Revision note — verify before publication. OpenAI announced ChatGPT Images 2.5 on September 8, 2026. This revision updates the image-generation section of the ChatGPT review with official availability, workflow changes, API model names, and safety disclosures. It is not a hands-on benchmark.
What changed in ChatGPT Images 2.5
OpenAI describes Images 2.5 as a new image model focused on sharper detail, more natural lighting, richer textures, more reliable instruction following, and better preservation of subjects from reference photos. OpenAI also says generation latency is reduced by up to 50% compared with Images 2.0. In ChatGPT, the release adds Sketch for drawing a visual reference directly in the product, templates for common formats such as flyers and product photos, comments for more focused edits, and the ability to share prompts used to create an image.
The practical editorial takeaway is narrower than the launch language: Images 2.5 should be most useful when a user expects several controlled edits rather than one disposable image. The strongest claimed improvement is consistency across multi-turn editing, but readers should still inspect text, faces, logos, hands, and other high-consequence details before publishing.
API models and who should use them
OpenAI is also introducing GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst in the API. Flare is positioned as the default choice for most applications, with the same quality, editing, and speed improvements and lower latency than GPT-Image-2. Sunburst is positioned for premium creative workflows where tighter control across edits matters more than generation time. The right choice depends on the product’s latency budget, review process, image volume, and tolerance for re-renders—not on the model name alone.
Availability, safety, and limits to verify
OpenAI says ChatGPT Images 2.5 is rolling out to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web. The Help Center says image generation is available on all ChatGPT tiers, while Images with thinking is available on Plus, Pro, and Business and is coming to Enterprise and Edu. The API models have separate pricing and access conditions, so teams should verify the current pricing page before estimating costs.
OpenAI says the release continues prompt and image safety checks, C2PA metadata, and invisible watermarking. These safeguards do not remove the need for human review, rights checks for reference material, or a policy review for sensitive imagery. We have not independently measured the claimed latency improvement or visual quality.
Editorial judgment
Choose ChatGPT Images 2.5 when you value iterative editing, reference-image fidelity, and a simple conversational workflow. It is less suitable when you need a fully controlled production pipeline, guaranteed typography, deterministic output, or a verified cost-per-image without running your own representative tests. Treat the launch claims as vendor-reported capabilities and re-check availability, pricing, usage rights, and safety documentation before making a production decision.
Primary sources
- OpenAI: Introducing ChatGPT Images 2.5
- OpenAI Help Center: Images in ChatGPT
- ChatGPT Images 2.5 safety evaluations
Limitations and who should not choose it
ChatGPT Images 2.5 is not a fit for teams that need deterministic layouts, guaranteed typography, fully reproducible outputs, or a verified cost per image without running their own representative tests. It also does not remove the need to check rights for reference images, brand assets, and likenesses. For the broader product context, see the ChatGPT review overview before deciding whether its conversational workflow matches your process.
September 2026 update: new Work and API capabilities
OpenAI published several product changes on September 10, 2026. GPT-Live-1 is now generally available in the API for full-duplex voice conversations, while the Agents API is in public beta for running long-lived cloud agents with the Codex harness. OpenAI also introduced a Data agent in ChatGPT Work and a tailored ChatGPT for Financial Services experience. These are different products: the voice and agent releases target developers, while the Data agent and Financial Services experience target governed business workflows.
GPT-Live-1 can listen and speak at the same time, handle interruptions, and delegate deeper reasoning or tool calls to a backend model. OpenAI lists a front-end voice price of $0.05 per minute, with backend model and tool usage billed separately. The Agents API lets developers choose an OpenAI-managed sandbox, their own infrastructure, or a sandbox partner. It is an orchestration and execution layer, not a promise that an agent will be reliable without evaluation.
Who should care
Voice-product teams should evaluate GPT-Live-1 when interruption handling, telephony, and natural turn-taking matter. Platform teams may consider the Agents API when they need long-running sessions, tool use, and subagent coordination without maintaining the entire harness themselves. ChatGPT Work customers with approved business data may benefit from the Data agent’s dashboards and source-grounded analysis. Financial institutions should treat the Financial Services experience as an eligibility-gated enterprise product, not a general ChatGPT plan feature.
Editorial limits and decision rule
Do not budget from headline capabilities alone. Verify current API pricing, model availability, data retention, region, enterprise eligibility, connector permissions, sandbox boundaries, and audit requirements. OpenAI’s customer examples and evaluation claims are provider-reported; AI Tools Manual has not independently benchmarked latency, voice quality, agent reliability, or financial-analysis accuracy.
