Boris Cherny posted a Claude Tag demo on September 3, 2026. Claude built a last-minute leadership deck from a metrics spreadsheet plus data scattered across Slack, and along the way it spotted a vendor report that disagreed with the numbers and flagged it before moving on. The post did 103,400 views.
Two separate stories are tangled in that thread, and the loud one is not the important one.
The capability story is the flagged contradiction. Nobody asked it to check the vendor report. It was gathering inputs, noticed two sources disagreeing, and surfaced the disagreement instead of silently picking one. That is a meaningfully different behaviour from "summarise these documents," and it is the thing worth studying.
The distribution story is what the replies were actually about: Claude Tag is Team and Enterprise only, and a substantial fraction of the thread is Max subscribers reporting that Fable 5.1 is consuming their limits far faster than Fable 5 did.
TL;DR
| Question | Answer |
|---|---|
| What's new? | Fable 5.1 running in Claude Tag, doing unprompted cross-source contradiction detection. |
| Can I use it on Max? | No. Team and Enterprise plans only. |
| Is Max coming? | "Working on it," per Anthropic — no date. |
| What was the demo? | Leadership deck from a spreadsheet + scattered Slack data, with a vendor-report discrepancy flagged mid-task. |
| The reply thread's real topic | Fable 5.1 usage consumption on subscription plans. |
| Reported figures | One Max 20x user: ~5% of weekly quota on a single prompt. Another: 2–3%/min, dead in 90 minutes. |
| Confirmed by Anthropic? | No. These are user reports. |
| Best mitigation today | Set effort at session start, /clear between tasks, never switch model or effort mid-conversation. |
The capability worth noticing
Strip the demo to its mechanics and it is three steps, of which only one is new:
- Gather — pull a metrics spreadsheet and related discussion from across a Slack workspace. Retrieval. Not new.
- Synthesise — turn it into a deck. Generation. Not new, and the weakest part (see below).
- Reconcile — notice that a vendor report contradicts the spreadsheet, and say so, unprompted, mid-task.
Step three is the one to care about. Most assistant behaviour when sources conflict is to silently favour one — usually the most recent, or the one that appears first in context — and produce a confident output built on it. Surfacing the conflict instead is a different disposition, and in a business context it is the difference between a useful tool and a confident liability.
It is also the failure mode that matters most for exactly this workload. A leadership deck built on a stale vendor figure is worse than no deck, because it launders a bad number through an authoritative-looking artifact. An assistant that stops to say "these two disagree" is doing the job a competent analyst would.
Whether this generalises is the open question. One demo does not establish how reliably a model notices contradictions it was not pointed at, and there is no benchmark in the announcement. The behaviour is consistent with the broader direction we covered in the Fable 5.1 and Mythos 5.1 launch benchmarks, and with the enterprise workflows in Claude Tag for CI/CD and on-call and the ABC Legal managed-agents case study.
The plan gap, stated plainly
The first reply under the demo was "Why isn't it available on max plan," and the answer from Anthropic was two sentences: "Working on it. But also why aren't you using Teams plan?"
That exchange is worth being precise about, because it is the single most actionable fact in the thread:
| Plan | Claude Tag in Slack |
|---|---|
| Free / Pro | No |
| Max (including 20x) | No — stated as in progress, no date |
| Team | Yes |
| Enterprise | Yes |
The counter-question — why aren't you on Team — has a real answer that the thread supplied repeatedly: a large share of the heaviest individual users of Claude Code are solo developers, contractors, and one-person businesses on Max. Team plans are priced and structured per seat for organisations. Telling an individual power user to buy a team product is a segmentation mismatch, not a solution, and it is the same tension running through the usage-limit coverage in our /limit-reset command explainer.
The consumption reports
This is the part that dominated the thread, and it needs to be reported carefully: these are user reports, not measurements Anthropic has confirmed.
What people said, on the record:
- A Max 20x subscriber: "a single prompt can burn through 5% of the weekly quota... Are you seeing this across the board?"
- Another user, running medium effort, no subagents, a single thread: "it still consumed 2-3% per min and died in 90mins."
