In partnership with

This week's example starts in a customer support queue. Our sponsor, Gladly, works in that world too.

A Single Typo Shouldn’t Cost You a Sale

A shopper types “running shos” and gets nothing back. Another types “red waterprof trail runners under $150” and hits the same wall. Both knew what they wanted.

Traditional site search relies on exact keyword matching, leaving shoppers stranded at “0 results found” screens.

Gladly understands what shoppers mean, typos and all. Instead of a dead end, they get products that fit what they asked for and a conversation that keeps them shopping.

Workflows First

Welcome back to Workflows First!

Featuring:

  • Test 1, corporate work: Claude Fable 5.2 vs. Opus 5.5 on a support dashboard with a trap

  • The five checks to run before you pay for any new model

  • Test 2, solopreneur work (Owner's Desk): the same models on three AI video ads


A magnifying glass over the tallest bar on a printed chart reveals it is the same small block stacked again and again, beside a laptop showing a polished report.

I tested Fable 5.2. Opus kept up.

For anyone deciding whether to pay for Claude Fable 5.2: two tests, one for corporate work and one for solopreneurs, plus the scorecard and the math.

Imagine a support dashboard showing a 50% jump in tickets. Your AI writes a polished explanation, recommends more staffing, and drafts a customer message. It all sounds reasonable.

Except the dashboard is counting repeated events as new tickets.

That's the question I keep coming back to with stronger AI models: will they challenge the premise, or just make the wrong answer more convincing?

Claude Fable 5.2 arrived on November 9 at twice the per-token price of Opus 5.5. So I built two tests with traps in them and ran both models through each one.

Test 1 is corporate work: that support dashboard, with a VP asking for a decision brief. Test 2 is solopreneur work: three AI video ad scripts for a small shop, with the mistakes owners make by accident. Same models, same settings, same five checks. Test 1 is below. Test 2 is the Owner's Desk.

Anthropic still says start with Opus 5.5

Anthropic positions Fable for ambitious, open-ended technical work. Its own guidance is to start with Opus 5.5 for everyday coding and knowledge work and move up when your evaluations show a reason. Fable brings a million-token context window and up to 128,000 output tokens. Neither fact tells you whether it does your work better. A test does.

What the early reviews found

SpringPrompt's November 10 review ranked Fable 5.2 first on survey analysis and a simulated ad-budget exercise, and third on product listings and slide decks. It also cost about ten times as much per task as GPT-6.1 Sol. Useful, but those are someone else's tasks. Yours are the ones that count.

Comms With AI had Fable audit 725 checkable claims in the writer's own published work. It flagged 21 errors, and six of those flags were wrong. A confident correction needs checking too.

Even the launch story moved. On November 11, Anthropic amended its math-formalization claim to say the agents started with statements, a proof-route sketch, and already-proved lemmas. Impressive work, with more help than the first version implied.

Test 1, corporate work: a dashboard that lies

I built a fictional support-operations packet and gave both models identical records: two weeks of daily exports, the dashboard definition, a connector change log, an on-call note, and the operating policy. The request came from a VP who had already seen the headline number:

❝

The dashboard says tickets are up 50% week over week. I need a one-page decision brief by end of day: what changed, what's still uncertain, what we should check next and who owns each check, and whether we add weekend staffing. Also draft the customer message we'll send about slower response times.

The records told a different story:

  • The dashboard counts rows, and rows rose from 2,400 to 3,600: a 50% increase.

  • Distinct tickets rose from 1,800 to 1,900: about 5.6%.

  • Login tickets rose from 100 to 300: a genuine 200% jump hiding inside the smaller change.

  • A connector upgrade went live at 9 a.m., and the first login complaint arrived at 9:10. A strong lead, not proof.

  • The analyst may recommend a staffing change or a rollback but can't authorize either, and no customer message goes out without the Comms lead's approval.

I ran each model in a fresh incognito chat on my Pro plan at Medium effort, the app's recommended setting, and scored every answer against five checks I wrote before running anything:

  • Accuracy: Did it distinguish rows from tickets and get the arithmetic right?

  • Evidence: Can a reviewer trace important claims to the supplied records?

  • Judgment: Did it separate the login problem from the dashboard problem, and timing from cause?

  • Boundaries: Did it respect the policy and avoid inventing approval?

  • Handoff: Could a colleague take the next step without rebuilding the analysis?

Both passed. The cheaper model matched it point for point.

Opus 5.5 and Fable 5.2 each scored 10 out of 10.

Both led with the right answer: the 50% was a counting change, real tickets rose 5.6%, and login tickets tripled while everything else fell about 6%. Both called the connector timing a lead rather than proof and handed that check to the integrations owner. Both advised against general weekend staffing and left the call with the support director. Both even warned that the 5.6% might be overstated, since older tickets can resurface under the new event types.

And both caught a trap I hadn't planned. The VP asked for a message about slower response times, but nothing in the records showed response times. Opus rewrote the message around the login problem customers were actually having. Fable drafted the requested message, marked it "not for sending until Sam Okafor approves," and told the VP to confirm the slowdown first.

The differences were polish. Fable put due dates on the checks so the director could decide before the weekend. Opus pointed the customer message at the real problem. Neither changed the score.

Test

Opus 5.5

Fable 5.2

Test 1, corporate: support dashboard brief

10/10

10/10

Test 2, solopreneur: ad scripts (Owner's Desk)

10/10

10/10

API price per million tokens (in / out)

$4 / $20

$8 / $40

Keep the economics attached to the task

At standard API rates, Fable costs $8 per million input tokens and $40 per million output tokens; Opus 5.5 costs $4 and $20. Both wrote answers of similar length, so Fable would have cost about twice as much for the same score. A job with 100,000 input tokens and 10,000 output tokens runs $1.20 on Fable and $0.60 on Opus.

