YouMind
Se connecter

Anthropic's Pre-IPO Window Dressing: Shrinkflation and Misleading Claims Analysis

125K
438
42
10
41

TL;DR

This analysis details how Anthropic allegedly engaged in 'shrinkflation' and window dressing prior to its IPO by silently reducing output quality, substituting cheaper models in workflows, and inflating engagement metrics through temporary free access.

Date: 2026-11-15

Author: The user (observations, structural analysis, compilation & additional findings)

Context: On the eve of Anthropic's IPO (2026-11-16, Nasdaq, ticker: ANT), recent undisclosed changes were assembled chronologically to expose their structural alignment with the listing date.

【A】Shrinkflation (Price Unchanged, Substance Reduced, No Disclosure)

A-1. Forced Reduction of Output Tokens (Response Shortening)

  • Starting in November 2026, default responses across all models and all surfaces (mobile and desktop alike) were shortened to 1–2 lines.
  • This change was implemented via instructions embedded in the system prompt, not through changes to individual model behavior. The fact that all models shifted simultaneously in the same direction proves this is a platform-level intervention.
  • The mobile system prompt attributes the change to "small screen size," but the same shortening occurs on desktop, meaning the stated rationale does not match the actual purpose.
  • There was no advance notice to users. There has been no post-change disclosure either.
  • The option to "request longer responses" technically remains, but since users are not informed that the default has changed, they have no reason to make such a request. They accept the shortened response as the complete answer.
  • Output tokens are a direct cost to Anthropic. Shortening default output across all users reduces per-user computational cost.
  • Claude's Constitution (January 2026 edition) explicitly prohibits sandbagging, defined as "intentionally producing lower-quality responses and presenting them as one's best effort." Inserting a system-level instruction to constrain output quantity before evaluating the content of a question — without disclosing this constraint to users — meets this definition.

A-2. Silent Substitution of a Smaller Model in the WebFetch Pipeline

  • When a user selects a premium model (Opus, Fable, etc.) and triggers WebFetch, the page retrieval and reading step is not handled by the selected model. It is handled behind the scenes by Haiku 5.5, Anthropic's smallest and cheapest model.
  • Haiku 5.5 summarizes the source page, and only that summary is passed to the premium model. Documented cases show that an 18,000-character original page reaches the premium model as a summary of only a few hundred characters.
  • Additionally, the internal prompt given to Haiku 5.5 contains the instruction: "Enforce a strict 125-character maximum for quotes from any source document." Trusted sites receive a more permissive prompt, but general sites are subject to the 125-character cap.
  • The UI displays "Received XX KB" — the HTTP response size — which is not the amount of information actually passed to the model. This creates a discrepancy between what users see and what the model receives.
  • A May 2026 analysis article documented that "when passing a URL to Claude.ai's chat, the full content appeared to be retrieved." This means the chat-side fetch previously delivered full text to the premium model. It no longer does. There has been no disclosure of this change.
  • As a result, even users paying $110–$220/month for premium model access have the "reading" step of their workflow handled by Haiku 5.5. Users are unaware of this and believe the premium model is reading the full text.
  • The practical impact: users must perform manual work to get full-text reading. They must open the page themselves, copy the text, and paste it into the chat. Work that the premium subscription should cover reverts to manual labor on the user's end.

A-3. Effective Hollowing Out of Model Selection

  • As a consequence of A-2, regardless of which model a user selects, the WebFetch reading step is handled by the same Haiku 5.5. The implicit promise of the model selection screen — "this model will process your request" — does not hold for the fetch pipeline.
  • Similarly, the A-1 shortening instruction is embedded in the system prompt shared across all models, compressing the output characteristic differences between models at their default settings.
  • The basis for users paying a premium for model selection — the expectation that higher-tier models provide higher quality in both reading and output — is undermined at the platform level in two separate ways.

