[AI Tool Updates] OpenAI Expands ChatGPT Instructions Limit (7.15)

OpenAI gave paid ChatGPT users more room for persistent instructions, while Google Cloud tightened Gemini Enterprise controls for Jira Data Center connectors.…

Table of contents

1. Overview
2. Details
3. Morning Breaking Updates
4. At a glance
5. FAQ
6. Sources

OpenAI Expands ChatGPT Instructions Limit (7.15)

Overview

• OpenAI raised the ChatGPT custom instructions limit from 1,500 to 5,000 characters for Plus, Pro, Enterprise, Business, and Education users.
• Google Cloud added Public Preview action filtering for Gemini Enterprise Jira Data Center federated data stores.
• OpenAI published GPT-Red, an automated red teaming system that uses self-play to test model robustness and prompt-injection resistance.
• Thinking Machines introduced Inkling through Hugging Face as part of an open source and open science push.

Details

OpenAI Raises ChatGPT Custom Instructions Limit to 5,000 Characters

OpenAI said it increased the ChatGPT custom instructions character limit for Plus, Pro, Enterprise, Business, and Education users from 1,500 to 5,000 characters. The change affects users who rely on persistent guidance to define tone, formatting rules, role assumptions, and preferred behavior across sessions.

The practical change is simple: paid and education users can now describe more context without compressing their working rules into a short prompt. That matters for teams that keep reusable instructions for code review style, writing standards, tutoring preferences, or enterprise response policies.

The update does not change the model itself, and OpenAI did not describe a new API endpoint or pricing change in the provided release-note evidence. It is a workflow change for ChatGPT users rather than a developer-platform migration.

▸ ChatGPT instructions deep dive

The old 1,500-character ceiling forced many users to choose between style guidance and task-specific constraints. A 5,000-character limit gives more space for durable preferences, especially when a user needs the assistant to follow house rules across many chats. For example, a developer can define review priorities, preferred test commands, naming conventions, and response format in one place instead of repeating those details in every thread.

The change also makes custom instructions more useful for organizations. Enterprise, Business, and Education users often need a shared working style that is more specific than “be concise” or “use professional tone.” Longer instructions can capture internal terminology, accessibility preferences, citation habits, and escalation rules. That does not replace governance controls, but it can reduce repeated prompt setup for routine work.

There is a tradeoff. Longer persistent instructions can become stale or overly broad if users treat them as a dumping ground. The best use is a compact operating profile, not a second system prompt filled with every possible exception. Teams should review these instructions when workflows change, because outdated standing guidance can quietly shape answers in later sessions.

For individual users, the update is most useful when instructions express stable preferences. Temporary project facts still belong in the chat, files, or a connected workspace. Persistent instructions should carry durable behavior: how to structure answers, what level of detail to use, which assumptions to avoid, and what constraints apply across most work.

Key takeaway: OpenAI turned custom instructions into a more practical workspace setting for paid ChatGPT users, with the limit rising to 5,000 characters but no pricing or API migration attached.

Google Cloud Adds Jira Data Center Action Filtering to Gemini Enterprise

Google Cloud said Gemini Enterprise added Public Preview action-filtering support for Jira Data Center federated data stores. The feature applies configured filters to both search queries and action execution, not just to retrieval.

That distinction matters for enterprise users. Search filtering controls what information a user can find, while action filtering affects what the system can do after it finds relevant data. In Jira Data Center environments, those two layers often need to match closely because tickets can contain customer data, security work, or internal planning details.

Google Cloud marked the feature as Public Preview, so teams should treat it as available for testing rather than as a mature default. The provided evidence does not describe a breaking change, deprecation, or price adjustment.

▸ Gemini Enterprise Jira deep dive

Federated data stores let Gemini Enterprise work across systems without forcing every source into a single new repository. That architecture is useful, but it raises permission questions. If an assistant can search Jira and also trigger actions, administrators need confidence that the same policy boundaries apply across both phases.

Action filtering addresses that gap. A search-only control can stop a user from seeing restricted results, but action execution introduces another risk. An assistant might draft, update, or trigger workflow steps based on data that should remain outside a user’s scope. By applying configured filters to both search and action execution, Google Cloud is moving the connector closer to ordinary enterprise access-control expectations.

The Public Preview label is important. It signals that teams can evaluate the behavior, but should still test edge cases before using it in critical workflows. Jira Data Center deployments often contain customized projects, fields, permission schemes, and automation rules. Those local differences can affect whether filtering behaves as expected.

For administrators, the first test should be simple: compare Gemini Enterprise behavior with Jira’s own permission model for the same user. If a user cannot view or act on an issue in Jira, the assistant should not create a workaround through search or action execution. That validation matters more than the feature announcement itself.

Key takeaway: Google Cloud’s Jira Data Center update is about enterprise control, not interface polish: Gemini Enterprise now has preview filtering that covers both discovery and action paths.

OpenAI Publishes GPT-Red for Automated Robustness Testing

OpenAI published GPT-Red, describing it as an automated red teaming system that uses self-play to improve AI safety, alignment, and prompt-injection robustness. The announcement sits closer to research and safety tooling than to a consumer product release.

The core idea is that models can be tested through adversarial interaction rather than only through static evaluation sets. OpenAI’s framing places GPT-Red in the safety workflow: generate attacks, examine model behavior, and use the results to improve robustness.

The provided source evidence does not say GPT-Red is a generally available product, paid feature, or API endpoint. Readers should treat it as an OpenAI research and safety update unless OpenAI later attaches it to a developer-facing release.

