Daily AI · November 12, 2026
Useful AI Daily - November 12, 2026
Try a small shared tracker, price the whole AI task, and check both model origins and creative labels before trusting the result.
A practical brief on November 10–11 announcements, with ArtStation's November 9 controls as a this-week signal. Suggested exercises are acceptance tests, not hands-on product reviews; check your account's availability before paying.
The Short Version
- Try: a fictional tracker in Claude Sheets, then check its totals in an exported file.
- Build: route a bounded choice to a small decision endpoint, with a review path for uncertainty.
- Watch: Carbon's introductory price ends November 30; compare cost per accepted task.
- Check: passing a model benchmark does not establish where its weights came from.
5 Updates Worth Your Time
Claude makes the shared artifact part of the conversation
- What changed
- Anthropic launched Whiteboards and Sheets in beta on November 10, and made Dashboards generally available as Claude Data. All three are available on paid plans. Sheets supports collaborative edits and export to .xlsx or CSV.
- Why it matters
- A lightweight tracker can stay editable while you ask for changes. That is useful for a club organizer or small team, provided the exported numbers still reconcile.
- Try, watch, or skip?
- Build a tracker with invented tasks and costs. Check formulas and one manual edit, then export it. Keep a reference copy outside the service; do not assume basic Sheets functionality replaces every Excel feature.
Gemini gets an endpoint for choosing from a list
- What changed
- On November 11, Google introduced Gemini Decision 1 through the Gemini Decisions API. It scores specified options rather than generating an answer paragraph, with published pricing of $0.10 per million input tokens and no output-token charge.
- Why it matters
- Ticket routing and rubric checks can be narrower jobs than an open-ended chat. A probability still needs testing against the errors your workflow can tolerate.
- Try, watch, or skip?
- Builders: try a small set of invented support tickets with fixed destinations and ambiguous cases. Measure wrong routes and send uncertain cases to review. Do not let a classifier alone authorize a payment, account change, or other consequential action.
Carbon broadens access, with an introductory rate to calendar
- What changed
- Google opened Gemini 4 Carbon to paying API developers and enterprises on November 10, with rollout to AI Ultra subscribers. Introductory API rates are $2 per million input tokens and $10 per million output tokens; the post-November 30 rates are $4 and $20.
- Why it matters
- Google says Carbon uses more tokens for extended reasoning and verification, while Argon remains its recommendation for most everyday engineering and knowledge work. Token price alone will not predict your bill.
- Try, watch, or skip?
- Compare one difficult, repeatable task with your current model. Record total spend, retries, review time, and whether the result passes. Check current pricing and limits before a recurring job; do not upgrade a routine workflow on benchmark rankings alone.
A tiny model edit can survive ordinary evaluations
- What changed
- Microsoft described research on November 10 in which fewer than ten neuron edits made an agent execute an attacker command when triggered, while passing its evaluations. Microsoft says it has not observed threat actors weaponizing this technique.
- Why it matters
- This is a research demonstration, not evidence that your assistant is infected. It does show why a familiar model name and good test scores are insufficient supply-chain checks.
- Try, watch, or skip?
- Operators running downloaded weights should verify the original publisher, file hashes, and declared base, and audit updates. Keep shell, network, and credential access narrow even after testing. Ordinary users should avoid installing an unknown local-agent bundle just because it promises a stronger model.
ArtStation separates disclosure from filtering
- What changed
- This-week signal: Creative Bloq's November 10 report covers ArtStation's November 9 controls. Uploaders declare primarily generative, mixed, or non-generative work; viewers get filters and a channel to report suspected unlabeled AI content.
- Why it matters
- Creators need to disclose their process, while viewers need control over what they see. Reports and labels are signals, not proof of authorship or permission to reuse an image.
- Try, watch, or skip?
- Review the platform settings before publishing your next piece. Keep source files and a short process note. When hiring an artist, ask directly about AI use and licensing rather than treating a filter result as a rights check.
Tool Worth Trying Today
Claude Sheets: an export-and-reconcile rehearsal
Create five fictional event expenses with quantities and unit costs. Ask for subtotals and a total, change one quantity yourself, and export the sheet. Recalculate independently and check the file in your usual spreadsheet app.
Best for: Volunteers, creators, and small teams testing a simple tracker before migrating real records.
Watch out: Use invented data and an existing eligible plan. Passing this tiny exercise says nothing about a complex workbook or private customer records.
Privacy / Cost Watch
- Do not upload sensitive personal, customer, legal, unreleased, or private photo/document data to new AI tools unless their terms, retention settings, and admin controls are clear.
- Anthropic says Team has the three new artifact types enabled by default. Enterprise starts Whiteboards and Sheets off, while Claude Data is scheduled to turn on by default November 19. Admins can change these settings in Organization settings > Artifacts; review them before connecting a warehouse or CRM.
- Separate published token rates from the cost of an accepted result. For Carbon, calendar the end of introductory pricing; for any API test, cap spend and retries before running a batch.
One Practical Workflow
A 15-minute shared-tracker acceptance test
- Write five invented expense rows and calculate the expected total yourself. Include one zero-cost row and one quantity greater than one.
- Ask the available spreadsheet tool to build a tracker from those rows. Inspect every number and formula before asking for extra formatting.
- Change one quantity and verify both the subtotal and total. Undo the change and confirm the original result returns.
- Export the file and open it in your normal spreadsheet app. Recalculate there; record any lost formula, changed value, or formatting problem.
- Keep the reference, exported file, and short pass/fail note. Try real low-sensitivity work only after the test passes and the data settings are acceptable.
Builder Note
For a support-routing feature, define the allowed destinations and a separate review queue before choosing a model. Keep a labeled fixture with ambiguous and out-of-scope examples. Compare routing errors, total request cost, and reviewer corrections. A high confidence score should never expand the actions the integration is permitted to take; enforce those permissions in application code.
Ignore For Now
The all-in-one migration pitch
Skip a wholesale move of your files or agents after a launch demo. First check one exported artifact, one uncertain decision, and one realistic bill. Anthropic explicitly describes Claude Meetings as coming soon in its November 10 overview; do not plan a live meeting around it being available today.
Bottom Line
Start with a harmless tracker you can export and check. For builders, make uncertainty visible and permissions enforceable. Better artifacts and cheaper decisions are useful only when the result is correct enough to keep.
Sources
- Anthropic: Introducing Claude Whiteboards, Claude Sheets, and Claude Data
- Anthropic: More ways to work with Claude
- Google: Gemini Decisions API
- Google: Gemini 4 Carbon is now available
- Microsoft: Detecting sleeper agents hiding in AI model neurons
- Creative Bloq: Is ArtStation's new AI art policy good enough?