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

Last Updated: September 1, 2026

Quick Answer: ChatGPT's biggest changes in August 2026: GPT-5.6 Sol is now the default model for Plus and Pro users with a new reasoning effort slider replacing the old Instant/Thinking split. Free and Go users now get unlimited text chats on GPT-5.6 Luna with a Think button for harder questions. OpenAI o3 retired August 26. The official DALL-E GPT retired August 30. Computer History, WebMCP site tools, and cloud browser sign-in launched. GPT-5.4 and GPT-5.4 mini retired from Codex on August 31.

ChatGPT in August 2026 is a fundamentally different product than it was in January. OpenAI has shipped four model generations, retired six models, launched ChatGPT Work as a standalone product, rebuilt memory with Dreaming V3, added security with Lockdown Mode, and made unlimited text chat free for the first time in ChatGPT's history. The pace of change has outrun most users' ability to track it.

This guide covers every significant ChatGPT update from January through August 31, 2026, organized chronologically with a clear explanation of what each change means for business users. Every retirement, every new capability, every model update - with the practical implication for teams using ChatGPT professionally.

The single most important update for business teams right now: as of August 6-7, 2026, ChatGPT's old Instant versus Thinking model split is gone. Plus and Pro users now have one model - GPT-5.6 Sol - with a slider controlling reasoning depth. And free users now get unlimited text chats for the first time.

📩 Like what you're reading? Get our free 2026 AI Tools Cheat Sheet plus the daily AI newsletter business leaders read every morning.

Get the Free Cheat Sheet →

Models retired in 2026 (complete list):

Upcoming: GPT-5.5 retirement timeline not yet confirmed - expect a 90-day sunset after OpenAI signals its successor fully deployed.

Sources: OpenAI ChatGPT Release Notes, OpenAI Model Release Notes, OpenAI What's New

For our complete ChatGPT statistics including user counts, revenue, and market share, our ChatGPT statistics guide covers every metric. For how ChatGPT pricing has changed across 2026, our AI pricing guide 2026 covers every plan.

The August 2026 updates are the most significant consumer-facing changes ChatGPT has made since GPT-5 launched - eliminating the Instant/Thinking model split, making unlimited text chat free, and launching a wave of new capabilities across Work, Voice, and Tasks.

Quick Answer: August 6-7, 2026 is the date that matters most. GPT-5.6 Sol replaced the old Instant/Thinking split for Plus/Pro with a reasoning slider. GPT-5.6 Luna became the free default with unlimited text chats. OpenAI o3 retired August 26. DALL-E GPT retired August 30. GPT-5.4 and mini retired from Codex August 31. Computer History, WebMCP, and cloud browser sign-in all launched.

The biggest ChatGPT interface change of 2026: OpenAI eliminated the separate Instant and Thinking models for Plus and Pro users. Both now run on a single GPT-5.6 Sol, with a reasoning effort slider that lets users control how much thought the model applies to each response.

Until this update, ChatGPT routed everyday conversations between a fast Instant model and a slower reasoning model - a split that required users to decide upfront whether a task needed "thinking" before knowing how complex the answer would be. As of August 6, 2026, one model handles both, with the slider making that decision explicit and user-controlled rather than model-determined.

The updates to GPT-5.6 Sol in ChatGPT are designed to give more direct responses, use tighter formatting, and avoid extra detail when it does not help. For a quick question, that means a direct answer with the context needed. For more involved work like multi-step planning, research, or writing, it means a fuller response that keeps the main recommendation clear.

For Free and Go users: GPT-5.6 Luna became the default model with unlimited text chats - the first time ChatGPT has offered truly unlimited text conversations at no cost. A Think button gives free users access to deeper reasoning on harder questions, subject to abuse guardrails. Limits still apply for file uploads, images, and other tools.

The practical implication: Plus and Pro users no longer need to choose between Instant and Thinking at the start of a session. Set the slider to low effort for quick questions, high effort for complex analysis. Free users now have genuinely unlimited text access to a GPT-5.6 generation model.

OpenAI previewed Ultrafast mode for GPT-5.6 Sol on August 13, 2026. The preview can run GPT-5.6 Sol at up to 14 times the speed. This is a preview feature, not yet generally available. For teams running time-sensitive workflows where response latency is a bottleneck, Ultrafast mode is worth watching as it approaches broader availability.

