ChatGPT New Features 2026: Everything That Changed and What's Coming

Last Updated: July 17, 2026

GPT-5.6 became generally available on July 9-10, 2026, rolling out to ChatGPT Plus users after launching in limited preview on June 26. GPT-5.4 is retiring July 23, 2026. ChatGPT Work launched July 9 as a new product powered by GPT-5.6 that integrates team tools to turn notes and drafts into finished work - available to all plans on desktop. Codex is now built directly into the ChatGPT desktop app on macOS and Windows rather than existing as a separate application. GPT-5.6 introduces a 1.5 million token context window on Enterprise tier - 3.75 times the GPT-5.5 ceiling of 400,000 tokens. The tool-call refusal rate dropped below 4% on agent benchmarks, down from approximately 12% on GPT-5.5.

The current default ChatGPT model is GPT-5.5 Instant for most users, with GPT-5.6 rolling out in stages. GPT-5.5 produces 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts in medicine, law, and finance. Dreaming V3 - ChatGPT's rebuilt automatic memory system - launched June 4, 2026. Lockdown Mode for prompt injection protection launched June 4. GPT-4.5 retired June 26. OpenAI o3 retires August 26, 2026.

If you read the March 2026 version of this article, your mental model of ChatGPT is already out of date. GPT-5.4 - the current frontier model as of March - has been superseded by GPT-5.5, which became the default model on May 5, 2026. GPT-5.6 launched in limited preview on June 26, 2026 - literally yesterday - with an entirely new naming system and a new ultra reasoning mode that goes beyond single-agent architecture.

The most important thing to understand about ChatGPT in June 2026 is that the pace of model releases has accelerated beyond what most users can track. Three model generations were retired in four months. Any workflow you built on a named model needs re-testing. This is not a criticism - it is a practical operating reality for teams using ChatGPT professionally.

This guide covers every significant ChatGPT update from March through June 2026 chronologically, explains what each change means practically for business users, and cuts through OpenAI's changelog language to tell you what actually matters for getting work done today.

🎯 Before you read on - we put together a free 2026 AI Tools Cheat Sheet covering the tools business leaders are actually using right now. Get it instantly when you subscribe to AI Business Weekly.

Get the Free Cheat Sheet

The GPT-5 model family has moved faster than any previous OpenAI release cycle. Here is where things stand as of June 27, 2026, sourced from OpenAI's official release notes:

Currently active in ChatGPT:

Recently retired:

Upcoming retirement:

The pattern is clear: OpenAI is running approximately 90-day model lifecycles. Models generally remain available in ChatGPT for 90 days after a successor is released, then retire. Any workflow you built on a specific named model should be tested quarterly.

Three significant retirements happened in the three months since the March article.

GPT-4o retired February 13, 2026. This is worth noting because GPT-4o had developed a devoted user base who valued its personality and conversational warmth. When OpenAI retired it alongside GPT-4.1, GPT-4.1 mini, and o4-mini, there was significant user backlash - some users described GPT-4o as their "best friend." OpenAI proceeded with the retirement anyway, routing existing conversations to GPT-5.2. For business teams, this was the final confirmation that GPT-4-era evaluation data is obsolete.

GPT-5.2 retired June 12, 2026. Existing conversations now automatically continue on the corresponding GPT-5.5 model. If your organization's deep research workflows were running on GPT-5.2, they are now on GPT-5.5 - which is an upgrade in most cases.

GPT-4.5 retired June 26, 2026 - yesterday. Existing conversations continue on GPT-5.5. This applies to ChatGPT only - no changes to the API.

The practical implication for enterprise teams: if you have custom GPTs, automated workflows, or documented prompting strategies that reference specific model names, audit them now. They may be silently running on a different model than intended.

For context on how ChatGPT's model evolution compares to Claude and Gemini's release cadences, our ChatGPT vs Claude comparison covers the current competitive picture.

