OP 03 September, 2026 - 05:20 PM
(This post was last modified: 03 September, 2026 - 05:20 PM by JVZU.)
Claude Fable 5.1 is out — what actually changed, and why it matters if you run Claude Code
Anthropic shipped Fable 5.1 on September 1st, three months after Fable 5. No new sticker price, no rebrand — but under the hood it's the biggest jump in the 5-series so far, and the pricing change nobody noticed is the real story.
What is Fable? It's Anthropic's top-end line above Opus — tuned for long agentic runs: terminal work, research, multi-hour coding sessions. 1M token context, 128K max output, adaptive thinking always on.
The benchmarks (Anthropic's numbers, independently discussed by Vellum and others):
Fable 5.1 takes first place in every single row. The Terminal-Bench-Science number is the headline: doubled vs Fable 5, nearly doubled vs Opus 5. This is the "give it a shell and let it work" benchmark — exactly the Claude Code use case. Artificial Analysis has it leading their Intelligence Index too: 66 vs Opus 5's 63 and GPT-5.6 Sol's 61.
The pricing change that actually matters: input/output stayed at $10/$50 per M, but cache reads dropped 75% — from $1.00 to $0.25/M. If you've used Claude Code you know why this is huge: on a long session, cache reads are 70–80% of your tokens. Anthropic estimates ~25% cheaper typical workloads, up to ~45% on context-heavy agentic ones. (Fair warning: Artificial Analysis notes that at max effort it burns ~1.7x the output tokens of Fable 5, so lazy comparisons cut both ways.)
Gotchas if you're migrating from Fable 5:
- tool_choice "any" / "tool" now returns 400 — forced tool use is gone, use "auto" + strict tools
- Thinking blocks are model-bound: 5.1 reads older models' thinking, not vice versa
- It narrates less and prefers whole-file rewrites over targeted edits — adjust your prompts
- Ships with built-in output watermarking (C2PA on files)
Where to try it without a $200 subscription: we run Fable 5.1 (claude-fable-5-1) on the real Anthropic Messages API — native prefill, real token counting, full streaming, no chat-completions translation layer in the middle (test any "cheap Claude" provider with an assistant-prefill request and watch half of them fall over). We also just cut all our usage rates by 33%. Setup for Claude Code / Cursor / anything that speaks Anthropic: tokies.cc — docs: tokies.cc/docs
Happy to answer questions about the breaking changes — we've been running 5.1 traffic since launch day.
![[Image: signature.gif]](https://external-content.duckduckgo.com/iu/?u=https%3A%2F%2Fi.postimg.cc%2Fbrxd4TZr%2Fsignature.gif)
Anthropic shipped Fable 5.1 on September 1st, three months after Fable 5. No new sticker price, no rebrand — but under the hood it's the biggest jump in the 5-series so far, and the pricing change nobody noticed is the real story.
What is Fable? It's Anthropic's top-end line above Opus — tuned for long agentic runs: terminal work, research, multi-hour coding sessions. 1M token context, 128K max output, adaptive thinking always on.
The benchmarks (Anthropic's numbers, independently discussed by Vellum and others):
Code:
+------------------------------+-----------+--------+---------+----------+
| Benchmark | Fable 5.1 | Opus 5 | Fable 5 | GPT-5.6 |
+------------------------------+-----------+--------+---------+----------+
| Terminal-Bench-Science | 52.6% | 29.0% | 24.7% | 22.4% |
| Terminal-Bench 4.0 | 55.8% | 52.3% | 42.0% | 37.3% |
| CursorBench 3.2.0 | 73.4% | 70.0% | 70.5% | 67.2% |
| Humanity's Last Exam (tools) | 65.0% | 63.6% | 63.8% | - |
| AutomationBench | 31.4% | 26.9% | 17.1% | 19.6% |
| OSWorld 2.0 (strict) | 41.7% | 39.6% | 36.1% | - |
| GDPval-AA v2 | 1853 | 1824 | 1723 | 1711 |
+------------------------------+-----------+--------+---------+----------+Fable 5.1 takes first place in every single row. The Terminal-Bench-Science number is the headline: doubled vs Fable 5, nearly doubled vs Opus 5. This is the "give it a shell and let it work" benchmark — exactly the Claude Code use case. Artificial Analysis has it leading their Intelligence Index too: 66 vs Opus 5's 63 and GPT-5.6 Sol's 61.
The pricing change that actually matters: input/output stayed at $10/$50 per M, but cache reads dropped 75% — from $1.00 to $0.25/M. If you've used Claude Code you know why this is huge: on a long session, cache reads are 70–80% of your tokens. Anthropic estimates ~25% cheaper typical workloads, up to ~45% on context-heavy agentic ones. (Fair warning: Artificial Analysis notes that at max effort it burns ~1.7x the output tokens of Fable 5, so lazy comparisons cut both ways.)
Gotchas if you're migrating from Fable 5:
- tool_choice "any" / "tool" now returns 400 — forced tool use is gone, use "auto" + strict tools
- Thinking blocks are model-bound: 5.1 reads older models' thinking, not vice versa
- It narrates less and prefers whole-file rewrites over targeted edits — adjust your prompts
- Ships with built-in output watermarking (C2PA on files)
Where to try it without a $200 subscription: we run Fable 5.1 (claude-fable-5-1) on the real Anthropic Messages API — native prefill, real token counting, full streaming, no chat-completions translation layer in the middle (test any "cheap Claude" provider with an assistant-prefill request and watch half of them fall over). We also just cut all our usage rates by 33%. Setup for Claude Code / Cursor / anything that speaks Anthropic: tokies.cc — docs: tokies.cc/docs
Happy to answer questions about the breaking changes — we've been running 5.1 traffic since launch day.