Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1 on September 1, its most capable coding and knowledge work models yet, both running a 1M token context window and 128k max output tokens as the standard, not an opt in
Fable 5.1 is generally available and Anthropic says it beats Fable 5 by a wide margin at a much lower cost per task, with Claude Code users seeing roughly 60 percent fewer cybersecurity false positives
Mythos 5.1 is the same underlying model with different safeguards, limited to trusted access programs for cybersecurity and life sciences work, and can now flag software vulnerabilities without being allowed to write exploits for them
I run Claude Code daily to build RAXXO tools, so I read past the launch language to find what actually changes this week versus what is just a bigger number on a spec sheet
What Anthropic Actually Shipped on September 1
Anthropic released two models on the same day, not one. Claude Fable 5.1 is generally available across the Claude Developer Platform, the consumer apps, and the major cloud marketplaces. Claude Mythos 5.1 is described as the same model underneath, but gated behind trusted access programs, with safeguards tuned specifically for cybersecurity and life sciences work rather than general use. That split matters more than it looks. Anthropic is not building two different brains, it is building one and then deciding who gets which set of guardrails depending on what they are allowed to do with it.
Both models carry a 1M token context window and 128k max output tokens as the baseline, with no smaller variant to fall back to. Adaptive thinking is on by default rather than something a developer has to switch on for harder problems, which means the model decides how much reasoning a request needs instead of always running at one fixed setting. Anthropic also lowered cache read pricing with this release, which quietly matters more than the headline numbers for anyone running long agent sessions, since cache reads are what most repeated tool calling actually costs.
The beta controls are the part most likely to get skipped in the coverage, so it is worth naming them directly. Fable 5.1 and Mythos 5.1 ship with new controls for tool use behavior, preserved thinking across turns, per-message effort levels, and turn-scoped system messages that apply to a single exchange instead of an entire session. For anyone building multi-step agent tools rather than just chatting with a model, that last one is the quiet upgrade. It means a system prompt can change mid-session without carrying old instructions forward by accident, which is exactly the kind of bug that used to eat an afternoon of debugging before this existed.
None of this changes what I already wrote about invisible watermarking on newer Claude models. That policy already covers anything launched after August 2, 2026, so Fable 5.1 and Mythos 5.1 inherit it rather than adding anything new to explain. Worth noting anyway, since it means anything either model writes for me, a draft, a script comment, a support reply, carries the same invisible mark I already accepted as a fact of building on Claude.
The Coding Numbers That Matter for Claude Code
Anthropic frames Fable 5.1 as a step up from Fable 5 in coding, knowledge work, and long-running problem solving, and says it reaches similar or better results at a noticeably lower cost per task. That is the same shape of claim Anthropic made when Opus 5 shipped in July, better output per unit of spend rather than a flashier benchmark chart, and it is the claim that actually matters if you run Claude Code for hours at a stretch instead of firing off a handful of prompts a day. Two releases in two months both leaning on the same argument, cheaper per task rather than just smarter on paper, tells me more about where Anthropic thinks the real bottleneck is than any single spec sheet does.
The number I paid closest attention to is the roughly 60 percent drop in cybersecurity false positives inside Claude Code. A false positive in a security scan is not a neutral cost, it is time spent chasing a warning that was never real, and it trains you to skim past the next ten warnings that might be. Fewer false positives means the warnings that do show up carry more weight, which is the difference between a tool you trust and a tool you tune out.
Fable 5.1 can also now discover software vulnerabilities in code, a capability Anthropic pairs with an explicit restriction: it is not permitted to develop exploits for what it finds. That is a narrow, deliberate line, find the crack, do not hand anyone the key. For a one person studio shipping several small tools instead of one large product, a model that can point at a real vulnerability without me having to specialize in security research first is a genuine capability upgrade, not a marketing line. It does not replace judgment about what to fix first, but it changes what gets found in the first place.
That restriction is also a useful way to read every capability Anthropic ships going forward. The interesting question is rarely whether a model can do something, it is where the line sits on what it is allowed to do with that ability once it has it. Vulnerability discovery without exploit development is a clean example of drawing that line in a way that is still genuinely useful to a developer instead of just being useful in theory.
Why Mythos Exists Separately From Fable
Mythos already had its own track record before this release. The earlier Mythos model was the one that found real structural weaknesses in two cipher designs, work that only made sense with looser restrictions than a consumer facing model should carry by default. Mythos 5.1 continues that pattern rather than starting it. Same underlying capability as Fable 5.1, different safeguards, and access limited to trusted programs instead of a general signup.
That structure answers a question I used to have about why Anthropic would maintain two names for what is functionally one model family. The answer is accountability. A model tuned to help find vulnerabilities or work through sensitive life sciences questions needs a different set of checks around who can use it and how, compared to a model helping someone write a landing page or debug a build script. Keeping that split visible as two named models, rather than one model with a hidden permissions flag, makes it obvious to anyone reading the announcement which tier they are actually getting.
For most people reading this, Mythos 5.1 will stay a name on a page rather than a model they ever open, and that is by design. What it tells you as a Fable 5.1 user is that the safety work behind your everyday model is being tested somewhere harder first, on cipher designs and vulnerability research, before the improvements show up in the version everyone can sign up for.
What Changes for a Small Studio Running Claude Daily
I do not run six figure infrastructure or a team pulling shifts on model evaluation. I run Claude Code most days to build and maintain the tools RAXXO sells, and this release changes a few concrete things about that routine. Lower cache pricing means a long session spent iterating on one file costs less than the same session did last week, which is the kind of change that never gets a headline but shows up every time I actually work.
Turn-scoped system messages solve a specific annoyance I have hit building tools that hand off between agents, where an instruction meant for one step of a task quietly kept influencing steps it was never meant for. Being able to scope a system message to one exchange instead of the whole conversation removes a class of bug I used to work around by starting a fresh session more often than I wanted to.
Adaptive thinking being on by default rather than a manual toggle also matters for a solo workflow, because it means I no longer have to guess upfront how hard a task will be before deciding how much reasoning budget to give it. The model makes that call per request. Before this, I would sometimes flip a task to a higher effort setting only after watching it stumble on something simpler, which cost a redo I did not need to pay for.
None of this is a reason to rebuild anything I already shipped. It is a reason the next tool I build starts from a slightly better baseline than the last one did, which is the actual pattern this whole blog keeps returning to release after release. The models keep getting cheaper to run for longer, the false positives keep dropping, and the actual day to day work of building small tools alone keeps getting a little less friction heavy. That is a slower story than any single release headline, but it is the one that has held up every time I check it against what changed in my own week.
Bottom Line
Two models in one release, one generally available and one gated for higher stakes work, both running the same 1M token context and both cheaper to lean on for long sessions than what came before. The headline number worth remembering is the roughly 60 percent drop in Claude Code's cybersecurity false positives, because it is the one change that shows up in daily use rather than a benchmark chart. Mythos 5.1 staying restricted to trusted programs is the right call, and it is a reminder that the safety work behind the model I actually use gets tested somewhere harder first. I will keep running Claude Code the same way I did last week, just with fewer warnings to second guess and a little more room in a long session before the cost adds up.
Top comments (0)