GPT-5.5 Instant became the default ChatGPT model on May 5, 2026, replacing GPT-5.3 Instant. Per OpenAI's GPT-5.5 Instant announcement, the model produces 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts covering medicine, law, and finance.

That hallucination reduction is the headline number for business users. In domains where accuracy matters most - legal analysis, financial modeling, medical queries - GPT-5.5 Instant is meaningfully more reliable than what most teams were using three months ago.

Other GPT-5.5 improvements:

Personalization from past chats: GPT-5.5 Instant introduced the ability to draw from past conversations, files, and connected Gmail to personalize responses. This feature rolled out to Plus and Pro users first, then expanded to Free and Go users on June 9, 2026. Per OpenAI's announcement, when a response is personalized, users can see what context was used and delete or correct it.

Canvas removed: As of the GPT-5.5 update, canvas is no longer available in GPT-5.5 Instant or GPT-5.5 Thinking. Writing and coding functionality now works directly in chat responses through writing blocks and code blocks - a cleaner integration that removes the context switch that canvas required.

GPT-5.6 launched in limited preview on June 26, 2026, per 9to5Mac's coverage. This is the most architecturally significant ChatGPT release since GPT-5.

New naming system: GPT-5.6 introduces three tiers named after celestial bodies:

OpenAI describes this naming system as identifying "durable capability tiers that can advance on their own cadence" - signaling that Sol, Terra, and Luna will persist as tier names even as model versions increment.

New reasoning modes: GPT-5.6 Sol introduces "max" and "ultra" reasoning effort options. Ultra mode goes beyond single-agent architecture by leveraging subagents to accelerate complex work - a significant step toward the multi-agent reasoning that has been available in dedicated research models.

API pricing for GPT-5.6:

Availability: Currently in limited preview to a select group of trusted partners and organizations. OpenAI plans broader availability to ChatGPT, Codex, and API users soon. The staggered release involves coordination with the US government.

This is too new to evaluate in production, but the architecture direction is clear: OpenAI is moving toward tiered intelligence levels with multi-agent ultra reasoning at the frontier.

July 2026 update: GPT-5.6 moved from limited preview to general availability on July 9-10, 2026. Enterprise API access expanded July 7-14. ChatGPT Plus users received GPT-5.6 access from July 10. The GA release introduced a 1.5 million token context window on Enterprise tier - 3.75 times the GPT-5.5 ceiling of 400,000 tokens - and a tool-call refusal rate below 4%, down from approximately 12% on GPT-5.5 per CursorBench June 2026 preview data. GPT-5.4 is retiring July 23, 2026, per OpenAI's model release notes. ChatGPT Work - a new product powered by GPT-5.6 that brings together context from team tools to turn notes and drafts into finished work - launched July 9 and is available to all plans on desktop, rolling out to Plus, Pro, Business, Enterprise, and Edu on web and mobile.

💡 Finding this helpful? Get bite-sized AI news and practical business insights like this delivered free every morning at 7 AM EST.

Subscribe Free

The biggest product change in ChatGPT since March 2026 is not a model update. It is memory.

On June 4, 2026, OpenAI launched Dreaming V3 - a completely rebuilt memory architecture that replaces the manually curated saved-memories list as ChatGPT's core memory system. Per OpenAI's Dreaming announcement, the old system required you to explicitly tell ChatGPT what to remember. The new system learns automatically from your conversations without explicit instructions.

What Dreaming V3 does differently:

The old saved-memories system wrote notes only when you prompted it with phrases like "remember that I..." and those memories went stale as circumstances changed. Dreaming V3 runs a background process that synthesizes memory from many conversations simultaneously, automatically captures context that arises naturally, and - crucially - updates memories as time passes.

OpenAI's canonical example: a memory reading "you're going to Singapore in July" automatically rewrites itself to "you went to Singapore in July 2026" after the trip ends. No user action required. That temporal awareness is the headline capability.