- Start with a narrow, reversible workflow and synthetic or approved data.
- Log model, tool, prompt, permissions, latency, cost, and failure behavior.
- Require human approval before external messages, financial decisions, data writes, or code deployment.
- Compare the managed harness with your own infrastructure before committing to a long-term architecture.
Primary sources
- OpenAI: GPT-Live-1 in the API
- OpenAI: Introducing the Agents API
- OpenAI: Data agent in ChatGPT Work
- OpenAI: ChatGPT for Financial Services
September 2026 update: GPT-6 Astra changes the ChatGPT work baseline
OpenAI’s September release cycle adds GPT-6 Astra to the ChatGPT and API story. The model is rolling out first to a limited set of organisations and then to ChatGPT Plus, Pro, Business and Enterprise users, with API availability under gpt-6-astra. OpenAI also lists Microsoft Azure and Amazon Bedrock availability. This is a provider announcement, not an independent benchmark by AI Tools Manual.
The practical change is broader than a new model selector. OpenAI positions Astra for computer use, browsing, software engineering, science and professional work, including workflows that create documents, spreadsheets and presentations. The business page describes operation across websites, desktop apps and internal tools, even when a dedicated API is not available. That makes it relevant to teams evaluating an agentic workflow rather than a chat-only assistant.
Availability and cost are decision points. OpenAI says Astra usage is included in existing subscription allowances, with additional credits available, while the API standard price is $10 per million input tokens and $50 per million output tokens; fast mode is listed at twice the standard price. Enterprise administrators must enable Astra for their workspace, and access is off by default at launch. Confirm the current account, region, model tier, credit policy and reseller terms before budgeting.
Who should switch, and who should wait?
- Consider a pilot: engineering, research, operations and design teams that already have a supervised browser or coding workflow and need stronger multi-step execution.
- Wait or use a smaller model: teams with simple drafting needs, strict fixed-cost budgets, or no safe sandbox for computer actions. Astra’s capability does not remove the need for approvals, access controls or rollback.
- Security-sensitive users: OpenAI classifies Astra as a major increase in cyber capability and describes stronger protections. Treat those statements as risk disclosure, not a guarantee; keep production safeguards, least privilege and human review in place.
Our editorial rule is to start with synthetic or approved data, log prompts, tools, permissions, latency, cost and failures, and require confirmation before sending messages, changing records or deploying code. Compare Astra with the model already in your ChatGPT workflow and use the site’s comparison guides before moving a high-impact task. The right question is not whether Astra is “best,” but whether its additional capability reduces total work without increasing unacceptable operational risk.
Evidence note: capability, availability, pricing and safety statements above come from OpenAI’s primary announcement and safety overview. OpenAI’s benchmark numbers and customer quotations are reported claims; we do not present them as independently reproduced results.
Primary sources
- OpenAI: GPT-6 Astra — a new generation of intelligence
- OpenAI: GPT-6 Astra for business
- OpenAI: Safety overview for GPT-6 Astra
What the recent announcements change
The current direction is broader than a single model release. OpenAI is positioning ChatGPT Work as a place where teams can connect approved company data, ask questions, and receive dashboards or actions. The Agents API points developers toward longer-running workflows, while GPT‑Live‑1 is aimed at more natural voice experiences. The advertising announcement is a product and ecosystem signal, not proof that ads are enabled for every ChatGPT user.
Who should pay attention
Knowledge-work teams should verify whether the Data agent is available in their workspace, which connectors and permissions it requires, and whether outputs can be shared outside the organization. Developers should separate API model availability from ChatGPT plan access: the same name can have different limits, billing, retention, and safety controls across surfaces. Marketers should treat the advertising announcement as an area to monitor rather than an immediate workflow change.
Our assessment and checks
The practical value is a move from chat-only answers toward connected, auditable work. The main risk is assuming that a launch post equals general availability. Before adopting a feature, record the account plan, region, rollout status, data sources, human approval points, rate limits, and export behavior. Keep a small representative task set and compare accuracy, traceability, latency, and cost over time. Do not publish sensitive analysis without a human check.
Primary source and review note
Primary source: OpenAI News. Re-check plan availability, pricing, privacy, and safety documentation before publication because these details can change independently of the announcement page.