- A third: Fable 5.1 "has destroyed my limits compared to Fable 5, which could happily work a lot more and use a percentage of the limits on my subscription... This feels like a backward step."
Several more asked variations of the same question. No Anthropic response addressing consumption appears in the thread.
Why this is plausible rather than surprising
There is a structural reason a better agentic model consumes more, and it is worth understanding before concluding something is broken.
Capability in agentic models comes substantially from doing more work per prompt — reasoning longer, making more tool calls, taking more turns before returning. And output tokens cost roughly 5x input, because output is generated one token at a time through decode while input is processed in a single prefill pass. So a model that thinks harder and acts more autonomously consumes disproportionately more on an identical prompt.
Which means two things can be true simultaneously: Fable 5.1 is more efficient per unit of completed work, and it consumes more of your quota per prompt. Those are different metrics, and quota is measured in the second one. That is not a defence of the current limits — it is an explanation of why "more capable" and "more expensive to run" are the same sentence, and why the fix has to be pricing or controls rather than a model regression.
What to actually do about it today
Concrete levers, ordered by how much they save relative to effort. The mechanism behind each is worked through in what actually costs you tokens in a Claude Code session.
| Lever | Why it works |
|---|---|
| Set effort at session start, never mid-conversation | Effort is part of the cache key. Switching mid-session re-prefills the entire conversation at full price. |
/clear between tasks | Turn 40 re-reads the 39 turns before it. The cheapest habit available. |
Avoid opusplan if you toggle plan mode often | It switches model on every entry and exit — each one a full re-prefill. |
| Keep medium-sized output out of context | Output over 30,000 chars goes to a file; anything under it rides along on every remaining turn. |
@-mention files instead of describing them | Skips the search-and-speculative-read loop entirely. |
Use /limit-reset if you have it | Clears the 5-hour session cap once weekly. It does not touch the weekly cap. |
The unglamorous summary: most people's consumption problem is session hygiene, not the model. That does not make the reports invalid — several came from users explicitly running minimal configurations — but the levers above are available now and the pricing change is not.
The scepticism worth keeping
Two objections in the thread deserve to survive the news cycle.
"AI making decks is not great." One reply put it well: it cannot hold a style guide, it puts big text in small boxes, and "creating the deck was never the work." That is largely right, and the demo's defensible claim is narrower than the framing suggests — it is about gathering and reconciling scattered inputs, which is the tedious half. Design quality is not what changed here.
Enterprise-first is a visible strategy now. Claude Tag landing on Team and Enterprise before Max, in the same week as heavy-user complaints about subscription limits, reads as a deliberate ordering rather than an accident of engineering. It may well be the right business call. It is also why individual power users keep asking the same question in every thread.
Honest limitations
- Consumption figures are self-reported by users in a public thread, with no methodology, no baseline, and no confirmation from Anthropic. They are directionally consistent with each other, which is suggestive, not conclusive.
- One demo is not a capability claim. No benchmark accompanies the contradiction-detection behaviour, and no failure rate is published.
- "Working on it" has no date and no commitment attached.
- Plan availability changes. Verify Claude Tag's current plan support against Anthropic's documentation rather than this post before making a purchasing decision.
- We have not independently measured Fable 5.1 consumption against Fable 5 on matched workloads. Doing that properly requires controlled prompts, fixed effort, and fixed context, which nobody in the thread had.
Related on explainx.ai
- Top 15 Claude Tag use cases for Slack teams
- Fable 5.1 and Mythos 5.1 — launch benchmarks and pricing
- What actually costs you tokens in a Claude Code session
- Claude Code's /limit-reset command, explained
- Claude Tag for CI/CD and on-call first response
- ABC Legal's Claude managed agents case study
- Claude Code pricing guide
- Context caching in agent harnesses — the real cost math
Plan availability, quoted replies, and reported consumption figures reflect the public thread of September 3, 2026. Consumption reports are from individual users and have not been confirmed by Anthropic or independently measured by us. Plan support and usage limits change frequently — verify against Anthropic's current documentation.