The performance numbers from OpenAI's memory eval:

Who gets it: Plus and Pro users in the US from June 4, 2026. Plus users get 2x memory capacity. Free and Go users are receiving it in the coming weeks. Not yet available in EU, UK, Switzerland, Norway, Iceland, and Liechtenstein due to regulatory requirements.

The privacy consideration: Dreaming V3 is more powerful and requires more active management than the old system. Users can view their memory summary page, edit or delete specific memories, correct details, or turn memory off entirely in Settings > Memory. For enterprise teams handling client data, audit your memory settings before the full rollout reaches your users.

For broader context on how AI memory and personalization capabilities are developing across platforms, our what is AI automation guide covers the underlying technology.

Launched June 4, 2026 alongside Dreaming V3, Lockdown Mode is an optional security setting now available to all logged-in ChatGPT users across account types and workspaces, per OpenAI's release notes.

What Lockdown Mode does: When enabled, ChatGPT restricts network-enabled capabilities including live web browsing, Deep Research, agent mode, file downloads, and some web-derived image support. The goal is to reduce the risk of data exfiltration from prompt injection attacks - where malicious content in a document or webpage attempts to manipulate ChatGPT into leaking your data to external sources.

When to use it: Enable Lockdown Mode when working with sensitive data - client files, financial documents, legal materials, proprietary company information. Disable it for everyday use where web browsing and Deep Research are needed.

What Lockdown Mode does not do: OpenAI is explicit that it is not a complete shield. Injected instructions in cached content or uploaded files can still influence response behavior or accuracy. It severs the exfiltration pathway, not the injection entry point.

Why this matters for enterprise teams: This is OpenAI's clearest acknowledgment that prompt injection is a real production security threat, not a theoretical concern. If your organization uses ChatGPT connectors to access company data in Google Drive, SharePoint, or Salesforce, Lockdown Mode is worth understanding before the next security review. For how AI security and data privacy considerations fit into the broader enterprise AI picture, our AI adoption statistics article covers what enterprises are doing to address these concerns.

When the March article was written, Codex was an emerging agentic coding tool. By June 2026, it has 5 million weekly active users - a 6x increase since the desktop app launched in February 2026 - per TechCrunch's developer coverage. Knowledge workers now represent approximately 20% of Codex users and are growing more than 3x faster than developers - a signal that Codex is crossing from developer tool into general-purpose agentic assistant.

Major Codex updates since March 2026:

Codex Remote generally available: As of June 2026, Codex Remote is generally available on all ChatGPT plans - including free. From the ChatGPT mobile app, users can start or continue work on a connected Mac or Windows host, review progress, and approve actions from their phone, per OpenAI's release notes. Remote Control uses authenticated one-to-one QR pairing between mobile device and host.

Goal mode generally available: Available across the Codex app, IDE extension, and CLI. Users can define an outcome and success criteria and let Codex keep working toward it autonomously. This is the most significant agentic capability addition - Codex can now pursue multi-step goals rather than waiting for explicit instruction at each step.

Appshots for macOS: Attach an app window to a Codex thread with a hotkey, including a screenshot and available text, so Codex can understand what you are looking at without a long setup prompt. Reduces the context-setting friction that slowed down complex Codex tasks.

GPT-5-Codex-Mini: A smaller, more cost-efficient version of GPT-5-Codex that provides up to 4x more usage as part of your ChatGPT subscription. Codex automatically offers to switch to Mini when you reach 90% of your 5-hour usage limit - letting you work longer without interruptions.

DigitalOcean Droplet Workspace plugin: Lets Codex provision a DigitalOcean Droplet, configure SSH access, and connect it as a remote workspace directly from Codex.

6 new enterprise plugins: OpenAI shipped plugins covering data analytics, creative production, sales, product design, equity investing, and investment banking. Partners include Wix, Replit, Figma, and Lovable.

For context on how Codex compares to Claude Code and GitHub Copilot in the developer productivity landscape, our AI coding tools statistics covers the full competitive market data.