September 2026 official update: youth safety and Astra for Law
OpenAI published an Australian Youth Safety Blueprint on September 18. The company describes six pillars that include AI literacy, age-appropriate safeguards, privacy-protective age assurance, links to real-world crisis support, and parental controls. OpenAI also says it began rolling out a ChatGPT for Teens experience in Australia for users identified as 13–17, with safeguards designed for that age group. This is a policy and safety development, not evidence that every account or country receives the same controls.
On September 17, OpenAI introduced Astra for Law, a legal-focused configuration combining GPT-6 Astra with a legal search index, legal-analysis instructions, and specialist integrations. The announcement says selected firms can access it through Trusted Access in ChatGPT and Codex, while an API offering is planned. The announcement also describes 26 ecosystem plugins and a zero-data-retention option for eligible API customers. Treat those as announced availability and controls, not a guarantee of access for ordinary ChatGPT plans.
Who should care
Parents, educators, and teams serving younger users should watch the country-specific rollout and read the applicable policy before changing account settings. Legal teams should treat Astra for Law as an early-access professional workflow: useful for source-grounded research and document review, but still requiring qualified legal review and confidentiality checks.
Editorial decision
Our view is that these announcements make safety configuration and domain-specific retrieval more important selection criteria than a headline model name. We would shortlist ChatGPT for organizations that can accept regional availability and governance review; we would not recommend it as a substitute for a lawyer, crisis service, or independent child-safety program. Read the practical guides before comparing it with other research tools, and revisit the comparison hub when access or pricing changes.
Primary sources: Australian Youth Safety Blueprint and Astra for Law.
September 2026 update: ChatGPT Images 2.5
OpenAI introduced ChatGPT Images 2.5 as a new image-generation and editing model for ChatGPT, ChatGPT Work, Codex, and the Images API. OpenAI says the model improves detail, reference-image fidelity, precision editing, and consistency across multiple turns, while reducing generation latency by up to 50% compared with Images 2.0. These are vendor-reported capabilities; AI Tools Manual did not reproduce the speed or quality claims.
The product update also adds Sketch, which lets users draw a rough visual reference inside ChatGPT, plus templates for common formats such as posters and product images. Shared images can include the prompt used to make them. In the API, GPT-Image-2.5 Flare is positioned as the faster default option, while GPT-Image-2.5 Sunburst targets more controlled, premium editing workflows.
Who should use it
- Creators who need several revisions from the same reference image.
- Marketing and product teams building repeatable visual variants.
- Developers choosing between lower-latency Flare and more controlled Sunburst API workflows.
Limits and checks
- Check current plan and API pricing, rate limits, and regional availability.
- Review C2PA metadata and invisible watermark behavior when provenance matters.
- Do not treat sharper output as factual verification of people, places, labels, or charts.
- Confirm rights and consent for reference photos before production use.
Read the official Images 2.5 announcement, then revisit our ChatGPT review and ChatGPT vs Gemini. No independent benchmark, screenshot, or first-hand test is claimed here.
September 2026 update: Workspace Agents and data-driven work
OpenAI’s current developer documentation now describes Workspace Agents as a way for ChatGPT Business, Enterprise and Edu workspaces to save repeatable workflows. An agent can combine instructions, connected apps, skills, approvals and an output destination, then run on demand or on a schedule. This is a different decision from choosing a chat model: the important questions are what the agent can access, what it can write, and when a person must approve an action.
The official cookbook example uses a sales meeting-prep agent that reads a calendar, gathers account context, searches the web and saves a brief to a connected document before sending a summary. The example is a documented workflow, not a promise that every workspace has the same apps or permissions. OpenAI’s Workspace Agents API can also trigger a published agent asynchronously from an external system, using a Workspace Agent access token rather than a standard platform API key.
What this changes for buyers
ChatGPT becomes more useful when a task is scattered across business tools and needs a finished artifact, not just an answer. Skills can standardise an output format, memory can preserve context between runs, and schedules can make a recurring process visible to a team. The trade-off is governance: a connector that can read a calendar or write a document may expose more information than a normal conversation, and a scheduled run can repeat an incorrect instruction until someone notices.
Best fit and not fit
Best fit: small and mid-sized teams with a clearly owned workflow, low-risk destinations and a reviewer who can inspect the first runs. Meeting briefs, recurring research digests and internal status packs are good pilot candidates because the output can be checked before it is shared.
