Three months after Claude Fable 5 became Anthropic's flagship model, a quieter but more practically useful update arrived. On September 1, 2026, Anthropic released Claude Fable 5.1 and its restricted-access twin, Claude Mythos 5.1. There's no keynote-worthy new capability here — no jump to a Fable 6. What there is: a meaningfully cheaper model for anyone running real workloads, and a handful of benchmark jumps that matter more to people building agents and writing code than to people asking for essay help. If you're the kind of user who tracks your daily AI spend down to the cent, this is the update to actually read.
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, live immediately on the Claude API (as claude-fable-5-1), AWS, Google Cloud, and Microsoft Azure. Base API pricing is unchanged at $10 per million input tokens and $50 per million output tokens, but cache-read pricing dropped 75%, from the old rate down to $0.25 per million tokens. Anthropic says that cuts real-world workload costs by roughly 25% for typical use and up to about 45% for heavily agentic work. On benchmarks, Fable 5.1 scored 52.6% on Terminal-Bench-Science 0.1 versus Fable 5's 24.7%, and 55.8% on Terminal-Bench 4.0 versus 42.0% — plus MacRumors reports roughly 60% fewer cybersecurity false positives in Claude Code.
Fable 5.1 vs. Mythos 5.1: same brain, different leash
As with the original June launch, Fable and Mythos aren't two different models trained separately — they're the same underlying weights running behind different safeguard levels. Fable 5.1 is the version anyone with an API key or a chat subscription can use today. Mythos 5.1 has fewer guardrails around dual-use domains and is restricted to vetted researchers in cybersecurity and life sciences, admitted through verification programs Anthropic runs in coordination with US government partners. For the overwhelming majority of readers, Fable 5.1 is the relevant release — Mythos 5.1 is really a story about how Anthropic is handling frontier-capability access control, not a model most people will ever touch.
The actual news: cache reads got 75% cheaper
The headline number that didn't change is the one people usually watch — $10 per million input tokens and $50 per million output tokens, identical to Fable 5's list price. What changed is quieter and, for anyone running the model repeatedly against the same context, more consequential: cache-read pricing fell to $0.25 per million tokens, a 75% cut. Prompt caching is what lets a model reuse a large chunk of context — a long system prompt, a codebase, a document you're chatting against repeatedly — without paying full input-token price every single call. Cutting that price by three-quarters doesn't touch the sticker price of a single one-off request, but it drastically lowers the cost of the workloads that actually rack up bills: coding agents that re-read the same files across dozens of steps, chat sessions with long-running context, or any tool that leans on caching to stay fast and cheap. Anthropic's own estimate is a roughly 25% cost reduction for typical workloads, climbing to around 45% for the most agentic, cache-heavy tasks.
Why this matters more than a raw price cut would
A straight price cut helps everyone equally. A cache-read cut disproportionately rewards the users doing the most repetitive, context-heavy work — which happens to be exactly the direction the industry is moving. Agentic coding tools, research assistants that re-read long documents, and multi-step workflows all lean hard on caching by design. If your usage looks like quick one-off questions, you may barely notice this change on your bill. If your usage looks like an agent grinding through a codebase or a long research session with a big pinned context, this is the kind of change that shows up directly in what you pay per day.
The benchmark gains: coding and agentic science jumped, not everything else
Anthropic didn't claim a broad-based leap across every benchmark category, and the numbers back that framing up. The two callouts worth knowing: on Terminal-Bench-Science 0.1, which tests agentic scientific research workflows run through a terminal, Fable 5.1 more than doubled Fable 5's score, 52.6% versus 24.7%. On Terminal-Bench 4.0, a coding benchmark, Fable 5.1 scored 55.8% against Fable 5's 42.0%. Those are big, specific jumps in agentic and coding tasks — not a claim that Fable 5.1 is uniformly smarter at everything. MacRumors also reports a roughly 60% reduction in cybersecurity false positives for Claude Code users, plus new invisible text watermarking, both squarely aimed at developers rather than casual chat users. Anthropic frames Fable 5.1 as outperforming Fable 5, Claude Opus 5, and OpenAI's GPT-5.6 Sol across multiple benchmarks, and matching or beating Fable 5's results even at lower effort tiers — meaning you may not need to dial the model up to maximum reasoning effort to get results that used to require it.