▸ GPT-Red deep dive

Red teaming has become a standard part of AI deployment because model failures often appear in interaction, not in isolated benchmark questions. Prompt injection is a clear example. The risk is not only that a model answers incorrectly, but that it follows hostile instructions embedded in documents, web pages, tickets, or tool outputs.

GPT-Red’s self-play framing suggests a scalable version of that testing loop. Instead of relying only on human testers to imagine attacks, an automated system can generate and refine adversarial prompts. That can widen coverage and expose failure modes earlier in development. It does not remove the need for human review, but it can make review more systematic.

The practical implication for builders is indirect but relevant. If OpenAI improves robustness testing upstream, downstream tools may become less brittle in agentic workflows. Systems that connect models to files, browsers, issue trackers, or code execution need stronger defenses against malicious context. Automated red teaming is one route toward finding those weaknesses before deployment.

There is also a reporting distinction. This is not the same kind of update as a new ChatGPT setting or a Gemini Enterprise connector control. GPT-Red is evidence of safety infrastructure development. Its value depends on whether the findings influence shipped models, public evaluations, or developer guidance in later releases.

Key takeaway: GPT-Red shows OpenAI investing in automated adversarial testing, with the clearest near-term relevance for prompt-injection defenses and agent safety rather than end-user settings.

Thinking Machines Introduces Inkling Through Hugging Face

Thinking Machines introduced Inkling in a Hugging Face post that framed the work around open source and open science. The provided evidence does not include version numbers, pricing, model limits, or an API contract.

For tool users, that means the update should be read as an ecosystem signal rather than as a ready operational migration. Hugging Face is the distribution and community context here, and the announcement’s available evidence emphasizes openness more than deployment mechanics.

The item still belongs on the watch list because open source AI releases can affect developer workflows quickly once model cards, code, weights, or demos become usable. But the immediate takeaway is cautious: the supplied evidence supports the existence and positioning of Inkling, not detailed claims about performance or cost.

▸ Inkling deep dive

Open source AI announcements can be difficult to evaluate from launch language alone. The useful questions are concrete: what artifact is available, under what license, with what hardware requirements, and how it compares against existing tools. The supplied source evidence does not answer those questions, so the responsible reading is narrow.

The Hugging Face context matters because it often serves as the place where developers first inspect runnable examples, model files, datasets, or community discussion. A project introduced there can move from announcement to experimentation faster than a closed product page. Still, experimentation depends on what Thinking Machines actually makes available.

For teams tracking AI tools, the next step is not adoption. It is classification. If Inkling becomes a model, library, benchmark, or workflow tool, the impact will differ. A model release raises questions about inference cost and quality. A library raises integration and maintenance questions. A research artifact raises reproducibility questions.

The safest conclusion is that Inkling is an early signal from Thinking Machines with an open-science posture. It should not displace existing production tools based on the provided evidence alone. It is worth monitoring for concrete artifacts, licensing terms, and practical examples.

Key takeaway: Inkling is best treated as an early open-science announcement until Thinking Machines provides concrete technical artifacts, license terms, or deployment details.

Morning Breaking Updates

• huggingface.co: Security incident disclosure — July 2026 - We’re on a journey to advance and democratize artificial intelligence through open source and open science.

At a glance

Fact	Publisher	Source
ChatGPT custom instructions rose from 1,500 to 5,000 characters for paid and education tiers.	OpenAI	help.openai.com
Gemini Enterprise added Public Preview action filtering for Jira Data Center federated data stores.	Google Cloud	docs.cloud.google.com
OpenAI described a state-and-federal AI safety model built around “reverse federalism.”	openai.com	openai.com
GPT-Red uses automated red teaming and self-play to test robustness and prompt-injection defenses.	openai.com	openai.com
Thinking Machines introduced Inkling in a Hugging Face post focused on open source and open science.	huggingface.co	huggingface.co

FAQ

Q1. What changed for ChatGPT users?

A. OpenAI raised the custom instructions limit from 1,500 to 5,000 characters for Plus, Pro, Enterprise, Business, and Education users, giving paid and school accounts more room for persistent response rules.

Q2. How should teams use the larger custom instructions field?

A. Use it for stable preferences, such as tone, formatting, review standards, and recurring constraints. OpenAI’s 5,000-character limit helps, but temporary project facts should still stay in the active chat or workspace context.

Q3. Does Google Cloud’s Jira update affect production deployments?

A. Google Cloud labels the Gemini Enterprise Jira Data Center action-filtering feature as Public Preview, so administrators should test permissions and action behavior before relying on it in sensitive Jira workflows.

Q4. How do the OpenAI and Google Cloud updates differ?

A. OpenAI’s ChatGPT change improves user-level instruction space, while Google Cloud’s Gemini Enterprise change focuses on enterprise access control. One affects prompting convenience; the other affects Jira search and action boundaries.

Q5. What should readers watch after these updates?

A. Watch whether OpenAI exposes GPT-Red findings in product safety guidance, whether Google Cloud moves Jira action filtering beyond Public Preview, and whether Thinking Machines publishes concrete Inkling artifacts on Hugging Face.

Sources

1. Increased custom instructions limit - OpenAI
2. Action-filtering support for Jira Data Center data stores in Public Preview - Google Cloud
3. The US is advancing AI safety through state and federal action - openai.com
4. GPT-Red: Unlocking Self-Improvement for Robustness - openai.com
5. Welcome Inkling by Thinking Machines - huggingface.co
6. Celebrating 25 years of visual search innovation - blog.google
7. Security incident disclosure — July 2026 - huggingface.co

Last updated: 2026-07-16T11:52:45.378Z