Starting August 19, 2026, OpenAI updated ChatGPT with a more capable model and expanded access for everyone. Free and Go users got a new default model for everyday chats. Plus and Pro users got an updated GPT-5.6 Sol with a slider that lets them choose how much effort ChatGPT uses for a response. These models replace GPT-5.5 Instant. The updated Sol produces 68% fewer errors than the previous version per ExplainX's August analysis - the most significant reliability improvement since GPT-5.5's 52.5% hallucination reduction in May. Openai

OpenAI updated ChatGPT Work on August 25, 2026, with support for websites that require sign-in on web and mobile. The user completes the sign-in without ChatGPT seeing the username or password. This cloud browser sign-in capability means ChatGPT Work can now assist with tasks inside authenticated web applications without users exposing credentials in chat. The sign-in flow runs in a separate cloud browser session from the user's local browser. Zen MCP Server

OpenAI o3 retired from ChatGPT on August 26, 2026, following its 90-day sunset period. OpenAI o3 was retired from ChatGPT on August 26, 2026 following a 90-day sunset period. Existing o3 conversations now continue on GPT-5.6 Sol. For teams that built workflows specifically on o3's reasoning architecture, audit those workflows now - GPT-5.6 Sol's reasoning slider covers the same capability space with a unified interface. OpenAI Help Center

On August 30, 2026, OpenAI retired the official DALL-E GPT in ChatGPT. Users are encouraged to download any images they want to keep. To continue creating or editing images, use ChatGPT Images. User-created GPTs with image generation enabled are not affected. ChatGPT Images is the replacement for all image generation and editing within ChatGPT - the DALL-E GPT wrapper is no longer needed as image generation is now native to the main interface. Releasebot

On August 31, GPT-5.4 and GPT-5.4 mini retire from Codex for users signed in with ChatGPT. Replace gpt-5.4 with gpt-5.6-terra and gpt-5.4-mini with gpt-5.6-luna in workspace defaults, saved model settings, managed configurations, custom agents, and scheduled tasks. Developers with API keys are not affected - only ChatGPT-authenticated Codex users need to update their configurations. ChatGPT Learn

Computer History turns activity across selected apps and websites into memories and a timeline that ChatGPT and Codex can reference later. You can return to recent work without reconstructing every open app, document, and next step. Computer History is off by default in the ChatGPT desktop app - users must explicitly enable it. For teams wanting to reduce context-setting friction at the start of each ChatGPT session, Computer History provides persistent workspace awareness without manual setup. Zen MCP Server

With Site tools (WebMCP), ChatGPT Work and Codex can use actions offered by a website in the desktop app's built-in browser. For example, a document editor can provide tools to find a section or add a comment. Update the desktop app and use GPT-5.6 Sol or GPT-5.6 Terra. Site tools are not available with GPT-5.6 Luna or in Enterprise or Edu workspaces. ChatGPT Learn

WebMCP is the most technically significant infrastructure addition in August 2026. It gives ChatGPT agents a direct interface with websites through shared live pages and signed-in sessions - moving from AI that reads web content to AI that can interact with web applications as a user would. For developers, this opens an entirely new category of workflow automation.

Scheduled tasks can now respond when something changes in Slack, Gmail, or GitHub for Plus and Pro users. OpenAI also began rolling out tasks to Free users, with a limit of three tasks, and added task sharing so other people can customize and use a shared task. Tasks have evolved from a scheduled prompt system into a trigger-based automation layer. A task that fires when a specific email arrives in Gmail, when a PR is opened in GitHub, or when a message appears in Slack is a genuinely different capability from a scheduled cron-style task. Zen MCP Server

ChatGPT Business now offers Premium seats with five times the usage of Standard seats and no five-hour usage limit. A Premium seat costs $100 per user per month with annual billing or $125 with monthly billing. For enterprise teams whose power users consistently hit usage limits on standard Business seats, Premium seats eliminate the five-hour cap that interrupted workflows. Zen MCP Server