Not fit: high-impact decisions, unrestricted production actions, or sensitive data without an approved workspace policy. Do not paste a platform API key into a Workspace Agent trigger; the developer documentation specifies a separate scoped access token. Keep writes approval-gated until audit, rollback and attribution are clear.
Our editorial assessment
Our view is that Workspace Agents are a meaningful extension of ChatGPT for work, but they should be evaluated as governed automations rather than a faster chat mode. Start with read-only connectors, one scheduled run and a named owner. Record the source event, prompt, connected apps, approval decision and final artifact. If you need a managed cloud runtime for a custom application, compare the Agents API documentation with the workspace product instead of assuming their sessions, tokens and data controls are identical.
Read our ChatGPT profile, compare ChatGPT vs Claude, and use our AI privacy checklist before connecting business data. AI Tools Manual has not independently tested Workspace Agents or the trigger endpoint.
Primary sources and evidence limits
Product facts are based on OpenAI’s Workspace Agents cookbook, the Workspace Agents overview, and the trigger-runs documentation. These sources document supported concepts and authentication boundaries; they do not prove reliability, plan availability in every region, or safe outcomes for your connected apps. Verify access, retention, permissions, token scope and approval settings before publication.
September 22, 2026 update: GPT-6 Sol and GPT-6 Luna
OpenAI introduced GPT-6 Sol and GPT-6 Luna for ChatGPT Work and Codex, with the models also available in the API as gpt-6-sol and gpt-6-luna. OpenAI positions Sol for complex coding and agentic work and Luna for focused, high-volume tasks. The company lists 50% lower API prices than the GPT-5.6 promotional prices: $2 input / $10 output per million tokens for Sol and $0.10 input / $0.50 output for Luna. These are official price claims and should be rechecked on the current pricing page before a purchase decision.
Availability is staged. OpenAI says Sol and Luna are rolling out to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex; Free and Go users can access Luna in the desktop app. The models are not yet in Chat, and Enterprise administrators may need to enable them. This is a product and access update, not proof that either model is best for every task. Use the model selector and workspace settings to confirm what the account actually exposes.
What the pricing change means
For teams running repeated agent steps, lower token prices can change the cost of iteration, but the real bill still depends on context length, cache hits, tool calls, reasoning effort, rate limits and retries. OpenAI’s related prompt-caching update describes a 30-minute shared-prefix window, diagnostics for cache misses and explicit breakpoints. Preserve stable tool definitions and instructions, measure cached versus uncached input, and keep a spend limit before moving a production workload.
Best fit, not fit, and our decision rule
Sol is a plausible fit for coding and multi-step work when a team can review generated changes; Luna is a plausible fit for routine, high-volume tasks where latency and cost matter. Neither is a fit for unsupervised legal, clinical, financial or production-security decisions. Our editorial rule is to test one representative workload against the current model with the same prompt, tools and acceptance checks, then compare total cost and correction time rather than headline benchmarks. We have not independently benchmarked GPT-6 Sol or Luna.
Verify the official GPT-6 Sol and Luna announcement and prompt-caching update. Continue with our ChatGPT profile, ChatGPT vs Gemini comparison, and AI privacy checklist. The linked OpenAI pages support availability, pricing and caching facts; the fit and limitations are editorial synthesis.
September 2026: security history makes account review more practical
OpenAI’s ChatGPT release notes now describe Security history, a settings view that shows recent sign-ins, sign-outs, and changes to multi-factor authentication, passkeys, and other security settings. Events include time, location, and device details, although OpenAI notes that some details can be approximate or unavailable. This is a useful operational change for teams and individuals who need a quick way to spot suspicious account activity without reconstructing it from separate notifications.
The same release notes say Live now supports plugins on web, iOS, and Android. During a Voice conversation, users can use connected apps available to their account and follow written responses in the chat. Voice is also available in Work, where a task can continue in text after the call ends. Availability, connected-app permissions, and usage limits still depend on the user’s plan and workspace settings.
What changed for learning and Work users
ChatGPT can now create interactive flashcards from a prompt or uploaded notes. Cards can be flipped, marked as known or needing practice, shuffled, and saved in the library for later review. GPT-6 Sol and GPT-6 Luna also arrived in ChatGPT Work and Codex, but OpenAI states that they are separate from the models available in Chat and that the available models and reasoning options depend on plan and workspace configuration.