Simplified model picker launched June 10, 2026. ChatGPT Business and Enterprise got an updated model picker on web, iOS, and Android. The new picker includes six options with clearer naming:

Thinking Light is no longer available. Instant can automatically switch to Medium when a request benefits from more reasoning. This is a meaningful UX improvement - the old Instant/Thinking framing confused users who had to decide upfront whether a task needed "thinking" before they knew how complex the response would be.

New speech-to-text model for dictation: Rolling out to all plans. Behind-the-scenes update with no change to how dictation works. Word error rate is at least 10% lower for top languages tested than the previous model. Improvements are largest in Japanese, Korean, Chinese, Urdu, Vietnamese, accented English, and long-form Spanish, French, Italian, and Portuguese.

Plugin admin controls for Business: Workspace owners and admins can now manage plugins from Workspace settings, including search and filters for status, installation policy, roles, and category. Brings plugin governance into a single management surface.

Pulse - ChatGPT's proactive daily updates feature - is being sunset as proactive updates move into scheduled Tasks, per OpenAI's release notes. Pro users had 14 days of continued access from the sunset announcement. To keep receiving daily updates, users can ask ChatGPT to schedule a daily briefing based on their interests and past chats.

Tasks - ChatGPT's scheduled prompt feature - is now the core infrastructure for recurring updates. Users can set one-off or recurring prompts and receive push or email notifications when complete, whether or not they are online. A daily news brief, weekly market summary, or recurring research task can be configured once and run automatically.

This is the quiet bridge between chatbot and assistant. For knowledge workers who previously relied on Pulse for daily AI briefings, migrating to a configured Task is straightforward and gives more customization than Pulse offered.

The aggregate picture of ChatGPT in June 2026 is a platform that is evolving faster than most organizations' evaluation cycles. Here is the practical summary for business teams.

Immediate actions:

The Dreaming V3 memory change is the most significant for daily users. ChatGPT is about to know more about your team's work patterns, preferences, and projects than any previous version. That is genuinely useful for organizational productivity and requires deliberate governance for data that should not persist.

Codex's expansion to knowledge workers is the most significant for non-developers. The 20% of Codex users who are knowledge workers growing 3x faster than developers signals that agentic task automation is crossing from developer infrastructure into general-purpose business tooling. If you have not evaluated Codex for non-coding workflows, the June 2026 state of the product warrants a fresh look.

For a complete breakdown of which ChatGPT plan provides access to which features, our ChatGPT Enterprise vs Plus comparison covers the full tier breakdown. For context on how ChatGPT's 2026 feature set compares to Claude and Gemini across specific business workflows, our ChatGPT vs Claude comparison covers the current competitive picture.

What is ChatGPT? Complete Guide 2026
The full ChatGPT background - how it works, model history, and capabilities overview.

ChatGPT Enterprise vs Plus: Which Plan?
Which plan gets you access to which GPT-5.5 model variants, Codex, Deep Research, and the latest 2026 features.

OpenAI Statistics 2026
The business context for ChatGPT's development - users, revenue, and the competitive pressure driving this pace of updates.

ChatGPT vs Claude: Which Is Better for Business?
How ChatGPT's 2026 updates affect the head-to-head comparison with Anthropic's Claude.

AI Coding Tools Statistics 2026
Where Codex sits in the AI coding landscape alongside Claude Code, GitHub Copilot, and Cursor.

What is Prompt Engineering?
How to get the most out of ChatGPT's latest models including GPT-5.5's improved instruction following.

AI Hallucinations: Causes and Solutions
Context on GPT-5.5's 52.5% hallucination reduction and what it means for enterprise reliability.

What is the current ChatGPT model in June 2026?
The current default ChatGPT model is GPT-5.5 Instant, which became the default on May 5, 2026, per OpenAI's release notes. GPT-5.5 produces 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts in medicine, law, and finance. GPT-5.6 launched in limited preview on June 26, 2026, with a new Sol/Terra/Luna naming system and ultra reasoning mode using subagents. GPT-5.2 was retired June 12, 2026, and GPT-4.5 was retired June 26, 2026.

