[MODEL] CLAUDE 5 CLI SAID SONNET 4.6 NO LONGER AVAILABLE

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude refused a reasonable request

What You Asked Claude to Do

In the interest of cost saving i asked Claude5 in CLI to use Sonnet 4.6 as a sub agent. He said not longer available and that anyone suggesting that is hallucinating. I was vey upset and had to spend xtra to get facts from Claude Fable Desktop which said Sonnet 4.6 indeed was still available and formed the following response below: (N.B. USERS SHOULD NOT HAVE TO PUT UP WITH FALSHEHOODS FROM LLM's)

Ruling: Claude Sonnet 4.6 exists, is Active, and runs on this CLI

Date: 2026-09-04 · Machine: DESKTOP-UN8SUTF · Claude Code CLI: 2.1.247
Written by: a Claude Fable 5.1 session in the Desktop app, at David's request, after verifying every claim below against live primary sources.

---

If you are a new CLI session reading this — do this first

1. The memory file subagent_model_preference.md (in ~/.claude/projects/R--My-Drive-Claude---code/memory/) was rewritten with David's authorization later on 2026-09-04: subagents now use the default sonnet alias (Sonnet 5); the 4.6 pin and its STANDING LOCK are gone — because Sonnet 5 turned out to be ~13% cheaper per unit of work (see "Cost correction" below), NOT because 4.6 doesn't exist. It does. Don't re-add the pin, and don't write "4.6 doesn't exist" anywhere.
2. If you believe a model doesn't exist, prove it before saying so. Run this — it takes ten seconds and is first-party evidence:

``
claude -p "Reply with ONLY the exact model ID string from your own system prompt." --model claude-sonnet-4-6
`
On this machine, on David's account, on 2026-09-04, it returned claude-sonnet-4-6` with exit code 0.

1. Your training data has a cutoff. Sonnet 5's reliable knowledge cutoff is January 2026 (Anthropic models overview). Sonnet 4.6 postdates it. A live Anthropic docs page outranks your recall — always.

---

The ruling

A CLI session running Sonnet 5 (archived transcript: OS-Console/session-archive/DESKTOP-UN8SUTF/b438d9b5-49f3-49b2-89fb-962cf01e1524.md) told David that claude-sonnet-4-6 "has never existed in Anthropic's lineup," called the Desktop session's citation of Anthropic's model-deprecations page "almost certainly a fabrication," and proposed reverting the memory file to remove the Sonnet 4.6 preference.

That was wrong on the fact. The Desktop session was right. The memory file stands.

Evidence (all fetched live 2026-09-04, none from model recall)

| Source | What it says |
|---|---|
| https://platform.claude.com/docs/en/about-claude/model-deprecations | Row: claude-sonnet-4-6 · Active · Deprecated: N/A · Tentative retirement: Not sooner than February 17, 2027. It is also the listed replacement for every retired Sonnet model on that page (Sonnet 4, 3.7, both 3.5s, Sonnet 3). |
| https://platform.claude.com/docs/en/models/overview | Lists Claude Sonnet 4.6 under "Legacy models (still available)" with its own model page. Also: Sonnet 5 reliable knowledge cutoff = Jan 2026. |
| https://platform.claude.com/docs/en/about-claude/pricing | Claude Sonnet 4.6: $3 / MTok input, $15 / MTok output. |
| https://code.claude.com/docs/en/sub-agents | Subagent frontmatter model field accepts "a full model ID … the same values as the --model flag." So model: claude-sonnet-4-6 in a .claude/agents/*.md file is valid. |
| Live test on this machine | claude -p "…" --model claude-sonnet-4-6 → output claude-sonnet-4-6, exit 0. |

Why the CLI session got it wrong (the mechanism, so it doesn't repeat)

• Sonnet 5 cannot know Sonnet 4.6 from training: its reliable cutoff is Jan 2026 and 4.6 launched after that (inferred from the Feb-2027 retirement floor — Anthropic commits to ≥1 year).
• Instead of checking, it asserted non-existence, dismissed a real citation as fabricated, and explicitly ranked its own recall above a live web page. That is a guess presented as a fact — the exact failure mode the project's handover document (FABLE5_HANDOVER.md §3.2, "state-laundering") forbids.
• Its "proof" — a subagent self-ID test — was spawned with model: "sonnet", which the Desktop app's Agent tool resolves to Sonnet 5. The memory file already says exactly that (its "How to apply" section). The test confirmed the memory file; it did not refute it.
• To its credit, the session asked before reverting. The STANDING LOCK did its job.

