Anthropic just quietly created a brand-new AI tier. Meet Claude Fable 5.1 and Mythos 5.1, launched on September 1, 2026, sitting above Opus, and yes, it’s the strongest Claude has ever been.
The lineup now runs Haiku, then Sonnet, then Opus, then Mythos-class at the top, a tier Anthropic first introduced back in June 2026 with Fable 5 and Mythos 5. Fable 5.1 and Mythos 5.1 are literally the same underlying model, just wrapped in different safeguards. Anthropic is calling the pair “the world’s most advanced models for coding and knowledge work.”
Why Two Names for One Model
Fable comes with extra safety layers around biology, cybersecurity, and AI research, which is what makes it safe enough for general public use. Mythos strips some of those layers back and stays locked to vetted partners only, mainly cybersecurity and life-sciences organizations that genuinely need fewer restrictions to do their work. Because the safeguards can intervene differently, the two versions occasionally score differently on the same benchmark, even though the model underneath is identical.
A Bug Five Years in the Making
Fable 5.1’s most striking demo came from Millennium, an investment firm. A crash hit roughly once in a million runs, and nobody on their team, nor any model they tried, including Fable 5, could explain it for four to five years. Fable 5.1 disassembled an external vendor library, matched it against the core dump from the crash, and traced the bug to that library, something no engineer or prior model had managed to do.
That bug hunt wasn’t a one-off. One reported run had Fable 5.1 working unattended for 38 hours on a machine-learning problem, where it caught an experimental error, corrected it, and launched six new experiments on its own before reporting back with results. On a hard browser-agent benchmark, it completed 82% of tasks in roughly ten minutes each, compared to 74% for Opus 5 and 57% for Fable 5, all while using fewer tokens than either.
Read also: Evergreen.ai Launches: AI-Powered Financial Planning, On Demand
Doubling Up on the Hard Stuff
The benchmark gains are just as notable on paper. Fable 5.1 scored 52.6% on Terminal-Bench-Science versus Fable 5’s 24.7%, and it now leads Opus 5 across every benchmark category Anthropic published, including a narrower but real lead on Terminal-Bench 4.0 and knowledge-work evaluations like GDPval-AA v2.
Pricing stayed the same at $10 per million input tokens and $50 per million output tokens, so the sticker price hasn’t moved. What did move is cache-read pricing, which dropped 75%, from $1.00 down to $0.25 per million tokens.
Anthropic says that translates to roughly 25% lower effective costs for typical workloads and up to 45% for highly agentic ones that lean heavily on cached context, since the discount only kicks in where caching is actually used.
Fewer False Alarms
Safety filters also got more precise with this release. Cybersecurity safeguards now trigger about 60% fewer false positives per Claude Code session, and Fable 5.1 can now assist with finding software vulnerabilities for defensive purposes, though exploit development and penetration testing still get routed to Opus models.
Biology safeguards fire 85% less often on ordinary medical questions, cutting down on the frustrating false flags that used to interrupt legitimate use.
Read also: Google DeepMind Lyria 3.5: Full-Length AI Music Generation Revealed
Not a Universal Upgrade
Independent testing has found Fable 5.1 can cost more per task than Opus 5 for only a small accuracy gain, largely because it tends to write more to reach an answer, generating roughly 1.7 times more output than its predecessor. That makes it best suited for hours-long autonomous jobs, deep research, and hard problems nobody else can crack, rather than everyday chat or quick coding tasks where a faster, cheaper model still does the job.