GPT-Live in ChatGPT Voice now supports file uploads and Projects. You can now upload files in a voice conversation and analyze its contents or ask questions. You can also use voice in Projects, referencing recent project chats, sources, and project instructions. Voice has crossed from a conversational interface into a full-capability workspace tool. Uploading a document and analyzing it through voice conversation within a Project context is a meaningfully different interaction than asking ChatGPT questions verbally. Releasebot

GPT-5.6-Cyber is a cybersecurity-specific model available to approved defenders for advanced vulnerability research, exploit validation, and authorized security testing. Available exclusively to approved Daybreak Red users. For security teams that qualify, this is the first GPT model purpose-built for the specific reasoning patterns that cybersecurity work requires. Zen MCP Server

ChatGPT adds Sign in with ChatGPT beta across select plugins and partner sites, including Airtable, GitLab, HubSpot, Notion, Supabase, and Vercel, making account setup faster and access simpler from ChatGPT and Codex. This OAuth-style sign-in positions ChatGPT as an identity layer across the productivity tool ecosystem - a notable infrastructure shift. Releasebot

July 2026 was the month ChatGPT became a work platform rather than a chat interface - with GPT-5.6 reaching general availability, ChatGPT Work launching, and the product splitting into three distinct experiences: Classic, Work, and Codex.

Quick Answer: GPT-5.6 hit general availability July 9-10, 2026. ChatGPT Work launched July 9 as a standalone agentic product powered by GPT-5.6. GPT-5.4 retired July 23. OpenAI cut Luna and Terra API pricing 80% on July 30. The existing ChatGPT app was renamed ChatGPT Classic.

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.

Key GPT-5.6 capabilities at GA:

ChatGPT Work launched July 9, 2026, available to all ChatGPT plans on desktop. Powered by GPT-5.6, it integrates context from team tools to turn scattered notes, drafts, and project context into finished deliverables.

ChatGPT Work and ChatGPT Codex share plugins and both live inside the ChatGPT desktop app on macOS and Windows. The existing ChatGPT app was renamed ChatGPT Classic. The desktop app now includes three distinct experiences: ChatGPT Classic (chat), ChatGPT Work (agentic task completion), and ChatGPT Codex (coding and development).

ChatGPT Work is built for knowledge workers rather than developers. Where Codex shows technical details, Work abstracts them away. Powered by GPT-5.6, it can break a goal into steps and work for hours while users follow progress, answer questions, change direction, and approve important actions.

GPT-5.4 retired from ChatGPT on July 23, 2026, per OpenAI's model release notes. Existing conversations continue on GPT-5.6. For teams with custom GPTs or workflows referencing GPT-5.4 by name, audit configurations and update to GPT-5.6-terra as the equivalent.

OpenAI cut Luna and Terra API pricing by approximately 80% on July 30, 2026. Luna dropped to approximately $1 per million input tokens. Terra dropped to approximately $2.50 per million input tokens. For developers who had been evaluating DeepSeek's $0.14 per million token pricing as a competitive alternative, the July 30 price cut significantly closed the cost gap between OpenAI's efficient tiers and the lowest-cost frontier alternatives.

For the full API pricing comparison including DeepSeek, Claude, and Gemini, our AI pricing guide 2026 covers every platform.

June 2026 brought three foundational changes that affect every ChatGPT user - a rebuilt memory system, a new security model, and the first preview of GPT-5.6's tiered architecture.

Quick Answer: June 4: Dreaming V3 rebuilt ChatGPT memory from scratch with automatic, self-updating memory (factual recall improved from 41.5% to 82.8%). June 4: Lockdown Mode launched for prompt injection protection. June 26: GPT-5.6 previewed with Sol/Terra/Luna tier naming. June 26: GPT-4.5 retired. June 12: GPT-5.2 retired.

Dreaming V3 is ChatGPT's rebuilt memory architecture. Unlike the old saved-memories system that required explicit "remember this" instructions, Dreaming V3 runs automatically in the background and updates as circumstances change.

OpenAI's performance numbers from internal evaluation:

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.

Available to Plus and Pro users in the US. Not yet available in EU, UK, Switzerland, Norway, Iceland, and Liechtenstein due to regulatory requirements.