What is ChatGPT Dreaming memory?
Dreaming V3 is ChatGPT's rebuilt memory architecture launched June 4, 2026. Unlike the old saved-memories system that required explicit instructions to remember things, Dreaming V3 runs a background process that automatically synthesizes memory from past conversations and updates it as circumstances change. A memory reading "you're going to Singapore in July" automatically updates to "you went to Singapore in July 2026" after the trip ends. Factual recall improved from 41.5% in 2024 to 82.8% in 2026. Available to Plus and Pro users in the US, expanding to Free and Go users in coming weeks. Not yet available in EU, UK, or Switzerland.

What is ChatGPT Lockdown Mode?
Lockdown Mode is an optional security setting launched June 4, 2026, available to all logged-in ChatGPT users. When enabled, it restricts network-enabled capabilities including live web browsing, Deep Research, agent mode, and file downloads to reduce the risk of data exfiltration from prompt injection attacks. Enable it when working with sensitive client or financial data. Disable it for everyday use where web browsing and Deep Research are needed. OpenAI notes it severs the exfiltration pathway but does not block injection entry through cached content or uploaded files.

What is GPT-5.6?
GPT-5.6 launched in limited preview on June 26, 2026. It introduces three capability tiers: Sol (highest intelligence), Terra (midrange), and Luna (cost-efficient). GPT-5.6 Sol introduces max and ultra reasoning effort options - ultra mode uses subagents to accelerate complex work beyond single-agent architecture. API pricing is $5/$30 per million tokens for Sol, $2.50/$15 for Terra, and $1/$6 for Luna. Currently available to a select group of trusted partners. Broader availability to ChatGPT, Codex, and API users is planned soon.

What happened to ChatGPT Canvas?
Canvas was removed as part of the GPT-5.5 update. Writing and coding functionality now works directly in chat responses through writing blocks and code blocks - eliminating the context switch that canvas required. This is a simplification rather than a feature removal; the same capabilities are available in a more integrated format.

How many users does Codex have in 2026?
Codex reached 5 million weekly active users by June 2026 - a 6x increase since the desktop app launched in February 2026. Knowledge workers represent approximately 20% of Codex users and are growing more than 3x faster than developers. Codex Remote became generally available on all ChatGPT plans in June 2026, enabling mobile app control of Codex tasks running on connected Mac or Windows hosts.

What is the new ChatGPT model picker?
The updated model picker launched June 10, 2026 for ChatGPT Business and Enterprise. It includes six options: Instant, Medium (replaces Thinking Standard), High (replaces Thinking Extended), Extra High (replaces Thinking Heavy), Pro Standard, and Pro Extended. Thinking Light is no longer available. Instant can automatically switch to Medium when a request benefits from more reasoning. The new naming makes it clearer what each option does without requiring upfront assessment of task complexity.

What is ChatGPT Codex Goal Mode?
Goal mode is a Codex capability that lets users define an outcome and success criteria and let Codex keep working toward it autonomously across multiple steps. It became generally available across the Codex app, IDE extension, and CLI in June 2026. This is the most significant step toward fully agentic coding workflows - Codex can now pursue goals rather than waiting for explicit instruction at each step.

What ChatGPT models were retired in 2026?
GPT-4o and GPT-4.1 were retired February 13, 2026. GPT-5.1 models were retired March 11, 2026. GPT-5.2 models were retired June 12, 2026. GPT-4.5 was retired June 26, 2026. OpenAI o3 will be retired August 26, 2026. Models generally remain available in ChatGPT for 90 days after a successor is released. These retirements apply to ChatGPT only - the API is not affected.

What is ChatGPT Work?

