Anthropic's Claude Fable 5.1: agentic work gets up to 45% cheaper
On September 1, 2026, Anthropic announced Claude Fable 5.1 and Claude Mythos 5.1. The models beat their predecessor Fable 5, Opus 5 and OpenAI's GPT-5.6 Sol across multiple coding and knowledge-work benchmarks, while a 75% cut to cache-read pricing lowers the total cost of heavily agentic workloads by up to 45%.
Anthropic announced two models on the same day
Claude Fable 5.1 is the publicly available version, running under standard safeguards. Claude Mythos 5.1 is the same underlying model with relaxed safeguards: it opens to vetted cybersecurity and life sciences professionals through a new Life Sciences Verification Program, and is largely restricted to US companies and trusted-access participants. Anthropic described the pair as "the world's most advanced models for coding and knowledge work."
Benchmark results and the pricing change
On performance, Fable 5.1 scored 52.6% on Terminal-Bench-Science 0.1, a benchmark for agentic scientific research tasks, more than double the 24.7% scored by the prior version, Fable 5. The model reached 55.8% on Terminal-Bench 4.0, 73.4% on CursorBench 3.2.0, and 60.9% on Humanity's Last Exam without tools. Claude Code users should see roughly 60% fewer cybersecurity false positives.
Base input and output pricing stayed the same, but cache-read cost dropped to $0.25 per million tokens, a 75% cut. Anthropic said this lowers typical workload costs by around 25%, and by up to 45% for heavily agentic work, where agents repeatedly re-read the same codebase, system instructions, tool definitions and conversation history. As a separate layer, all Claude-generated text now carries an invisible watermark, with detection APIs rolling out in private preview for EU compliance. Both models are available immediately on Claude.ai, Amazon Web Services, Google Cloud, Microsoft Azure and the Claude API.
The real story is the pricing logic, not a benchmark
The real story here is not a single benchmark record but a shift in pricing logic. AI agents repeatedly re-read the same context as they work, the codebase, documentation, tool definitions, conversation history, so cutting cache-read cost by 75% targets exactly that repeated cost. It moves the "running agents is expensive" assumption: the same budget now runs longer, more multi-step agent work.
What it means for businesses in Türkiye
For businesses in Türkiye, token cost for long-running agents has been one of the most concrete barriers to experimenting with AI agents. This drop in cache pricing makes continuously running agent scenarios, customer service, code review, operations automation, more reachable on an SMB budget. But as the model gets cheaper, the real difference shows up in which business process, which data and which guardrails the agent is connected to; setup quality now matters more than model price.
The UNALSOFT take
Our reading is clear: as agent costs fall, it becomes more realistic for businesses to move AI from "pilot" into a system that actually runs in production. In our Agentic AI service, our job is not tracking which model is cheapest; it is wiring the agent into your real process (booking, orders, support) with the right boundaries. Model prices change; this design layer is what keeps carrying the work.
Models get cheaper. Setup quality is what decides.
Let's talk about what agent costs mean for your business and clarify which step fits you.