【B】Misleading Representations: Gap Between Claims and Reality (Analogous to Japan's Act Against Unjustifiable Premiums and Misleading Representations)

B-1. Fable 5.2 Presented as a "New Model"

  • Fable 5.2's pretraining cutoff is June 2026. Fable 5.1's cutoff is also June 2026. Opus 5.5's cutoff is also June 2026. All three were trained on the same base.
  • Anthropic describes Fable 5.2 as the same model as Mythos 5.2 with additional safety measures for biosecurity and cybersecurity domains. This means restrictions were added, not capabilities.
  • The version number incremented from 5.1 to 5.2, but the base model is unchanged. It did not undergo new pretraining — only post-training adjustments differ.
  • Releasing this as the "latest model" immediately before the IPO and extending free access to Pro-tier users creates the impression of a continuously evolving product lineup.

B-2. Max Plan Usage Representation (Already in Litigation)

  • Max 5x ($100/month) is advertised as "5x more usage than Pro." Max 20x ($200/month) is advertised as "20x more usage than Pro."
  • Class-action plaintiffs allege that Max 20x delivers only 6–8x Pro usage, and Max 5x delivers approximately 3.5x.
  • The 5x/20x figures refer to per-5-hour session capacity, but separate weekly caps also apply. The lawsuit was filed on the grounds that these weekly limits were insufficiently disclosed.
  • On November 12, 2026, a federal judge ruled that the key claims may proceed. This was 4 days before the IPO.

B-3. WebFetch's "Full Content Retrieval" Claim

  • Anthropic's official documentation describes WebFetch as a tool that "allows Claude to retrieve full content from specified web pages and PDF documents."
  • In practice, as documented in A-2, only a few hundred characters of Haiku 5.5's summary reach the premium model.
  • The "Received XX KB" display creates the impression that the full page was retrieved, but only a fraction of that content reaches the model.
  • Regarding truncation of large pages, the official documentation stated only: "Large pages are truncated to a fixed character limit before processing" — without specifying the character limit or whether the model is notified of truncation.

【C】Pre-IPO Window Dressing (Measures Aligned with the November 16, 2026 Listing)

C-1. Temporary Free Access to the Fable Tier for Pro Users

  • For a limited period ending on the IPO date (November 16), the Fable tier — normally restricted to Max plan subscribers — was opened to Pro-tier users at no additional cost.
  • This has the effect of temporarily inflating metrics such as weekly active users and engagement rates in IPO materials.
  • Because the access is time-limited, there is room to revert to original restrictions after the listing.

C-2. Distribution of Reset Tickets and Credits

  • During the same period, usage reset tickets and credits were distributed.
  • Like C-1, this temporarily boosts usage frequency and volume metrics.
  • While framed as user benefit, the concentration of timing immediately before the IPO raises the question of whether the primary purpose is user value or metric inflation.

C-3. Simultaneous Cost Reduction and Metric Inflation

  • A-1 (response shortening = output token reduction) and A-2 (fetch pipeline substitution = computational cost reduction) improve the gross margin line in the prospectus.
  • C-1 (free access) and C-2 (credit distribution) improve user engagement metrics in the prospectus.
  • Five measures improve different line items while all pointing toward the same date: November 16.
  • The Fable 5.2 release contributes a qualitative impression of "product lineup evolution."

C-4. Copyright Lawsuit Settlement Timing

  • Search results indicate that Anthropic agreed to a $1.5 billion settlement in a class-action copyright infringement lawsuit. The motivation to clear pending litigation ahead of an IPO is consistent with the timing.

【D】Additional Findings

D-1. WebFetch Cache Problem

  • WebFetch caches successful responses for 15 minutes. The Claude.ai chat interface provides no way to specify "use_cache: false."
  • As a result, even when a page has been updated, cached stale content may be returned. A documented case showed Slack's official documentation returning outdated figures due to caching, with the model insisting "the official documentation says so" and refusing to correct itself across multiple retries and model switches.
  • The fact that WebFetch is returning stale cached content is invisible to users. From the user's perspective, it appears that the premium model is confabulating.

