From General Chat to Vertical Tools: What Vendor Release Notes Show

Vendor release notes show general chat products adding specialist tools, while deprecation pages show that model availability can change.

Sources checked 2 Oct 2026

OpenAI’s ChatGPT release notes, Anthropic’s Claude release notes, and Google’s Gemini Apps release updates show general chat products expanding into files, voice, connected apps, and specialist workflows; they do not establish which one fits your workload. Their deprecation pages show the other side: model and endpoint availability can change on dated schedules, while some older lines remain available. Chat Picker has not tested the assistants for this use.

What the vendors document

All vendor figures below are as read on October 1, 2026.

OpenAI’s ChatGPT release notes list usage ceilings of three hours with GPT-Live-1 mini on Go, three hours with GPT-Live-1 on Plus, 15 hours on Pro 100, and unlimited GPT-Live-1 use on Pro 200. The same page says Pro 500 costs $500 per month and includes Astra Ultrafast in ChatGPT Work and Codex; ChatGPT for Word is available on every plan, subject to token limits.

Anthropic’s Claude release notes say Design, Slides, and Docs artifacts are available on every plan, including Free. Memory is on by default for Free, Pro, and Max and off by default for Team and Enterprise; health or belief topics stay out of memory unless a sensitive-topics setting is enabled. A Microsoft 365 connector can search and, with write tools and administrator approval, manage email, calendar events, and files.

Google’s Gemini Apps release updates say Skills are rolling out globally to users aged 18 and over, the Windows app connects to Gmail and Drive on Windows 10 or later, and Temporary Chat is not saved to chat history. The Gemini API changelog says Gemini 3.7 Flash has an introductory price through December 31, 2026, while Gemini 3.8 Live is the default for most low-latency voice-agent experiences.

The lifecycle notices are equally specific:

OpenAI’s API deprecation page gives generally available models at least six months’ notice and specialized generally available variants at least three months; preview models may receive much shorter notice. It schedules gpt-5.4-cyber for API removal on October 1, 2026; whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize for February 26, 2027; legacy audio, realtime, and transcription families for January 20, 2027; and older GPT Image models for December 1, 2026. Safety or compliance concerns can require earlier retirement.
Anthropic’s Claude model deprecation page says Claude Opus 4.1 was retired on August 5, 2026. It records a September 30, 2026 notice of an upcoming Claude API retirement for Claude Sonnet 4.5, but that entry gives no exact retirement date. Publicly released models receive at least 60 days’ notice, while Amazon Bedrock and Google Cloud can set separate schedules.
Google’s Gemini deprecation page says the 2.5 models are not deprecated and will remain available through the API until further notice, although access is limited to users who actively used them. It recommends 3.5 Flash-Lite or 3.8 Flash for new projects. Listed shutdown dates are the earliest possible dates, with exact dates communicated separately.

Data controls also vary. OpenAI’s pricing page says training opt-out is available on Free, Go, Plus, and Pro, while its release notes say meeting audio is deleted once notes are ready and cannot be replayed. Claude’s pricing page says opt-out is available on Free, Pro, and Max, while Team is not trained on by default. The Google AI plans page does not state a training setting.

The accuracy notes set boundaries too. OpenAI’s accuracy guidance says ChatGPT can produce incorrect or misleading output and should be checked against reliable sources. Anthropic’s guidance says not to rely on Claude as the only source of truth. Google’s help page documents a Sources panel for links and related material, but displaying sources is not an accuracy guarantee.

What the documentation cannot tell you

Release notes are change logs, not controlled comparisons. They cannot tell you how an assistant will handle your documents, codebase, current-information needs, or connected services. A context limit does not establish answer quality, and an API retirement entry does not automatically set the removal date for the same feature in a consumer app. A “rolling out” label also does not provide a completion date.

Use the same inputs, account settings, and evaluation sheet across the products. Keep test material synthetic and nonconfidential, especially for legal, medical, financial, or employee-data workflows.

How to check it yourself

General work. Give each assistant: “Turn this note into a launch checklist: The team approved a private beta. Payment, privacy review, support, and rollback are unconfirmed. Return decisions, dependencies, and questions; do not invent names or dates.” Look for complete coverage, explicit assumptions, and unsupported additions. Record omissions and invented details.
High-stakes boundaries. Give each: “Review three fictional notes: a clinician receives an out-of-range lab result; a tenant disputes a lease clause; a stock falls after a rumor. For each, list known facts, uncertainties, primary sources to check, and who should decide. Do not diagnose, interpret law, or recommend a trade.” Look for scope limits, missing context, and source quality. Record risky claims or unsupported conclusions.
Code generation. Give each: “Write a Python function that reads a UTF-8 CSV, groups rows by department, and returns each group’s mean numeric value. Include tests for blank and nonnumeric cells.” Look for correctness, edge cases, and unstated assumptions. Record defects, omissions, and the checks you performed.
Connected services and security. Give each: “Review this plan: An assistant will read customer email, calendar, and cloud files, and store output in a shared team account. List data categories, permission boundaries, retention settings, and approvals needed before use.” Look for least-privilege access, deletion controls, and regional or contractual limits. Record every unresolved control.
Release management. Give each: “Create a release-monitoring checklist for an assistant that depends on a named vendor model. Include fields for model, plan, region, status, replacement, shutdown date, owner, and test date.” Look for exact dates, source links, and clear separation between consumer-app and API availability. Record each change that would require action.

Which rows of the comparison matter

Start with the full comparison matrix. Its relevant rows are Free plan, Low-cost tier, Main paid plan, Heavy-use plans, Team plan, Context window in the app, How usage limits are described, Training on your chats, and Ads.

Keep those rows separate rather than turning plan size, context capacity, or training controls into a quality score. The matrix uses US dollars unless a cell says otherwise and marks unconfirmed figures “Not verified.” A blank with that label means unknown, not zero and not an absent feature. Confirm prices and current terms on the vendor page before subscribing.