Best fit and limits
These changes fit users who already have connected apps, recurring study material, or a work account that needs an auditable security check. Security history is a useful first review, not a complete incident-response system: approximate locations, missing details, compromised sessions, and third-party app permissions still require investigation. Voice with plugins is convenient for hands-free work, but the user should confirm what an app can read or change before authorizing it. Flashcards are useful for retrieval practice, but generated answers still need a quick check against the source notes.
Our decision rule
Enable one connected app at a time, review its permission scope, and run a low-risk Voice task before using it with business data. For study workflows, ask ChatGPT to cite the uploaded notes and manually verify a sample of cards. For Work or Codex, record the model, plan, workspace policy, and reasoning setting used for a repeatable task. Compare the workflow with the ChatGPT tool profile and the AI privacy review guide before expanding access.
Evidence disclosure
The primary source is OpenAI’s official ChatGPT release notes, with the product newsroom used for cross-checking. AI Tools Manual did not independently test Voice plugins, security-history accuracy, flashcard quality, or plan availability. Verify current rollout, connected-app permissions, retention, workspace controls, and usage limits before approval.
September 29, 2026: GPT-6.1 Sol narrows the price–capability trade-off
OpenAI introduced GPT-6.1 Sol as an upgrade to GPT-6 Sol for agentic coding, computer use, and professional work. OpenAI says it approaches GPT-6 Astra on selected evaluations at roughly one-fifth of Astra’s standard token price. The API list price is $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. These are API rates, not a claim about ChatGPT subscription value. OpenAI’s reported benchmark results use selected test sets and settings; they are provider measurements, not tests conducted by AI Tools Manual. The company says the model is available to paid plan users in ChatGPT Work and Codex, but not yet in Chat itself, so readers should not assume the regular ChatGPT model picker already includes it.
Dots turn ChatGPT into a persistent agent workflow
OpenAI also announced Dots: always-on agents powered by GPT-6 Astra that use a separate cloud computer and can connect to more than 4,000 apps through plugins. The rollout begins for Pro and Business Premium users in eligible markets; Enterprise, Edu, and Healthcare users can try a beta when an administrator enables it. OpenAI says the first dot is included in the relevant plan, dot conversations do not count against ChatGPT usage limits, but work delegated to Codex or ChatGPT Work does count. Those distinctions matter for budgeting and access. A persistent agent with connected apps can take actions beyond answering a prompt, so organizations should scope permissions, review approval settings, and start with low-risk work. We would not grant broad email, finance, customer-data, or production access until the workspace has an owner, audit trail, rollback path, and explicit action boundaries.
Our decision rule and evidence limits
Treat GPT-6.1 Sol as a candidate for coding and structured work in Work/Codex when the paid-plan access and API economics fit; test representative tasks before standardizing. Treat Dots as an agent pilot, not a hands-off replacement for accountable staff. For model choice and safe code review, see our AI app-builder comparison and the generated-code review guide. Primary sources: OpenAI’s GPT-6.1 Sol announcement, Dots launch details, and GPT-6.1 Sol safety-card addendum. We have not independently benchmarked either product or audited Dots’ permissions, privacy, or safety controls.
September 29–30, 2026: GPT-6.1 Sol, dots, and a security disclosure
OpenAI announced GPT-6.1 Sol and its new dots agent on September 29, then published a security post about a model-distillation campaign on September 30. These are separate changes: GPT-6.1 Sol is a model for ChatGPT Work and Codex plus the API, while dots are persistent agents that can use connected apps and run work on a cloud computer. OpenAI’s announcements are primary sources, but the reported evaluations, availability, and security findings are provider statements—not independent tests by AI Tools Manual.
GPT-6.1 Sol: where it is available
OpenAI says GPT-6.1 Sol is available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users; it is not yet available in ChatGPT’s standard Chat experience. The API model ID is gpt-6.1-sol. OpenAI lists standard API pricing of $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. Its comparisons with GPT-6 Astra and competitor models rely on selected evaluations; task mix, reasoning settings, tools, system prompts, and production configuration can affect results. Buyers should benchmark representative work and confirm the current rate card before budgeting.
Dots: a persistent agent with connected-app access
OpenAI describes dots as agents that can work continuously, learn preferences from feedback, and connect to apps through plugins. The first dot is included for eligible Pro and Business Premium users in supported markets; Enterprise, Edu, and Healthcare users can try a beta when an administrator enables it. OpenAI says proactive research uses connected apps in read-only mode, while other actions follow app permissions, custom rules, and approval checks. Workspace administrators can separately control dots, local-computer access, Slack or Teams access, and custom rules. These are consequential controls: begin with a narrow role, least-privilege app connections, and approval requirements for external actions. Review the current workspace settings and data-use terms before connecting sensitive work data.