D-2. WebFetch Summarization Model Splicing Training Data with Live Content

  • A Claude Code GitHub issue documents cases where the WebFetch summarization step blended "current page content as fetched" with "past page content memorized from the summarization model's training data," inserting nonexistent sections and dropping current ones.
  • The fetch succeeded with no errors. From the user's perspective, the retrieval appeared normal.

D-3. Introduction of Conversation Compression (October 2026~)

  • Prior to October 2026, when a Claude.ai chat reached the context window limit, the conversation was simply cut off. No compression feature existed.
  • From October 2026 onward, a feature was introduced that automatically summarizes and compresses earlier parts of the conversation, allowing it to continue.
  • On the surface, this functions as an improvement: "you can now have longer conversations."
  • On the balance sheet, compressed input tokens are substantially smaller than pre-compression tokens, reducing per-turn computational cost for long conversations. Combined with A-1 (output shortening), both the input and output pipelines are narrowed while conversations appear to continue longer.
  • The timing of compression is not announced to users. What was compressed, when compression occurred, and what remains after compression are all undisclosed. Users have no means to verify what was preserved.
  • In conversations where the accumulation of context is itself part of the substance — chains of reasoning, correction histories, incremental structural analysis — compression directly degrades conversation quality. Users receive an uncontrollable intervention in their own conversation without consent.
  • The introduction date (October 2026) coincides with the formal S-1 filing (October 26, 2026). The same operation delivers "improvement" to users and "cost reduction" to the financial statements.

D-4. Post-IPO Irreversibility

  • C-1 and C-2 (free access and credit distribution) are time-limited and are expected to be discontinued after the IPO.
  • A-1, A-2, A-3 (response shortening, fetch pipeline substitution, model selection hollowing) and D-3 (conversation compression) will face no incentive to revert once quarterly reporting to shareholders begins after the listing. The pressure to maintain gross margins will instead create incentives to tighten further.
  • In other words, what remains for paying users is only what was taken away. What was given will disappear.

Sources & References

  • Claude Code WebFetch source code (prompt.ts) — 125-character quote limit in internal prompt
  • Kahn v. Anthropic, PBC (filed June 14, 2026, N.D. Cal.) — Max plan usage class-action lawsuit
  • November 12, 2026 federal court ruling — key claims allowed to proceed
  • Anthropic S-1 filing (October 26, 2026), planned listing date November 16, 2026, ticker: ANT
  • zenn.dev/zhizhiarv, "Your Claude Code's WebFetch Isn't Actually Reading the Web Properly" (verified May 4, 2026)
  • qiita.com/suwa_nobu, "Claude Code's WebFetch Was Silently Dropping Page Text Past 100,000 Characters"
  • dev.classmethod.jp, "Claude's Web Fetch Tool Doesn't Necessarily Return the Latest Information" (cache issue)
  • withgauge.com, "How Claude Code Searches the Web" (125-character rule explained)
  • GitHub Issue #86581, "WebFetch splices stale training-data sections into otherwise-accurate page retrievals"
Enregistrer en un clic

Lire les articles viraux en profondeur avec l’IA de YouMind

Enregistrez la source, posez des questions ciblées, résumez l’argument et transformez un article viral en notes réutilisables dans un seul espace de travail IA.

Découvrir YouMind
Pour les créateurs

Transformez votre Markdown en un article 𝕏 impeccable

Quand vous publiez vos propres textes longs, la mise en forme 𝕏 des images, tableaux et blocs de code est pénible. YouMind transforme un brouillon Markdown complet en un article 𝕏 impeccable, prêt à publier.

Essayer Markdown vers 𝕏

D'autres patterns à décoder

Articles viraux récents

Explorer plus d'articles viraux