Claude Fable 5.1 Holds LiveBench Lead at 83.4 as Opus 5.5 Thinking Debuts at #2
Claude Fable 5.1 remains at the top of the September 2026 LiveBench run with 83.4 overall, despite the arrival of Claude Opus 5.5 Thinking four days ago. The newer model debuted directly at #2 with 83.2.
Where Each Model Wins
| Category | Fable 5.1 | Opus 5.5 Thinking |
|---|---|---|
| Overall | 83.4 | 83.2 |
| Reasoning | 91.7 | 92.2 |
| Coding | 86.4 | 89.3 |
| Agentic Coding | 66.1 | 71.7 |
| Math | 97.0 | 97.1 |
| Data Analysis | 80.3 | 80.3 |
| Language | 89.5 | 86.3 |
| Instruction Following | 73.0 | 65.7 |
| Cost per task | $1.212 | $0.799 |
Opus 5.5 Thinking leads on four of seven task categories — reasoning, coding, agentic coding, and math. Data analysis is a tie at 80.3. Its agentic coding score of 71.7 is the strongest among closed frontier models on this run, outpacing Fable 5.1 by 5.6 points. Its weakness is instruction following at 65.7 — 7.3 points behind Fable 5.1. That deficit on a category Fable 5.1 wins by a wide margin is what keeps the newer model in second place overall.
Extended-thinking models routinely underperform on strict instruction following. The internal reasoning chain can override precise format constraints, a trade-off that shows up predictably on LiveBench’s instruction following tasks.
The Cost Gap
Fable 5.1 at $1.212 per successful task is 52% more expensive than Opus 5.5 Thinking at $0.799. For the same overall score tier, the economics favor Opus 5.5 Thinking in production.
Claude Fable 5 (original) costs $1.439 per task at 83.0 overall — more expensive than either and now third behind its own update.
Full Top-6 Leaderboard
| Rank | Model | Overall | Cost/Task |
|---|---|---|---|
| 1 | Claude Fable 5.1 Max Effort | 83.4 | $1.212 |
| 2 | Claude Opus 5.5 Thinking Max Effort | 83.2 | $0.799 |
| 3 | Claude Fable 5 Max Effort | 83.0 | $1.439 |
| 4 | GPT-6 Astra Max Effort | 82.2 | $0.736 |
| 5 | Muse Spark 1.3 xHigh Effort | 81.6 | $0.219 |
| 6 | DeepSeek V4.1 Flash Max Effort (open) | 81.1 | $0.029 |
Muse Spark 1.3 at $0.219 per task is the best-value model in the top five. DeepSeek V4.1 Flash, open-weight, sits at 81.1 with the lowest per-task cost on the board. Its agentic coding score of 77.3 is the highest of any model in the top six — open or closed.
Opus 5.5 Thinking’s debut at #2 within four days of release is consistent with how frontier thinking-model launches have performed on LiveBench since early 2026. High-ceiling reasoning benchmarks reward extended compute. Whether the instruction-following deficit narrows in later runs will determine if Opus 5.5 Thinking eventually overtakes Fable 5.1.