blank page. For anyone shipping investor updates, board material, and internal comms on repeat, this alone covers the Plus subscription.
Code, with a running interpreter. GPT-6 Astra and GPT-5.6 Sol sit at or near the top of the major coding benchmarks. Inside a Project, ChatGPT does not just suggest code; it runs it, reads the error output, and iterates. It is genuinely useful for scaffolding, tests, and working through an unfamiliar library. It still stumbles on hard architecture decisions and can invent library functions that do not exist, so you verify before you ship. If your work is coding-first, the head to head across the IDEs in the Cursor, Copilot, and Windsurf comparison is the better starting point.
Agents and Codex. This is the biggest change since early 2026. OpenAI folded its old Operator browser tool, deep research, and core intelligence into a single agent that takes real multi-step actions: browsing sites, filling forms, chaining tools, and building a deck from a research brief. Codex, its coding agent, now keeps notes across context windows under Astra instead of repeatedly compressing everything into one summary, which is a real fix for long jobs that used to lose the thread. Codex also went fully remote, so the agent runs from your phone; I walked through that in the Codex Remote writeup. For the broader picture of what separates an agent from a chatbot, start with what an AI agent actually is.
Sites, new with Astra. With Sites in ChatGPT, Astra can create, host, and share a website, web app, or small game directly from a prompt. It is early, and the output needs review like any generated artifact, but for a quick internal tool or a landing page it collapses a multi-step workflow into one.
Full-duplex voice. The old turn-taking Advanced Voice Mode is gone. GPT-Live-1, which launched July 8, 2026, listens and speaks at the same time, and since late July it lives in the Mac and Windows desktop apps wired straight into Codex and ChatGPT Work. You can talk through a decision on a commute, or issue instructions to coding agents hands-free and hear them report back. Building a production voice agent is still a separate engineering job, a distinction I drew out in the GPT-Live breakdown.
Projects and persistent files. Projects give the model standing context (company background, your style, a fixed instruction set) so you stop re-explaining yourself. The file library saves what you upload and create so your work compounds across sessions instead of resetting each time. Both reward the habit of loading context once and reusing it, and both depend on how well you frame the instructions in the first place.
Where it still falls short
The honest limits, several of which are structural rather than a patch away.
Hallucination is reduced, not solved. The reasoning models are more careful than GPT-4o was, especially on factual claims. But ChatGPT will still state wrong things confidently, particularly on niche topics, obscure sources, and specific data points like dates, statistics, and citations. Running 20-plus years in IT operations taught me the same rule that applies here: never wire an unverified output into anything where an error carries a consequence. Keep a human verification step on the last mile.
The Critical-capability shadow. Astra is the first model OpenAI has flagged at the Critical level for cybersecurity capability under its Preparedness Framework. That is a striking admission about raw power, and it came with a self-imposed slowdown and a heavy safety stack before the wider release. I covered that decision when OpenAI paused Astra over its cyber capability, and separately when it published ten decade-old math proofs to introduce the model. For a buyer, the takeaway is not fear; it is that governance posture is now part of the model spec sheet.
Long context still degrades. A million-token window sounds like it ends the problem. It does not. On tasks that require tracking many constraints across a very long document or conversation, the model still drops earlier instructions and can contradict itself. If that is your core workload, Claude has built its reputation specifically on holding long context more consistently.
The walled garden. ChatGPT does not see your CRM, codebase, or internal systems unless you connect them through uploads or enterprise connectors, and that setup is real IT work, not a toggle. Budget for it if you expect the model to reason over your own data.
Cost stacks up. Plus at $20 is reasonable. But a Pro seat at $200, plus API spend for integrations, plus per-seat Business licenses, escalates fast for a team. Map the spend against what each task actually needs before you buy everyone the top tier.
How it compares in September 2026
The field has converged, and there is no single winner. A practitioner’s read of the current flagships:
- Versus Claude Opus 5. Anthropic’s flagship lists at $5 in, $25 out and leads coding, posting around 87.6% on SWE-bench Verified against strong but lower numbers elsewhere. ChatGPT wins on ecosystem breadth, agents, and consumer polish; Claude wins on coding accuracy and long-context reliability. Anthropic also shipped Fable 5.1 and Mythos 5.1 in early September, so the gap moves week to week.
- Versus Gemini. Gemini 3.1 Pro is Google’s shipping flagship into September (the long-promised 3.5 Pro kept slipping) at $2 in, $12 out with a 2-million-token window and GPQA Diamond around 94%. It is the multimodal and cost champion; ChatGPT counters with the agent stack. The full picture is in the Gemini guide.
- Versus Grok. xAI’s Grok 4.x is the value and real-time pick at roughly $2 in, $6 out, strong when live web context matters.
- Versus DeepSeek. DeepSeek keeps undercutting everyone on price with open weights, a serious option when you are cost-sensitive and do not need OpenAI’s agent features.
If cited sources are the point of your workflow rather than a chat, Perplexity is built around that, and the full platform landscape maps the rest.
Getting real value out of it
Four habits separate people who get leverage from ChatGPT from people who use it as a fancier search box:
- Pick the model on purpose. Luna for quick questions, Sol or Astra for anything that needs reasoning, analysis, or accuracy. The default is rarely the right call on a paid plan. The reasoning-effort settings guide covers the dials that come with these models.
- Load context once. Put every recurring workflow in a Project with standing instructions, and use the file library for brand guides, templates, and reference material so it compounds.
- Give agents real jobs. Do not just chat. Hand the agent multi-step work with a checkable outcome, then review what it produced.
- Keep the last mile human. Verify facts, numbers, and citations before anything leaves your hands. The model is a strong first draft, not a signed one.
ChatGPT in September 2026 is a genuine productivity platform, not the chatbot it started as. The smartest posture is not loyalty to one vendor; it is knowing what each tool does best and routing each task to the right one. Set up your Projects, pick your model deliberately, and treat every frontier release as a reason to re-check your routing rather than a reason to switch everything overnight.