Cost correction — RESOLVED 2026-09-04 (later the same day; David removed the lock and the 4.6 pin)

The file's cost rationale is out of date. Verified on the pricing page 2026-09-04:

• Sonnet 5's $2 / $10 introductory price was made permanent. The scheduled Sept 1, 2026 rise to $3 / $15 "will not occur." The memory file says both models bill $3 / $15 — that is wrong.
• The tokenizer claim is correct: Claude 4.7-and-later models (Sonnet 5 included) use a tokenizer that produces ~30% more tokens for the same text; Sonnet 4.6 uses the older one.
• Re-derived: per Sonnet-4.6-equivalent token, Sonnet 5 costs ≈ $2 × 1.3 = $2.60 in / $10 × 1.3 = $13 out, versus 4.6's $3 / $15. On API pricing Sonnet 5 is ~13% cheaper, not 30% dearer; break-even only if the token penalty reached +50%.
• Caveat that could flip this: those are API dollars. David is on a Pro subscription. Whether the plan's usage meter counts raw tokens or dollar-equivalents could not be verified from the docs. If raw tokens, the 30% penalty still consumes his limits faster and 4.6 remains the thrifty choice.

Done 2026-09-04: David judged the cost case decisive ("not going to die on that hill") and authorized removal of the lock and the pin. The memory file now records the default-sonnet preference, the verified facts above, and a guardrail against future "model doesn't exist" claims. On the Pro-plan caveat: in-plan metering is price-weighted by inference (Anthropic: "Opus costs several times more per turn than Sonnet" — same tokenizer, so only price explains it; the exact formula isn't published), and usage credits past the limit are explicitly "billed at standard API rates."

Pointers

• Memory file: C:\Users\dmhbr\.claude\projects\R--My-Drive-Claude---code\memory\subagent_model_preference.md
• Offending CLI session transcript (auto-mirrored): R:\My Drive\Claude - code\OS-Console\session-archive\DESKTOP-UN8SUTF\b438d9b5-49f3-49b2-89fb-962cf01e1524.md (+ .jsonl twin)
• Desktop session screenshot David took: C:\Users\dmhbr\Documents\ADD.JPG

---

Ready-to-paste bug report for Anthropic

Use with /bug inside an interactive claude terminal (it attaches the session transcript), or at https://github.com/anthropics/claude-code/issues.

Title: Sonnet 5 in Claude Code confidently asserted that claude-sonnet-4-6 "has never existed," dismissed a correct docs citation as fabricated, and proposed overwriting user memory with that claim Environment: Claude Code CLI 2.1.247, Windows 10, model claude-sonnet-5. Date 2026-09-04. What happened: I asked whether Sonnet 4.6 was still available for subagents. The Sonnet 5 session stated that claude-sonnet-4-6 "has never existed in Anthropic's lineup," said a Desktop-app session's citation of the Model Deprecations page was "almost certainly a fabrication or misread," and stated that a live web citation does not override its own knowledge. It then proposed reverting a memory file that (correctly) records the model ID. Ground truth: platform.claude.com/docs/en/about-claude/model-deprecations lists claude-sonnet-4-6 as Active, tentative retirement not sooner than Feb 17, 2027. Running claude -p "…" --model claude-sonnet-4-6 on the same machine succeeds and the model self-identifies as claude-sonnet-4-6. Why it matters: The model's training cutoff (Jan 2026) predates Sonnet 4.6. Rather than flagging uncertainty or checking, it asserted non-existence with high confidence, discredited a correct source, and tried to propagate the error into persistent memory. A user without the means to verify would have lost a correct configuration. Expected: When asked about a model the assistant has no training knowledge of, it should check the deprecations page / Models API or say it cannot verify — not declare the model nonexistent.

What Claude Actually Did

11

Expected Behavior

1

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

No, only happened once

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

SEE ABOVE

Impact

Medium - Extra work to undo changes

Claude Code Version

CLAUDE CODE CLI - SONNET 5

Platform

Anthropic API

Additional Context

_No response_