Anthropic shipped Claude Fable 5.2 and Claude Mythos 5.2 on Nov. 9. Each one takes on the hardest technical jobs a team can hand it, from codebase migrations to research proofs1. The same document named a third model, one the company kept to itself2.
Its system card, the safety and benchmark report Anthropic publishes beside every model, lists an "Internal Model" at 178.78 points on the company's own capability scale. Mythos 5.2, the model that shipped that day, scored 177.122.
Fable keeps safety filters for cybersecurity and biology, open to any user. Mythos drops some of those filters for vetted labs and researchers. Both run on one shared model underneath, the way earlier Fable and Mythos pairs did too1.
Anthropic put the comparison in plain words: "Unlike in prior system cards, we also report AECI scores for a separate, unreleased internal research model that is somewhat more capable than Claude Mythos 5.2," the company wrote2.
The admission set off a fight within hours. Andrew Curran tracks lab announcements and leaked benchmarks on X daily. He called it proof of a stronger system behind Anthropic's walls3.
Peter Gostev leads AI capability work at the model evaluator Arena, the company once known as LMArena. He read the same gap and called it small5.
What the Chart Actually Shows
AECI is Anthropic's version of the Epoch Capabilities Index, built by the research group Epoch AI2. Picture a single number standing in for dozens of tests: coding, science, robotics, cybersecurity. AECI folds every one of those scores into one figure, so a reader can watch one line climb instead of fifteen.
Anthropic's table runs back to June. Mythos 5 scored 168.68 that month. Opus 5 followed in July, at 169.40.
Mythos 5.1 reached 172.62 in September. Opus 5.5 hit 174.78 three weeks later2.
Nov. 9 added one more point. Mythos 5.2 posted 177.12 that day, about eleven points above where June's pace alone would have placed it2. The chart's last point sits higher still. Anthropic marked it with a hollow circle, apart from the filled circles standing for every model it sells.
Anthropic's AECI score rose 10 points since June, then a hidden model added 1.66 more
Anthropic's own capability index (AECI), by model, in order of release; higher is better
Each solid point is one model Anthropic has released, in order of release; the hollow point is the unreleased internal model named in its Nov. 9 system card.
Source: Anthropic [2]. Chart by Kelly Dennis.
The numbers behind this chart
| Period | Value |
|---|---|
| Mythos 5 | 168.68 |
| Opus 5 | 169.40 |
| Mythos 5.1 | 172.62 |
| Opus 5.5 | 174.78 |
| Mythos 5.2 | 177.12 |
| Internal model | 178.78 |
A Pattern From Past System Cards
Model 1 and Model 2 got names once already, in Anthropic's Sept. 1 card for Mythos 5.1. The company said both would stay inside Anthropic, kept for research alone6.
That card skipped the scores. This one prints one. The shift gives Curran and Gostev a number to argue over, where the September card gave only a name26.
Anthropic left its reason for the new disclosure unstated, in the card and in its launch post alike12.
The two cards together describe a company that tests ahead of what it sells, a step Anthropic frames as ordinary lab practice2.
Benchmarks like these run inside machine rooms, built for one job. They train and test a model before anyone past the lab sees it.
Is It Really Hidden Away
Andrew Curran read the chart first. "Anthropic's most capable model remains hidden deep within their tower. Source: Fable/Mythos 5.2 System Card," he posted on Nov. 9, hours after the system card went online3. His post reached 755 favorites within a day.
Gostev ran his own math on the same chart5.
Gostev wrote that on Nov. 6, answering a different, earlier report. METR is an independent outside evaluator. It had already found an Anthropic model it named Model 4. That model scored 178.6 on the same index, four days ahead of the system card4.
"It also doesn't seem to be the case that Anthropic has much more powerful models hidden away," Gostev wrote. "These 4.0 points are only c.7 weeks of progress"5.
The math lines up. Anthropic's own card puts its internal model's 178.78 exactly 4.0 points above Opus 5.5's 174.78, the same gap Gostev had already called modest2.
Anthropic's hidden model scores 178.78, just 4 points above Opus 5.5
AECI score: Opus 5.5, Mythos 5.2 and Anthropic's unreleased internal model, as of Nov. 9, 2026; higher is better
Each column is one model's AECI score; the gray column is Anthropic's prior flagship, and the magenta column is the model it keeps inside the company.
Source: Anthropic [2]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Opus 5.5 | 174.78 |
| Mythos 5.2 | 177.12 |
| Internal model | 178.78 |
Andrew Curran@AndrewCurran_Anthropic's most capable model remains hidden deep within their tower. Source: Fable/Mythos 5.2 System Card. pic.twitter.com/y7yBUpqGXY
Anthropic's own words called the model "somewhat more capable," a hedge from the company that built it2. METR's wider study found a faster pace too. Anthropic's models now gain ground on well-defined tasks. The pace now runs roughly twice its March 2024 to October 2025 speed4.
Model 4 cleared half of every benchmark METR tried against it. Earlier Anthropic models fell short on at least one test each4. The launch also reached Hacker News within hours, where readers picked through the same system card line by line7.
What to do
Read a capability chart for the step between two models. A 1.66-point rise over five months reads differently than a ten-point jump2.
Compare a new score against the model right before it, instead of the oldest one on the page. Opus 5.5 to the internal model is four points. Mythos 5 to the internal model is ten2.
Treat an unreleased score as a forecast. Anthropic built it for research alone, with a price, a release date and a safety review still pending2.
By the numbers
- 178.78: Anthropic's unreleased internal research model's score on the company's AECI scale, printed in its own Nov. 9 system card2.
- 177.12: Claude Mythos 5.2's score on that same scale, the model Anthropic shipped the same day2.
- 1.66 points: the smallest step up Anthropic's six-model table has shown since June2.
- 174.78: Claude Opus 5.5's AECI score, Anthropic's flagship before Mythos 5.22.
- 178.6: the score METR gave an Anthropic model it named Model 4, reported four days earlier4.
- 755: favorites on Andrew Curran's post naming the chart a hidden model, within a day3.
- Sept. 1: the date Anthropic's prior system card first named two internal models, Model 1 and Model 2, each left unscored6.
What to watch
Watch whether Anthropic ever names its internal model in public. A number missing a name invites exactly the kind of guessing Curran did on Nov. 923. METR may get the next turn. It caught this same climb early, and it could test the model directly if Anthropic grants access4.
Sources
- Anthropic, "Introducing Claude Fable 5.2 and Claude Mythos 5.2," Anthropic, Nov. 9, 2026, https://www.anthropic.com/claude-fable-and-mythos-5-2
- Anthropic, "System Card: Claude Fable 5.2 & Claude Mythos 5.2," Anthropic, Nov. 9, 2026, https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card
- Andrew Curran, X post, Nov. 9, 2026, https://x.com/AndrewCurran_/status/2119930955888746627
- Tom Cunningham, Nate Rush, Parker Whitfill, Hjalmar Wijk, "Independent Assessment of AI R&D Acceleration at Anthropic in October 2026," METR, Nov. 5, 2026, https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/
- Peter Gostev, X post, Nov. 6, 2026, https://x.com/petergostev/status/2118779677829677331
- Anthropic, "System Card: Claude Fable 5.1 & Claude Mythos 5.1," Anthropic, Sept. 1, 2026, https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32c215d14235/Claude%20Fable%205.1%20&%20Claude%20Mythos%205.1%20System%20Card.pdf
- "Claude Fable 5.2 and Mythos 5.2," Hacker News, Nov. 9, 2026, https://news.ycombinator.com/item?id=50429906