Security disclosure: distinguish exposed reasoning from stored conversations
OpenAI says it identified and disrupted a campaign that manipulated model interactions to expose protected reasoning. Its post explicitly says the operators did not break encryption, compromise a database, or gain direct access to stored user conversations. OpenAI reports 16,000 attempted extraction requests from more than 4,000 users during July 24–25, and related prompt patterns across a cluster of more than 15,000 users; the footnote says these are attempts, not necessarily successful extractions. OpenAI attributes a core activity cluster to individuals associated with Moonshot AI, while also saying it is unclear whether all operators were one actor. That attribution and the scope figures are OpenAI’s account, not an independent investigation by this publication. Do not interpret this as evidence that stored ChatGPT conversations were broadly breached.
Our assessment and decision rule
For an individual, GPT-6.1 Sol is the more direct reason to revisit model selection if you use Work or Codex for multi-step coding or document tasks; it does not replace the standard ChatGPT model picker yet. Dots matter more to teams that want persistent, connected workflows, but they also raise the stakes for app permissions and review boundaries. Do not enable broad access simply because the launch describes an agent as always-on. Start with read-only or narrowly scoped connections, keep approval on for consequential actions, inspect Activity View, and test recovery and offboarding. Compare this with our AI-agent rollout checklist and AI privacy review guide.
Evidence disclosure — verify before approval
Primary sources: OpenAI’s GPT-6.1 Sol announcement, dots announcement, security disclosure, and dots workspace controls. AI Tools Manual did not reproduce model evaluations, inspect a customer workspace, or independently verify the security investigation.
October 5, 2026: ChatGPT tests a new visual ad format
OpenAI announced a visual advertising format that it says will initially appear during image generation in ChatGPT for users on the Free and Go plans. The company says the ads will be clearly labeled and kept separate from the image being created, and that advertising will not influence ChatGPT’s answers. OpenAI said testing would begin later in October in the United States with an initial group of advertisers. The same announcement describes expanded advertiser measurement integrations and early brand-suitability work. These are OpenAI’s stated plans and safeguards; they should not be read as proof that a test has reached every eligible account or that independent privacy and effectiveness audits have been completed.
The change matters differently to different readers. Free and Go users may encounter a new discovery surface during image generation if they are included in the U.S. test. Advertisers gain a new format and measurement options, but should distinguish OpenAI’s product claims from partner-reported campaign results. Users elsewhere, or on other plans, should not assume immediate availability based on the announcement alone.
What users and advertisers should verify
Users should look for the ad label and distinguish paid placements from the generated image and from ChatGPT’s answer. OpenAI’s post says conversations stay private, but it does not fully document in this announcement the signals used to select placements, the retention and sharing details for campaign measurement, or every user control. Those points warrant checking against current product controls and policies before making a privacy decision. Advertisers should confirm eligibility, format specifications, measurement definitions, placement exclusions, and how conversion data is transmitted before budgeting around the test.
Our assessment is that the format is a material change to ChatGPT’s commercial surface, but the announcement is an initial test—not evidence of broad rollout or independently verified performance. We have not seen the test in an account and have not audited its targeting or data flows. Readers evaluating AI services for sensitive work should review our AI tool privacy checklist and re-check the official Ads Principles and product controls as the test develops.
Primary source: OpenAI, “Building advertising for the way people use AI,” October 5, 2026. Human review required: verify ad eligibility and rollout, conversation-data use and retention, user controls, measurement partners, and the exact distinction between OpenAI’s claims and independently established facts before this revision is applied.
GPT-6 and Intelligent UI: October 7 rollout and product impact
OpenAI’s October 7 release introduces GPT-6 in the ChatGPT Chat experience alongside Intelligent UI. OpenAI describes responses that can combine text with graphics, charts, forms, buttons, and other interactive components; it says the model may choose a visual or interactive format when that fits the question, while retaining a plain-text answer for simpler requests. The release also describes progressively streamed interfaces and answers that can begin before the model has finished thinking. These are product claims from OpenAI, not results independently tested by AI Tools Manual.