The governance consideration: Dreaming V3 requires active management for enterprise teams. Users can view their memory summary page, edit or delete specific memories, and turn memory off in Settings > Memory. Any team handling client data should audit memory settings before Dreaming V3 reaches their full user base.

Lockdown Mode is an optional security setting available to all logged-in ChatGPT users. When enabled, it restricts network-enabled capabilities including live web browsing, Deep Research, agent mode, file downloads, and some web-derived image support to reduce the risk of data exfiltration from prompt injection attacks.

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.

When to use it: Enable for sessions involving sensitive client or financial data. Disable for everyday use where web browsing and Deep Research are needed.

For how AI security and prompt injection fit into the broader enterprise security picture, our AI for cybersecurity guide covers the full threat landscape.

GPT-5.6 launched in limited preview on June 26, 2026, introducing the Sol/Terra/Luna tier naming system. Sol is the highest intelligence tier, Terra is the midrange, and Luna is the cost-efficient tier for high-volume tasks. OpenAI describes these as "durable capability tiers that can advance on their own cadence" - meaning Sol, Terra, and Luna persist as tier names as model versions increment.

May 5, 2026: GPT-5.5 Instant replaced GPT-5.3 Instant as the default ChatGPT model, delivering 52.5% fewer hallucinated claims on high-stakes prompts in medicine, law, and finance - the most significant reliability improvement in ChatGPT's history to that point.

Key GPT-5.5 improvements:

GPT-5.5's 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 delivered meaningfully more reliable outputs than the model most teams were running on three months prior.

For our complete guide to AI hallucinations and how GPT-5.5's improvement fits into the broader accuracy picture, our AI hallucinations guide covers every mitigation method.

The first four months of 2026 saw three model generations - GPT-5.1, GPT-5.2, GPT-5.3 - launch, serve as default, and begin their retirement cycle in rapid succession, establishing the 90-day model lifecycle that now defines ChatGPT's development cadence.

Model retirement pattern established:

OpenAI is running approximately 90-day model lifecycles in 2026. Models generally remain available in ChatGPT for 90 days after a successor is released, then retire. Any workflow built on a specific named model should be tested quarterly.

February 2026:

March 2026:

April-May 2026:

The pattern for enterprise teams: any evaluation data you collected on a named GPT model before May 2026 is obsolete. The current baseline is GPT-5.6 Sol.

For our complete guide to what ChatGPT is, how it works, and its full history, our what is ChatGPT guide covers every generation.

Six actions that matter immediately for any team using ChatGPT professionally in August 2026.

1. Update your model references in Codex configurations
GPT-5.4 and GPT-5.4 mini retired from Codex on August 31, 2026. If your team has workspace defaults, saved model settings, managed configurations, custom agents, or scheduled tasks referencing gpt-5.4 or gpt-5.4-mini, update them to gpt-5.6-terra and gpt-5.6-luna respectively. API-authenticated Codex sessions are not affected.

2. Learn the reasoning slider - it replaces your model choice
Plus and Pro users no longer choose between Instant and Thinking. The GPT-5.6 Sol reasoning slider handles both. Low effort for quick questions. High effort for complex analysis. This is a simpler interface than the old model picker but requires adjusting the habit of selecting a model before starting a session.

3. Evaluate WebMCP for workflow automation
Site tools (WebMCP) lets ChatGPT Work and Codex interact with websites as a user would - through shared live pages and signed-in sessions. For teams using ChatGPT alongside web-based tools like HubSpot, Notion, or GitHub, this is a new category of automation worth evaluating. Requires GPT-5.6 Sol or Terra and the latest desktop app.

4. Configure trigger-based Tasks for recurring monitoring
Scheduled Tasks now fire on Slack messages, Gmail arrivals, and GitHub events for Plus and Pro users. A task that monitors a Gmail label for client updates and drafts a summary response is a different category of assistant than a cron-style scheduled prompt. Set up two or three trigger-based tasks this week and evaluate the workflow change.

5. Audit memory settings before broader rollout
Dreaming V3's automatic memory is significantly more powerful than the old saved-memories system. For teams handling client data, proprietary research, or any information that should not persist across sessions, audit Settings > Memory now and establish team standards for what should and should not be stored.

