Replacement map
The "named replacement" column reproduces the replacement Anthropic names in the official deprecation ledger. Those entries were written when the notice was published, so several still point at Opus 4.8. As of September 19, 2026 the current Opus-class model is Opus 5 and the current highest-capability generally available model is Fable 5.1, so a migration done today should usually land on the newer ID.
| Old model or family | Status in official ledger | Named replacement | Action |
|---|---|---|---|
| Claude 1 and Claude Instant IDs | Retired November 6, 2024 | claude-haiku-4-5-20251001 | Replace hard-coded IDs and re-test prompts. |
| Claude 2.0 / 2.1 | Retired July 21, 2025 | claude-opus-4-8 | Re-baseline quality, latency, and cost against a current model. Context and behavior differ materially. |
| Claude Sonnet 3.5 IDs | Retired October 28, 2025 | claude-sonnet-4-6 | Apply Sonnet 4.6 migration notes, especially around assistant prefilling and tool versions where relevant. |
| Claude 3.7 Sonnet | Retired February 19, 2026 | claude-sonnet-4-6 | Move to Sonnet 4.6 or newer and account for reasoning/tool behavior changes. |
| Claude Opus 3 | Retired January 5, 2026 | claude-opus-4-8 | Re-test high-capability workflows and cost assumptions. |
| Claude Opus 4.1 | Deprecated June 5, 2026; retired August 5, 2026 | claude-opus-4-8 | Already failing. Move to claude-opus-5 rather than the named 4.8 replacement and verify parameter compatibility. |
Fable 5 to Fable 5.1 breaking changes
The Fable 5.1 platform overview lists breaking changes relative to Fable 5 rather than calling it a drop-in swap. Forced tool use returns errors in cases where Fable 5 accepted it. Earlier models cannot read Fable 5.1 thinking blocks, which matters for multi-model pipelines that pass transcripts between generations. Editing an earlier turn invalidates thinking for that conversation.
The same overview lists additive betas that do not break existing calls: per-message effort, turn-scoped system messages, display: "updates" progress, lower cache reads, and content provenance. Cache reads on Fable 5.1 bill at $0.25 per MTok, so a Fable 5 cost model carried over unchanged will overstate spend on cache-heavy workloads.
Sonnet 5 migration notes
The official migration guide calls Sonnet 5 a drop-in upgrade from Sonnet 4.6 and calls out breaking API changes for manual extended thinking and non-default sampling parameters. On pricing, the $2/$10 per MTok rate originally announced as introductory through August 31, 2026 did not rise to $3/$15 on September 1; it is now the standard Sonnet 5 rate, which makes the move from Sonnet 4.6 at $3/$15 a price cut rather than a wash.
The same guide says Sonnet 5 uses a new tokenizer, so token counts for the same text can be higher. Roughly 30% more tokens against a 33% lower rate is close to neutral on paper, which makes measured pre-migration cost checks part of the migration rather than an afterthought.
Audit method for old codebases
Search for literal model IDs, not only visible model names. Then classify each ID as active, deprecated, retired, or partner-platform-specific. The deprecation page recommends using console usage exports to locate remaining deprecated-model usage.
For large engineering migrations, keep implementation practice out of this archive and use Claude Ships Code. This page is for historical and source-backed migration facts.
Parameter deprecations
The model deprecations page also records request-parameter deprecations. It says temperature, top_p, and top_k set to non-default values return a 400 error on Claude Opus 4.7 and later, including Opus 4.8, and on Claude Sonnet 5. That is a model-history detail with direct production consequences.
Questions this page answers
What replaced Claude 3.5 Sonnet?
The official deprecation ledger recommends Claude Sonnet 4.6 for the retired Claude 3.5 Sonnet model IDs.
What replaced Claude 2?
The official deprecation ledger recommends Claude Opus 4.8 for Claude 2.0 and Claude 2.1. That notice predates Opus 5, which launched July 24, 2026 and is the current Opus-class model.
What should developers check before moving to Sonnet 5?
Check tokenizer-driven cost changes, manual extended thinking, non-default sampling parameters, max output sizing, and refusal handling for cybersecurity-sensitive workloads.
Is Claude Fable 5.1 a drop-in replacement for Fable 5?
Not quite. The platform overview lists breaking changes: forced tool use can error, earlier models cannot read Fable 5.1 thinking blocks, and editing an earlier turn invalidates thinking.