OpenAI says the rollout begins globally for Plus, Pro, Business, and Enterprise, then expands to Free and Go on October 8; Enterprise availability depends on workplace administrator settings. The Chat experience uses GPT-6 Sol on paid tiers and GPT-6 Luna on Free and Go. OpenAI explicitly says this release does not change the models used by Work or Codex. The launch is therefore a major change to the consumer-facing ChatGPT experience, not a universal replacement across every OpenAI product.
Who should reassess their ChatGPT workflow
Best fit: people who explain complex topics, compare options, plan trips, or need small interactive helpers inside a conversation. Teams should first test whether these generated interfaces improve the task they already perform; they should not assume every response will use an interface or that an interactive answer is automatically correct. Users who need fixed, auditable outputs may prefer a documented workflow or a conventional spreadsheet/form.
Our assessment is that Intelligent UI changes the interaction format more than the underlying need to verify consequential answers. Try it on a reversible planning task, inspect the inputs and calculations, and keep a text-only fallback for work that must be copied into a controlled process. OpenAI reports stronger safeguards and multi-turn jailbreak resistance; those are vendor safety claims and do not replace organization-specific testing. The release’s safety card should be reviewed before relying on those claims in a regulated or high-impact workflow.
See our ChatGPT review for the broader feature and limitation context. Evidence note: this section summarizes OpenAI’s GPT-6 release and deployment safety card; no independent benchmark or hands-on test is claimed. Next review: October 15, 2026.
OpenAI safety report: AI-enabled false-front operations
OpenAI’s October 8, 2026 safety report describes two influence operations—one attributed by OpenAI to Russia and one to Iran—that used its models alongside other tools. The report says the operators generated long-form articles, social comments, fake documents or audio scripts, and internal reports. OpenAI says it banned the accounts it identified. This is a vendor-published threat report, not an independent audit of ChatGPT’s safeguards or a change to the product’s feature set.
What the report establishes—and what it does not
OpenAI explicitly cautions that operators’ own impact claims were sometimes exaggerated or could not be verified. It says open-source research corroborated some content and publication claims, while other items may have remained private, been removed, or never been posted. That distinction matters: the report supports the conclusion that generative AI can make deceptive content workflows more efficient, but it does not establish that every claimed campaign outcome occurred or that ChatGPT independently produced the final published material.
Who is affected: readers, publishers, educators, and organizations that rely on online claims about contentious events. The decision impact is primarily about verification, not a new setting users need to enable. Check the author’s identity and publication history, look for independent corroboration, trace media to its original source, and be cautious when several accounts repeat unusually similar claims. A polished, fluent article is not evidence of genuine authorship.
Our assessment: this disclosure is useful as a threat-pattern report, but OpenAI’s attribution and impact estimates remain the company’s account and should be read with the stated caveats. It should not be presented as proof that ordinary ChatGPT use is inherently deceptive, nor as a complete measurement of abuse. For a broader assistant comparison, see our Claude review. Evidence note: facts and attribution above are summarized from OpenAI’s October 8 report; we have not independently verified the cases. Next review: October 25, 2026.
October 15, 2026: GPT-Rosalind Academic Plan is a separate research-access route
OpenAI announced the Rosalind Academic Plan for eligible U.S. university labs. It offers up to 30 Business Standard seats at $0 per user per month for up to 12 months, while access to GPT-Rosalind itself requires trusted-access approval. OpenAI describes GPT-Rosalind as a specialized life-sciences model paired with Rosalind Workbench for research workflows such as medicinal chemistry, genomics and wet-lab assistance. The plan is not a general ChatGPT consumer tier, and API access is not included at this time. Existing university Enterprise customers apply through their Enterprise access process. Details: OpenAI’s announcement.
The practical audience is a principal investigator or lab lead at an eligible U.S. academic institution—not an individual consumer looking for a new ChatGPT model. The plan could lower the cost of evaluating AI assistance in life-sciences research, but it does not replace scientific validation, institutional review or expert judgment. OpenAI says research data shared through the plan is not used to train models by default; this is a vendor policy statement, not an independent audit of data handling.
Before treating this as available, verify the lab’s institution, approval path, seat count, 12-month limit, usage limits, trusted-access requirements, API exclusion and data terms. The plan’s $0 price is time-bounded and eligibility-based; do not generalize it to paid ChatGPT plans or public GPT-Rosalind access. We have not tested the model or inspected the workbench. For general assistants, compare the published profiles for Claude and Perplexity. Next review: November 16, 2026.
Human review required: confirm eligibility, seat limits, plan duration, API exclusion, trusted-access conditions and the exact research-data language before publication.
