Mythos-Class September 2026 Model Briefing

Claude Fable 5.1:
Anthropic’s Most Powerful AI, Explained

Claude Fable 5.1 is a new tier above Opus. It beat Pokemon from screenshots, remapped Venus, and cracked a five-year-old bug — then got switched off worldwide for 18 days.

Watch the 5-minute breakdown
01 — The Lineup

Claude Fable 5.1: A New Tier Above Opus

For the first time, the Claude family has something sitting above Opus. Anthropic calls it Mythos-class, and Claude Fable 5.1 is the first version the public can actually use.

HaikuFast and light
SonnetEveryday work
OpusDeep reasoning
Fable / MythosHours-long autonomous work — the new ceiling
02 — The Twist

One Model, Two Names

Claude Fable 5.1 and Mythos 5.1 are the same underlying model. Same brain. Anthropic is explicit that the safeguards — not the capability — are what separate them. If you want the full technical detail straight from the source, Anthropic’s official access announcement covers exactly how the two tiers are gated.

PUBLIC

Fable 5.1

  • Available to everyone
  • Extra safeguards on biology and cybersecurity
  • Flagged requests fall back to Opus
GATED

Mythos 5.1

  • Vetted organizations only
  • Cyber and Life Sciences programs
  • Looser safeguards for real defensive work
03 — The Specs

Claude Fable 5.1 Under the Hood

1M
Context window
128K
Max output tokens
5
Thinking effort levels
$10/$50
Per M tokens (in/out)

API model string: claude-fable-5-1. Knowledge cutoff: June 2026. This is built for jobs measured in hours, not seconds.

04 — The Backstory

It’s Had a Wild Year

Apr 2026
Mythos Preview goes to roughly 150 vetted organizations under Project Glasswing
Jun 9
Fable 5 and Mythos 5 launch publicly
Jun 12
U.S. export controls force a full worldwide suspension — access pulled almost overnight
Jun 30 – Jul 1
Controls lifted; access restored globally after 18 days
Sep 1
Fable 5.1 and Mythos 5.1 launch
05 — The Proof

Three Demos Worth Knowing

DEMO 01 — OFFICIAL

It beat Pokemon using only screenshots

No maps, no navigation aids, no custom helper scripts. Fable played through Pokemon FireRed reading raw game screenshots the way a person would. Earlier Claude models needed an elaborate support harness just to find their way around.

DEMO 02 — OFFICIAL

It built a sharper map of Venus

Fable 5.1 trained a neural network on 30-year-old NASA radar data and produced a higher-resolution elevation map covering about a third of the planet — detail down to 2–3 km, versus 10–20 km before. Anthropic released it free, ahead of NASA’s and ESA’s next Venus missions.

DEMO 03 — CUSTOMER STORY

It solved a bug that stumped engineers for five years

At Millennium, a one-in-a-million crash had gone unexplained for four to five years. Fable 5.1 disassembled a vendor library, matched it against the crash data, and traced the cause. At Ramp, one unattended 38-hour run caught a bad result, fixed it, launched six new experiments overnight, and reported back by morning.

06 — The Honest Part

Claude Fable 5.1: It’s Powerful, and It Isn’t for Everything

It’s expensive
Independent testing found it can cost more per task than Opus 5 for a small accuracy gain — largely because it writes more.
It’s verbose
It generates roughly 1.7× more output than its predecessor to reach the same answer.
Watermarking has limits
Its invisible text watermark signals Claude was likely involved. It doesn’t prove authorship, and it’s weak on short or code-heavy text.

So — When Should You Actually Use It?

SONNET / OPUS
Everyday chat, quick coding, and the vast majority of day-to-day work. Faster and cheaper.
FABLE 5.1
Hours-long autonomous jobs, deep research, and hard bugs nobody else can crack. That’s where the price tag earns itself.
07 — Keep Reading

More AI Model Breakdowns

Curious how Claude Fable 5.1 stacks up against the competition? Read our breakdown of GPT-6 Astra, OpenAI’s flagship model, covering its own computer-use, coding, and cybersecurity benchmarks.

Sources: Anthropic’s official Fable and Mythos access announcement, launch demos, and customer accounts from Millennium and Ramp; independent third-party cost testing.
Decoding Data Science · AI Model Briefings

Leave a Reply

Your email address will not be published. Required fields are marked *