6. Test GPT-5.6 Sol's 68% error reduction on your highest-stakes workflows
The August 19 GPT-5.6 Sol update delivered 68% fewer errors than the previous version. For teams using ChatGPT in legal analysis, financial modeling, or medical contexts where the May GPT-5.5 reliability improvement already changed what was acceptable to use ChatGPT for - retest those workflows. The bar has moved again.

For our complete guide on implementing AI effectively across business workflows, our how to implement AI in business guide covers the organizational framework.

What Is ChatGPT?
The complete ChatGPT guide - how it works, every model generation, and capabilities overview.

ChatGPT Statistics 2026
900 million weekly users, $40 billion ARR, and every ChatGPT metric in one place.

ChatGPT Enterprise vs Plus
Which plan provides which GPT-5.6 capabilities, Codex access, context windows, and governance controls.

ChatGPT vs Claude
How ChatGPT's August 2026 updates affect the head-to-head comparison with Anthropic Claude.

AI Pricing Guide 2026
GPT-5.6 Sol, Terra, and Luna pricing after the July 30 cut - plus Claude, Gemini, Grok, and DeepSeek.

AI Hallucinations: Causes and Solutions
GPT-5.5's 52.5% and GPT-5.6's 68% hallucination reductions in the broader context of AI reliability.

AI Coding Tools Statistics 2026
Where Codex sits in the AI coding landscape after its August 31 model update.

How to Use ChatGPT
Practical workflows for GPT-5.6 Sol including the reasoning slider, ChatGPT Work, and Codex.

AI for Cybersecurity
How Lockdown Mode and GPT-5.6-Cyber fit into enterprise AI security strategy.

OpenAI Statistics 2026
The business context behind these product decisions - $40 billion ARR and the IPO timeline.

What is the current ChatGPT model in August 2026?
GPT-5.6 Sol is the current default model for ChatGPT Plus and Pro users as of August 6-7, 2026, replacing the old Instant and Thinking model split with a single model and a reasoning effort slider. GPT-5.6 Luna is the default for Free and Go users with unlimited text chats as of August 7, 2026. Enterprise users have GPT-5.6 Sol with a 1.5 million token context window. GPT-5.6 Terra is available for API use and Codex. GPT-5.6-Cyber is available to approved Daybreak Red cybersecurity users only. GPT-5.5 Instant is still accessible but no longer the default. OpenAI o3 retired August 26, 2026. GPT-5.4 and GPT-5.4 mini retired from Codex August 31, 2026. Source: OpenAI Release Notes, 9to5Mac August 6, 2026

What is the GPT-5.6 reasoning slider?
The reasoning effort slider is a new interface control for ChatGPT Plus and Pro users that replaced the old Instant versus Thinking model split on August 6, 2026. Instead of choosing between a fast Instant model and a slower reasoning model before starting a session, users set the slider to control how much effort GPT-5.6 Sol applies to each response. Low effort produces fast, direct answers. High effort applies extended reasoning for complex analysis, multi-step planning, and research tasks. The slider is available across ChatGPT web, mobile, and desktop apps. This unification means Plus and Pro users always use one model - GPT-5.6 Sol - with tunable depth rather than switching between models based on task type. Source: 9to5Mac August 6, 2026, ExplainX August 2026

Is ChatGPT free in 2026?
ChatGPT's free tier now includes unlimited text chats on GPT-5.6 Luna as of August 7, 2026 - the first time ChatGPT has offered truly unlimited text conversations at no cost. Free users also have access to a Think button for harder questions that applies deeper reasoning. Limits still apply for file uploads, image generation, Codex access, and other tools. ChatGPT Plus at $20/month provides GPT-5.6 Sol with the reasoning slider, higher limits across all features, and access to ChatGPT Work and Codex. For business teams and enterprise deployments, ChatGPT Business starts at $40/user/month with a new Premium seat option at $100/user/month annual providing 5x usage and no five-hour limit. Source: 9to5Mac August 6, 2026, OpenAI Release Notes