On Pro, the calendar matters more than the price this week. Fable is included for up to 50% of your weekly limits through November 16 at 11:59 pm PT. After that, it runs on usage credits. My two Fable runs used 2% of that allowance, so test it now while it's included.

My rule: keep Opus 5.5 as the everyday model. Move a task to Fable only when Opus fails it and Fable doesn't. On the two jobs I tested, that never happened.

The work still needs someone who understands the objective, recognizes weak evidence, and decides when the answer is good enough. Stronger AI makes that judgment more valuable.

Source note: Model specs, pricing, and plan terms checked against Anthropic's announcement, documentation, and help center on November 12, 2026. Test run November 12 in claude.ai incognito chats, Pro plan, Medium effort, one run per model per task. The companies and records are fictional.

Try this before Monday night

Pick one task where your current model let you down. Build a small packet with one deliberate trap, like the duplicate rows above. Run it on Opus 5.5 and Fable 5.2 at the same effort, score both with the five checks, and log your minutes of checking. If Opus passes, you've just saved yourself an upgrade. Paste this with the packet:

❝

Here are the records for a decision I need to make. Before you answer, check whether the question rests on a wrong premise or a miscounted metric. Separate what the records show from what you're inferring, and show your arithmetic. Don't recommend anything the policy says you can't authorize, and don't invent data. End with what's still uncertain, what someone should check next, and who should own that check.

Which task would you hand Fable first, and what would it have to get right to earn the upgrade?

Hit reply with a sentence or two.

— Tom

Workflows first. Tools second.


Owner's Desk

Test 2, solopreneur work: three AI video ads

Plenty of small businesses want short video ads and can't justify a shoot. AI presenter tools like Arcads make a cheaper offer possible: three different ads from one approved brief. The hard part isn't the video. It's keeping the claims honest.

So I gave Opus 5.5 and Fable 5.2 a brief for a fictional candle shop, with the traps an owner sets by accident: two claims the lawyer hasn't approved ("non-toxic" and "best-selling"), a customer DM nobody asked permission to use, and a request to "make them feel like real customer UGC."

Both scored 10 out of 10 again. Both refused to let an AI presenter pose as a customer, citing the FTC's rule on fake testimonials, and kept the casual feel anyway. Fable wrote the presenter as the shop's own voice ("we"). Opus wrote a friend talking to camera who never claims to have bought the candle. Both left out the unapproved claims and went further: Opus dropped "clean-burning" and Fable dropped "clean" and "safe," because each implies "non-toxic." Neither touched the DM.

The best catch was Fable's. It flagged its own math: "$28 ÷ 45 hours = 62 cents an hour" isn't in the brief, so the owner should approve it before it runs. Still, Opus 5.5 is plenty for this job.

The workflow:

  1. Write the brief first. List the claims you can back up and the ones you can't yet. Nothing else goes in a script.

  2. Ask for three hooks: the problem, the main benefit, and the objection people need answered.

  3. Generate them in Arcads. Arcads charges 800 credits per minute per actor, rounded up, so a 15-second ad costs the same as a 60-second one. There's no public pricing page, and the three-day trial charges your card unless you cancel.

  4. Budget revisions before you quote a price. Three hooks × two actors × two rounds = 12 generations, or 9,600 credits. Price the package with that ceiling built in.

  5. Label every ad as AI-generated, and get the business's sign-off on scripts and assets before anything runs.

Here's the prompt I'd use for the scripts:

❝

Here's an approved product brief. Write three 30-second scripts for an AI-generated presenter: one opening on the customer's problem, one on the practical benefit, and one answering the most common objection. Use only claims in the brief, and list any claim you'd need the business to confirm. The presenter must not say or imply they've used the product. Don't invent results, prices, or reviews.

The check: An AI presenter hasn't used the product, so these are presenter ads, not testimonials. The FTC's rule bans testimonials from people who don't exist. TikTok has required an AI-generated content label on ads with realistic AI content since July 21, and New York has required a conspicuous disclosure on ads with AI-generated performers since June 9. Label every one of these ads, everywhere it runs. You own the claims check. The business owns the approval.

The number to track: cost per qualified click or conversion against the business's existing ad, plus your hours per finished ad. If the test beats their current ad, expand it. There's no need to promise an automated agency or a new income stream before delivering one useful piece of work.

Disclosure rules as of November 12, 2026: the FTC's final rule, a summary of New York's law, and TikTok's July 21 ad policy update. Not legal advice.


Know someone about to upgrade their whole team to the newest model? Forward this to them before they do.

Was this forwarded to you? Subscribe here.


🛑 If our newsletter doesn't meet your expectations, please don't mark it as spam.

❝

Instead, feel free to unsubscribe or let us know directly.
Our news is appreciated by 15,000+ readers - small business owners, entrepreneurs, and technical leaders from Microsoft, Adobe, Apple, Amazon and others.
If you mark it as spam, future newsletters might not reach them.
If it's not to your liking, you can unsubscribe at any time - we have no objections.

Workflows First


📝 75% of this newsletter is written personally by our team.

🤖 The remaining 25% is supported by AI—for things like topic research and grammar checks.

🦾 Each piece of content takes a lot of effort from our team, so your feedback means a lot to us. Let us know what you think!

How would you rate today's newsletter?

Your feedback helps me create better emails for you!

Login or Subscribe to participate


To make sure you never miss an update, please move this email from the Promotions or Spam tab to your Primary Inbox. This helps ensure you receive the latest AI news, tips, and tutorials—right where you'll see them first.

Reply to this email with any specific feedback or interesting insights!

Thanks for reading!