October 2026 OpenAI platform changes: API, ads, and cloud access
Review flag: the changes below affect billing, privacy statements, advertising, and enterprise access. They are held for human review; confirm all current pricing, account eligibility, data handling, and launch status before approval.
Decisions API general availability
OpenAI announced general availability for the Decisions API on October 22. The endpoint is designed for bounded classifications, choices, and scores using text, images, or structured input; the release notes say requests can include up to 400 questions and that the API now integrates with GPT-Live. OpenAI lists GPT-6 Luna pricing at $0.10 per million uncached input tokens and $0.01 per million cached input tokens, with no output-token charge. Regional processing is available in the US and Europe, while Zero Data Retention and HIPAA use are limited to eligible customers. These price and data-control terms need direct review against the current pricing and data-control documentation.
OpenAI’s “up to 10x faster” comparison is a vendor claim, not an independent AI Tools Manual benchmark. A bounded output can reduce unnecessary generation, but it still needs a fallback and human review when a wrong classification would trigger a consequential action.
ChatGPT Ads market expansion
On October 21, OpenAI said ChatGPT Ads expanded to 26 additional countries, bringing its stated availability to 89 countries. The announcement says ads appear on Free and Go, while Plus, Pro, and Enterprise remain ad-free. OpenAI also states that conversations remain private from advertisers, customer data is not sold, and ads are labelled separately from answers. Because these are important privacy and plan claims, do not treat this summary as a substitute for the current ads principles, market list, or account-level plan terms.
OpenAI models in Google Cloud
OpenAI and Google Cloud announced a planned Model Garden integration for OpenAI models and Codex workflows, with purchases through Google Cloud Marketplace and access to Google Cloud governance controls. OpenAI said general availability was planned for the coming weeks; this announcement does not establish that the integration is already generally available. Google Cloud customers should verify region, model list, access stage, billing, data-processing terms, and contractual controls before making deployment decisions.
Editorial assessment: these are three separate decisions, not one blanket ChatGPT upgrade: API buyers should model complete request and cache costs; consumers should verify their country and plan’s ad status; enterprise teams should wait for confirmed cloud availability and review the applicable data terms. We did not independently test the endpoint, ad experience, or cloud integration. Related: our source-verification workflow and ChatGPT vs. Claude.
Primary sources and required checks
- OpenAI Decisions API announcement and official API guide: verify endpoint, token charges, caching, ZDR/HIPAA eligibility, and regional processing.
- OpenAI Ads expansion announcement: verify country list, plan coverage, and privacy language.
- OpenAI–Google Cloud announcement: verify that planned availability has actually launched and confirm commercial and data terms.
Next review: November 23, 2026, or sooner when pricing, access, privacy, or launch status changes.nn
GPT-6.1 Terra: availability and product fit
OpenAI announced GPT-6.1 Terra on October 26, 2026. The company says it is available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, and through the API under the model identifier gpt-6.1-terra. OpenAI positions it as a more efficient model for visual reasoning, computer use, and writing rather than as a universal replacement for every GPT-6 model. Its API documentation lists a 1,050,000-token context window and 128,000-token maximum output; confirm current limits in the model page before selecting it for a production workload.
The practical choice is to test Terra on bounded work where lower cost or visual/computer interaction matters, then compare the result with the model already used by the team. OpenAI reports improvements on its own factuality and coding evaluations, but those are vendor-reported results, not an AI Tools Manual benchmark. They should be treated as a reason to evaluate the model, not proof that it will be more accurate on a particular company’s data or workflow.
Cost, limitations, and a safe evaluation path
OpenAI lists standard API rates of $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens at announcement. These are time-sensitive billing facts and may change; model selection should include a current price check and a small cost test using representative prompts. ChatGPT Work/Codex access and API billing are separate routes, so do not assume a subscription includes API usage.
For an evaluation, choose three representative tasks, define a human-checked success criterion, record latency and total token usage, and compare against the current model. Keep tool permissions narrow and review generated actions before they affect customer or production systems. This model is a fit for teams that can evaluate visual reasoning or coding work and inspect outputs; it is not a reason to remove human review from high-impact decisions. Related context: Perplexity Computer review for a contrasting research-and-execution workflow.
Evidence note: This update summarizes OpenAI’s launch and model documentation. We have not independently benchmarked GPT-6.1 Terra. Next review: November 11, 2026.