What is ChatGPT Work?
ChatGPT Work is a standalone agentic 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 project context into finished work. Powered by GPT-5.6, it can break a goal into steps and work for hours while users follow progress, answer questions, change direction, and approve important actions. ChatGPT Work and ChatGPT Codex share plugins and both live inside the ChatGPT desktop app on macOS and Windows. ChatGPT Work got website sign-in support on August 25, 2026, allowing it to assist with tasks inside authenticated web applications through a cloud browser without exposing credentials in chat. WebMCP (Site tools) launched in August 2026, letting ChatGPT Work interact with websites as a user would through shared live pages and signed-in sessions. Source: OpenAI What's New, Releasebot August 2026

What is ChatGPT Dreaming memory?
Dreaming V3 is ChatGPT's rebuilt memory architecture launched June 4, 2026. Unlike the old saved-memories system requiring explicit "remember this" instructions, Dreaming V3 runs a background process that automatically synthesizes memory from past conversations and updates it as circumstances change. A memory reading "going to Singapore in July" updates automatically to "went to Singapore in July 2026" after the trip ends. Factual recall improved from 41.5% with 2024 saved memories to 82.8% with Dreaming V3, preference adherence from 31.4% to 71.3%, and time-sensitive context updates from 9.4% to 75.1% per OpenAI's internal evaluations. Available to Plus and Pro users in the US - not yet available in EU, UK, or Switzerland due to regulatory requirements. Users can view, edit, or delete memories in Settings > Memory. Source: OpenAI Release Notes

What ChatGPT models were retired in 2026?
The complete 2026 ChatGPT model retirement list: GPT-4o and GPT-4.1 retired February 13, 2026. GPT-5.1 family retired March 11, 2026. GPT-5.2 family retired June 12, 2026. GPT-4.5 retired June 26, 2026. GPT-5.4 retired from ChatGPT July 23, 2026. OpenAI o3 retired August 26, 2026. Official DALL-E GPT retired August 30, 2026. GPT-5.4 and GPT-5.4 mini retired from Codex August 31, 2026. OpenAI is running approximately 90-day model lifecycles in 2026 - models generally remain available in ChatGPT for 90 days after a successor is released before retiring. These retirements apply to ChatGPT only - the API model availability follows a separate schedule. GPT-5.5 retirement has not been announced as of August 31, 2026. Source: OpenAI Model Release Notes, OpenAI Release Notes

What is WebMCP and how does it work in ChatGPT?
WebMCP - also called Site tools - lets ChatGPT Work and Codex use tools provided by websites in the ChatGPT desktop app's built-in browser. A website can expose actions directly to the AI agent through a shared live page and signed-in session, enabling the agent to interact with web applications as a user would rather than just reading content. For example, a document editor can provide tools to find a section or add a comment. WebMCP requires GPT-5.6 Sol or Terra and the latest desktop app - it is not available with GPT-5.6 Luna or in Enterprise or Edu workspaces. Cloud browser sign-in, launched August 25, 2026, complements WebMCP by allowing ChatGPT Work to sign into authenticated websites through a cloud browser session without exposing credentials in chat. Source: OpenAI What's New, Releasebot August 2026

ChatGPT on August 31, 2026 is a platform that has changed more in eight months than it did in its first two years. The model that launched in November 2022 was a chat interface. The platform in August 2026 is a tiered work environment with three distinct experiences - Classic, Work, and Codex - a rebuilt memory system, a trigger-based task automation layer, website interaction through WebMCP, and a reasoning effort slider that replaced the entire Instant/Thinking model architecture.

The most consequential August change for business teams is not the GPT-5.6 Sol update or the DALL-E GPT retirement. It is the trigger-based Tasks capability. Scheduled Tasks that fire on Slack messages, Gmail arrivals, and GitHub events move ChatGPT from a tool you query to an assistant that monitors and responds. For teams that have been waiting for ChatGPT to behave more like an autonomous agent and less like a search box, August 2026 is when that capability arrived.

The practical guidance remains constant: OpenAI is running 90-day model lifecycles. Any named model you are testing today will retire in approximately 90 days. Any workflow evaluation you complete on GPT-5.6 Sol this month is valid for roughly a quarter before GPT-5.7 or equivalent changes the baseline. Test quarterly. Update configurations when retirements are announced. And if you have not re-evaluated ChatGPT's reliability since May 2026's 52.5% hallucination reduction - let alone August's 68% error reduction on Sol - your mental model of what ChatGPT gets right and wrong is already outdated.