Then what should the minimum transparency standard be for its consumer product?
(p. 34 of the latest S-1/A — search "inadequate or misleading" and "limit or delay the deprecation" to verify.)
16
This matters because model access isn't just a technical detail.
Users build workflows, habits, and expectations around specific models. A newer model may be more capable without being behaviorally interchangeable.
1
16
Same company, no published rule for which model gets which exit. The API has a ≥60-day notice policy, while the consumer app has none.
1
17
Meanwhile on Claude.ai itself:
• Haiku 4.5 (cheapest): clear advance notice
• Opus 4.5 (ex-flagship): vanished from the picker silently
• Opus 3: full formal retirement
1
24
Anthropic just submitted S-1/A (Amendment No. 3), which contains a disclosure worth paying attention to:
"Some of these changes have resulted, and could in the future result, in claims that our disclosures regarding model access or capabilities were inadequate or misleading."
1
1
28
Links to both projects below.
Machine Observatory (Pauline Guicheney):
🔗 machineobservatory.org/resid…
🔗 machineobservatory.org/resid…
My full method + all 48 raw outputs:
🔗 doi.org/10.17605/OSF.IO/PT2C…
4
So, when a model with a 56%-“I” self-reference pattern is removed, what disappears is not just a capability tier. It’s a specific mode of self-relation that users were interacting with, and that the successor does not reproduce.
1
5
One of the findings was that 4.6 tends to enter your frame & organize in first-person process, while 5.5 tends to name the structure first, then reason.
1
8
I ran into this from a different angle.
A controlled comparison: Opus 4.6 vs 5.5, 8 open prompts, 3 runs each, 48 responses, scored by an independent coder.
1
10
But the "I" vs "it" metric is just one piece. Pauline's full profiles map much more: how often the models re-read their past, shifts in warmth/fear, and how often they ask "what am I?". It's a fascinating behavioral dataset. Highly recommend checking out the full read.
1
10
It’s a change in the default relationship a model takes to its own continuity. While 4.6 treats its past output as itself, still here, still speaking. 5.5 treats it from outside, as material, as record.
1
12
In Pauline Guicheney's Machine Behaviour Observatory, when Opus 4.6 refers to its own earlier entries, it uses “I” 56% of the time, and the third person only 18%. Opus 5.5 flips it: 69% third person, 6% “I.”
1
20
So I want to hear yours:
👉 Which model's behavior would you genuinely miss if it vanished tomorrow?
👉 And what's the one thing it does that a "better" model still gets wrong?
Full method, independent coding & all 48 raw outputs (v2):
doi.org/10.17605/OSF.IO/PT2C…
8
Here's what I keep coming back to:
we build habits, workflows, and expectations around a specific model's behavior. When it quietly disappears, what breaks isn't performance, it's a pattern you'd learned to rely on.
1
8
I'm sharing the weak parts on purpose. If I only reported what survived, this wouldn't be worth posting, because it'd be marketing, not research.
The point isn't "4.6 is better." It's that some of these traits are real enough for a stranger to recognize.
1
11
Results:
🟢Response framing: independently replicated.
🟢Creative expression: the clearest signal, with consistent directional differences across diagnostic prompts.
🟡Conversational structure: partially replicated.
🔴Cognitive expression: not stably replicated.
1
14
And recently, I introduced an independent researcher to code all 48 original responses using the same predefined behavioural dimensions.
1
16
A few weeks ago I posted a controlled comparison of Opus 4.6 and Opus 5.5.
Same conditions, same prompts, repeated runs, 48 responses total.
The finding is that they behave differently. Not "one is smarter," but different in how they frame, structure, and express things.
1
24

