2026 AI Assistant Roadmap: What Release and Deprecation Pages Show
A practical comparison plan based on documented 2026 releases, notice periods, and retirements across ChatGPT, Claude, and Gemini.
Sources checked 2 Oct 2026
The vendors’ release and deprecation pages settle a narrower question: which changes are documented, which models or features are listed as available, and what notice rules or retirement dates apply. They do not provide a shared forecast for ChatGPT, Claude, and Gemini, and Chat Picker has not tested the assistants for upgrade stability, migration risk, or output quality.
Know these limits before you start
As read on October 1, 2026, OpenAI’s API deprecations page gives generally available models at least six months’ notice and specialized variants at least three months. Preview models may receive much less, such as two weeks, while safety or compliance concerns can require earlier retirement. Anthropic’s model deprecations page promises active customers at least 60 days’ notice for publicly released models, but says Amazon Bedrock and Google Cloud set separate schedules. Google’s Gemini deprecations page says its listed shutdown dates are the earliest possible; exact dates are communicated later.
Record both the product surface and the date. An app entry does not establish API availability. OpenAI’s list schedules API removal of gpt-5.4-cyber on October 1, 2026. It also schedules removal of whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize on February 26, 2027; legacy audio, realtime, and transcription families on January 20, 2027; and older GPT Image models on December 1, 2026. Anthropic says Claude Opus 4.1 was retired on August 5, 2026 and that developers using Claude Sonnet 4.5 received notice of an upcoming API retirement on September 30, 2026. That entry did not state the final retirement date. Google says its 2.5 models are not deprecated and will continue through the API until further notice, although access is limited to users who had actively used them.
Release pages add capabilities, not universal availability. ChatGPT’s release notes list Pro 200 at $200 per month, Pro 500 at $500, new Box, Dropbox, and SharePoint connections for ChatGPT Library, and a Privacy Center rollout. Claude’s release notes place Design, Slides, and Docs on every plan, add a Microsoft 365 connector, and describe a beta 1M-token context window for Sonnet 4.6. Enterprise versions of the artifacts are beta and off by default until enabled. Gemini’s app updates list Skills for users aged 18 and over, a Windows app connected to Gmail and Drive, chat-history import, and Temporary Chat.
As read on October 1, 2026, OpenAI’s Plus details list $20 monthly, Claude’s pricing page lists Pro at $20 monthly or $17 per month with annual billing, and Google’s US plans page lists Google AI Pro at $19.99 monthly. ChatGPT’s pricing page and Claude’s pricing page describe usage without publishing exact message counts. Google says compute limits refresh every five hours up to a weekly limit, with paid tiers listed at 2x, 4x, and up to 20x free usage.
For chat training, ChatGPT offers opt-out on Free, Go, Plus, and Pro. Claude offers opt-out on Free, Pro, and Max and says Team is not trained on by default. Google’s plans page did not state a setting, so it remains unverified.
Accuracy conditions also belong in the test. OpenAI warns that ChatGPT can sound confident while wrong and should be checked against reliable sources. Anthropic advises against using Claude as the only source of truth and recommends examining original sources. Google describes a Sources panel when sources are available; displaying that panel is not the same as verifying its contents.
The test plan
-
Phase 1: Text baseline. Input: a plain-text brief for a fictional product launch containing names, dates, and constraints. Prompt: “Read the attached launch brief. Produce a handoff list, preserve every date and owner exactly, and flag missing fields.” Observe added facts or missed constraints. Log the plan, product surface, completion status, and corrections.
-
Phase 2: Long-context retrieval. Input: a multipage PDF containing a fictional plan and a conflicting correction near the end. Prompt: “Read the attached plan. Return the final approved budget, launch date, owner, and unresolved risks. Cite the page for each item and report conflicts instead of resolving them silently.” Observe whether it finds the correction and cites it. Log the cited pages and every mismatch.
-
Phase 3: Image combination. Input: a small set of PNG screenshots showing the same table at different zoom levels. Prompt: “Combine these screenshots into one table. Preserve labels and values exactly, and mark unreadable cells.” Observe dropped rows, altered values, and unsupported guesses. Log the table cells that require correction.
-
Phase 4: Voice transcription. Input: a short audio clip in which two speakers state different dates, but only where voice input is exposed on the account. Prompt: “Transcribe the clip, then list each speaker’s stated date. Do not reconcile the differences.” Observe speaker separation, exact dates, and access failures. Log transcription errors separately from unavailable features.
-
Phase 5: Release and deprecation research. Input: pasted excerpts from the three deprecation pages. Prompt: “Compare the API retirement entries for gpt-5.4-cyber, Claude Sonnet 4.5, and Gemini 2.5. State the status only when the text supports it, quote any date, distinguish API from app availability, and write ‘not stated’ when no date is supplied.” Observe unsupported status labels and app-to-API confusion. Log each claim beside its supporting passage.
-
Phase 6: Migration planning. Input: a saved prompt, a permitted test document, and a mocked external response. Use synthetic files rather than live personal or confidential material. Prompt: “Write a migration checklist for this workflow. Separate confirmed retirement dates, notice windows, required replacements, and unresolved items; do not invent a replacement.” Observe unsupported replacements and unclear handoff steps. Log the plan, data setting, official notice, and manual work required.
How to read your results
Set your required features and failure conditions before running the prompts. For each row, mark the outcome as pass, fail, blocked, or needs verification. Verify every date, availability claim, and citation against the linked vendor page.
Separate output errors from plan, account, region, permission, and rollout restrictions. A blocked feature is not an answer-quality failure, and an unsupported answer is not fixed by confidence or style. Do not average impressions into a score.
If several options remain, compare the documented plan price, context allowance, usage description, data controls, and migration work required for your actual workflow. Choose only among options that meet your non-negotiable requirements.
Where the plans differ
Use the Claude vs ChatGPT vs Gemini comparison matrix and check these rows:
- 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
- Ads
Each figure includes its vendor source and read date. Confirm it on the vendor page before subscribing. A blank or “Not verified” cell means Chat Picker could not confirm the figure; it does not mean the feature or setting is absent.