ChatGPT Work is a new product powered by GPT-5.6 that launched July 9, 2026, available to all ChatGPT plans on desktop. It integrates context from team tools to turn scattered notes, drafts, and ideas into finished work while keeping projects moving. ChatGPT Work and ChatGPT Codex share plugins and both live inside the ChatGPT desktop app on macOS and Windows. The existing ChatGPT app is now called ChatGPT Classic. ChatGPT for desktop includes both ChatGPT Work and ChatGPT Codex, with Codex mode showing more technical details that Work abstracts away from the user.

When is GPT-5.6 available to ChatGPT Plus users?

GPT-5.6 became generally available to ChatGPT Plus users from July 10, 2026, after launching in limited preview on June 26. Enterprise API access expanded July 7-14. GPT-5.6 introduces three capability tiers - Sol, Terra, and Luna - plus a 1.5 million token context window on Enterprise, a tool-call refusal rate below 4% on agent benchmarks, and ultra reasoning mode using subagents. GPT-5.4 is retiring July 23, 2026. API pricing: Sol at $5/$30 per million tokens, Terra at $2.50/$15, Luna at $1/$6.

What is new in ChatGPT in June 2026?
The biggest ChatGPT updates in June 2026: GPT-5.5 Instant is now the default model with 52.5% fewer hallucinations than its predecessor on high-stakes prompts, per OpenAI. GPT-5.6 launched in limited preview June 26 with Sol/Terra/Luna tiers and ultra multi-agent reasoning. Dreaming V3 rebuilt ChatGPT's memory from scratch on June 4 - now automatic and self-updating. Lockdown Mode added June 4 for prompt injection protection. Codex Remote became available on all plans. GPT-5.2 retired June 12 and GPT-4.5 retired June 26.

What is the current ChatGPT model in 2026?
GPT-5.5 Instant is the current default ChatGPT model as of June 2026, released May 5, 2026 by OpenAI. It produces 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts in medicine, law, and finance. GPT-5.6 launched in limited preview June 26, 2026, with three tiers: Sol, Terra, and Luna. GPT-5.2 was retired June 12, 2026. GPT-4o was retired February 13, 2026. GPT-5.3 Instant remains available to paid users for three months before retirement.

What is ChatGPT Dreaming and how does it work?
ChatGPT Dreaming V3 is OpenAI's rebuilt memory system launched June 4, 2026. A background process automatically synthesizes memory from past conversations without requiring explicit "remember this" instructions. Memories self-update as time passes - a note about a future trip updates itself after the trip ends. Factual recall improved from 41.5% with 2024 saved memories to 82.8% with Dreaming V3, per OpenAI's internal evals. Available to Plus and Pro users in the US, expanding to Free and Go users in coming weeks.

ChatGPT in June 2026 is a different product than it was in March - not incrementally different, but meaningfully so. Dreaming V3 changes how ChatGPT knows you. Lockdown Mode changes how safely you can use it with sensitive data. GPT-5.5's hallucination reduction changes how reliably you can trust it on high-stakes professional tasks. And GPT-5.6's ultra multi-agent reasoning, while in limited preview, signals where the frontier is heading.

The practical recommendation for any business team: update your mental model quarterly, not annually. The organizations building genuine competitive advantage with AI are the ones whose evaluation cycles match OpenAI's release cadence. If you last seriously evaluated ChatGPT before May 2026, you are missing Dreaming V3, Lockdown Mode, Codex Goal Mode, and GPT-5.5's reliability improvements. That is three months of changes that affect daily workflows meaningfully.

Schedule a structured two-week re-evaluation. Focus specifically on memory management with Dreaming V3, security posture with Lockdown Mode, and whatever workflows felt inadequate at your last assessment. In most cases, at least two or three of those limitations have been directly addressed.

📨 Don't miss tomorrow's edition. Subscribe free to AI Business Weekly and get our 2026 AI Tools Cheat Sheet instantly - bite-sized AI news every morning, zero hype.

Subscribe Free