Should you actually switch?
If you're mostly asking one-off questions — draft this email, explain this concept, summarize this page — Fable 5.1 will feel almost identical to Fable 5 day to day, since the sticker price for a single request hasn't moved. The switch is much more clearly worth it if any of the following describe how you use AI: you run coding agents or terminal-based tools that repeatedly reference the same files or context; you keep long-running chat sessions with a large pinned system prompt or document; or you're doing agentic, multi-step work where the model calls itself, checks its own output, and iterates. In those cases the 25-45% cost reduction Anthropic cites is a real, compounding saving, and the Terminal-Bench gains suggest you'll also get more correct output per dollar, not just cheaper output.
- Fable 5.1 and Mythos 5.1 launched September 1, 2026 — same model, different access levels (Mythos is restricted to vetted cybersecurity/life-sciences researchers).
- Base pricing unchanged: $10/M input, $50/M output tokens.
- Cache-read pricing cut 75%, to $0.25/M tokens — the actual cost story of this release.
- Anthropic estimates 25% lower cost for typical workloads, up to ~45% lower for heavily agentic tasks.
- Terminal-Bench-Science 0.1: 52.6% (Fable 5.1) vs. 24.7% (Fable 5).
- Terminal-Bench 4.0: 55.8% (Fable 5.1) vs. 42.0% (Fable 5).
- ~60% fewer Claude Code cybersecurity false positives, plus new invisible text watermarking, per MacRumors.
- Biggest winners: coding agents, research workflows, and anyone running long, cache-heavy sessions — casual one-off chat users will notice less.
01Is Claude Fable 5.1 a completely new model, or an update to Fable 5?
It's an update, not a new generation — Anthropic itself frames it as incremental. The underlying capabilities, pricing structure, and general behavior carry over from Fable 5, launched in June 2026; this release focuses on caching costs, agentic coding and science benchmarks, and security false-positive reduction rather than a ground-up rebuild.
02Will my API costs actually go down with Fable 5.1?
It depends on your usage pattern. Base input/output token pricing is unchanged, so simple one-off requests cost the same as with Fable 5. The savings come from the 75% cheaper cache-read pricing, so if your workload reuses context heavily — coding agents, long chat sessions, repeated document analysis — you should see a real reduction, potentially in the 25-45% range Anthropic cites.
03What is Claude Mythos 5.1, and can I use it?
Mythos 5.1 is the same underlying model as Fable 5.1 but with reduced safeguards, restricted to vetted individuals in cybersecurity and life-sciences research through verification programs run with US government partners. Most everyday users, developers, and students will only have access to Fable 5.1.
04How does Fable 5.1 compare to Fable 5 on coding tasks?
On Terminal-Bench 4.0, a coding-focused benchmark, Fable 5.1 scored 55.8% compared to Fable 5's 42.0%. Anthropic also says Fable 5.1 can match or beat Fable 5's results even when run at lower reasoning-effort tiers, which can itself translate into lower cost per task.
05Does this replace the need to compare Fable 5.1 against other models like GPT or Gemini?
No — Anthropic's own claim that Fable 5.1 beats GPT-5.6 Sol on multiple benchmarks is worth treating as one data point among several, since benchmark performance varies by task type and different labs optimize for different things. The practical move is still to test the models you actually use against your own workload rather than relying on any single vendor's benchmark table.
Updates like this are exactly why comparing models by sticker price alone misses half the picture — a model with the same headline rate can still cost meaningfully less (or more) depending on how you use it. That's part of the reasoning behind LumiChats' pay-per-day model: instead of locking into one provider's pricing quirks, you get Claude, GPT, Gemini, DeepSeek, and more under a single login, for less than $1 a day, so you can just switch to whichever model is actually cheaper or sharper for the task in front of you that day.
