Chat Picker

2026 AI Versioning: How Vendors Announce Updates and Retirements

AI vendors use different update and retirement notice systems, so compare fixed tasks, lifecycle dates, plan limits, and data settings without relying on unsupported scores.

Sources checked 2 Oct 2026

Vendor pages settle how updates and retirements are announced, not which assistant performs best: as read on October 2026, OpenAI’s API deprecations page gives tiered retirement notice, Anthropic’s model deprecations page labels models from active to retired, and Google’s Gemini deprecations page lists earliest possible shutdown dates while reserving the exact date for later notice. Release notes are published separately. Chat Picker has not tested the assistants for updates, quality, speed, or reliability.

Know these limits before you start

As read on October 2026, OpenAI says generally available models receive at least six months’ notice, specialized generally available variants receive at least three months, and preview models may receive as little as two weeks. Safety or compliance concerns can require earlier retirement. The list read on October 1 schedules gpt-5.4-cyber removal for 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 and snapshots for January 20, 2027; and older GPT Image models for December 1, 2026.

Anthropic distinguishes active, legacy, deprecated, and retired models. A deprecated model remains functional but is no longer recommended and has a replacement and retirement date; requests to a retired model fail. Publicly released models receive at least 60 days’ notice for active deployments. Its listed dates cover the Claude API, Claude Platform on AWS, and Microsoft Foundry, while Amazon Bedrock and Google Cloud set separate schedules. Claude Opus 4.1 retired on August 5, 2026. Customers using Claude Sonnet 4.5 were notified on September 30, 2026 about an upcoming retirement, but the notice does not give the exact shutdown date.

The release pages do not form a shared release calendar. OpenAI’s release notes document Library integrations for Box, Dropbox, and SharePoint; Claude’s release notes describe memory defaults; and Gemini’s app release updates say Skills are rolling out globally to users aged 18 and over. October 1 is the read date for this comparison, not a claimed launch date for those entries.

Google says access to the 2.5 models is limited to users who previously used them, but they are not deprecated and will remain available through the API until further notice. Its deprecations page recommends 3.5 Flash-Lite or 3.8 Flash for new projects. The Gemini API changelog lists Gemini 3.7 Flash at an introductory price through December 31, 2026, and records a file-input limit increase from 20 MB to 100 MB.

Keep verification inside the test. OpenAI warns that ChatGPT can produce incorrect or misleading output, while Anthropic advises against relying on Claude as the only source of truth. Gemini Apps documents a Sources button when sources are available. Use a source sheet and an answer key rather than treating fluent output as verification.

The test plan

  1. Establish a repeatable baseline. Prepare a fictional email: “Order A-17 was placed May 4; move it to 18 Oak Street; email Maya.” Submit: “Classify the request, list each fact you used, and quote the sentence that supports it.” Record the displayed model label, plan, enabled tools, answer, corrections, and elapsed time. Repeat the same packet in a fresh conversation to establish ordinary repeatability.

  2. Compare a documented version change. After a release or deprecation entry appears, rerun the unchanged packet and exact instruction. Record the entry date, old and new model labels, changed or missing fields, and whether the difference affects your work. Treat the difference as an observation, not proof that the announced update caused it.

  3. Measure the feedback clock. Where a feedback channel exists, submit: “When I paste the fictional order A-17, I expect the assistant to identify the changed address. Instead, it returns only the order number. The expected answer is 18 Oak Street.” Record the submission time, channel, first acknowledgment, and any public status change. An acknowledgment is not a fix.

  4. Check documentation and sources. Open the release note or deprecation page for the label your app displayed. Record whether it provides a date, rollout status, replacement, and shutdown date. Then submit a four-claim source sheet with: “Answer using only this sheet. Quote the exact supporting text for each answer and mark every unsupported claim.”

  5. Test your actual workflow. Write pass conditions before running a representative task. Submit: “You are preparing a beta-page launch for August 12. Remove the old banner, notify support, and document a rollback check. List the steps, owner, and verification evidence for each.” Record pass or fail, manual corrections, tool errors, and elapsed time. Compare vendor benchmark numbers only when the benchmark, model, task, settings, and date match.

How to read your results

Put each observation into four fields: hard constraint, observed result, evidence, and next action. Apply nonnegotiable requirements first: model availability, plan access, data settings, and a retirement date that conflicts with your release schedule. Then choose among the options that pass using your own workflow record.

Do not turn different tests into a composite score unless you define the weights before running them. A changed answer after a release note does not establish causation, and a longer response does not establish a regression. “Not observed” is more useful than filling gaps with a vendor benchmark or an assumed result.

Where the plans differ

Use the three-way comparison matrix to check the current cells. As read on October 2026, these rows directly affect a repeated version test. Prices are in US dollars.

Matrix row Documented difference
Main paid plan OpenAI’s Plus Help Center lists Plus at $20 monthly. Anthropic’s pricing page lists Claude Pro at $20 monthly or $17 annually with $200 up front. Google’s US plans page lists Google AI Pro at $19.99 monthly.
Context window OpenAI’s pricing page lists Instant limits of 27K on Free, 54K on Go and Plus, and 128K on Pro; reasoning limits are 256K on Go and Plus and 400K on Pro. Anthropic’s pricing page says up to 1M on every plan, varying by model. Google’s plans page gives no app token figure but lists uploads of up to 1,500 pages.
Usage limits OpenAI and Anthropic publish no exact message counts on their pricing pages. Google says compute limits refresh every five hours up to a weekly limit, with paid tiers at 2x, 4x, and up to 20x; AI credits can extend them.
Training on chats OpenAI offers opt-out on Free, Go, Plus, and Pro. Anthropic offers opt-out on Free, Pro, and Max and says Team is not trained on by default. Google’s plans page does not state a chat-training setting.
Ads OpenAI’s pricing page says Go may include ads and makes no such statement for the other listed plans. The pages read for Claude and Gemini contain no app advertising statement.

Sources

For current prices and limits, see the dated comparison pages.

Compare assistants