Claude Fable 5.1 Launch: 2x GPT-5.6 on Science, 75% Cache Cut (2026)

▲ Claude Fable 5.1 launch, 2x GPT-5.6 score

Claude Fable 5.1 is the flagship AI model Anthropic released on September 1, 2026. For coders and researchers, it means a cheaper model that works longer unsupervised. It arrives 81 days after a US government order cut off its predecessor, posting 2.3x GPT-5.6 Sol's score on a science-research test and a 75% cut in cache-read prices.

Why Claude Fable 5.1 Counts as a Comeback

Fable 5 launched June 9. On June 12, a US government order on foreign nationals led Anthropic to cut off every customer. Controls lifted June 30; service returned July 1. OpenAI's GPT-5.6 Sol (July 9) and Google's Gemini 3.7 Flash (Aug 13) arrived cheaper. On Aug 31, Reuters reported a $35 billion Nvidia-backed Lambda compute deal in Texas.

▲ Fable 5.1 benchmarks: 52.6% vs GPT 22.4%

Fable 5.1 Benchmarks vs GPT-5.6 and Gemini

Anthropic's table shows the gap. On Terminal-Bench-Science, a science-research test, Fable 5.1 scored 52.6%, versus 22.4% for GPT-5.6 Sol and 24.7% for Fable 5. On Terminal-Bench 4.0 (agentic coding) it hit 55.8% vs GPT's 37.3%, and led CursorBench (73.4%), AutomationBench (31.4%), and GDPval-AA v2 (1,853 Elo). Gemini 3.7 Flash self-reports 30.4% on AutomationBench, just behind. At fintech Ramp it ran unattended for 38 hours, completing six experiments.

▲ 75% cache cut shifts the AI race to price

What a 75% Cache Price Cut Means for the AI Market

Base prices hold at $10 per million input tokens and $50 per million output tokens, but cache reads (re-loading a document it already saw) fall from $1 to $0.25. Agents (AI that runs multi-step work without a person) re-read files constantly, so Anthropic expects 25% savings on typical workloads and up to 45% on agent-heavy ones. Once scores diverge, the contest becomes who works cheaper and longer, with Texas compute deals, EU AI Act watermarking, and Enterprise Frontier Safeguards as the new selling points.

What to Watch Next

The next move is OpenAI's and Google's. GPT-5.6 Sol lists at $5 in and $30 out, roughly half Fable 5.1's rate; Gemini 3.7 Flash runs an introductory $0.75/$3.75 through year end. Whether rivals cut prices again or answer with performance is the second-half question for US AI.

Key Takeaways

① 81-day comeback - Fable 5.1 scores 52.6% on a science-research test, 2.3x GPT-5.6 Sol's 22.4%

② 75% cache price cut - reads fall from $1 to $0.25, cutting agent workloads up to 45%

③ New battleground - the race shifts from scores to price, compute, and compliance

AI report cards no longer end at benchmark scores. How cheaply and how long a model works without a human is the new number that matters.

👉 Claude Fable 5 Returns as US Lifts Export Controls, Sonnet 5 Launches - also worth a read.

---

📌 Sources: Anthropic, VentureBeat, TechCrunch, SiliconANGLE, Reuters, Google DeepMind, OpenAI (2026)