Anthropic shipped two new models on Nov. 9: Claude Fable 5.2 and Claude Mythos 5.21. Fable handles general coding and research work. Mythos stays limited to Anthropic's approved cybersecurity and life-sciences partners.
Three days later, two outside trackers say the newer model keeps winning. Arena and LiveNerf both run apart from Anthropic, and both point the same way45. One gives Fable a net improvement score of 30.4 percent against Opus's 19.5 percent. The other has scored Fable ahead by double digits every day it has measured.
Two Trackers, One Gap
LiveNerf freezes 78 questions and asks each model the same questions every day. A reader can watch a score move over many days, past any single launch-day chart5. The set spans math competitions, a graduate-level science test and a broad knowledge exam. Fable scored 80.8 percent on launch day, climbed to 88.5 percent on Nov. 10, and settled at 84.6 percent on Nov. 11.
Opus answered the same questions, scoring 61.5, 57.7 and 65.4 percent on Nov. 9, 10 and 11, at least 19 points behind Fable5. The gap held steady even as both scores moved up and down. A chart below traces both lines side by side, day by day.
Fable 5.2 Beats Opus 5.5 by 19 to 31 Points a Day
LiveNerf's frozen 78-question panel, percent scored correct, Nov. 9 to 11, 2026; higher is better
Each line is one model's daily score on the same 78 questions; the magenta line stays at least 19 points above the blue line across LiveNerf's first three days of tracking Fable 5.2.
Source: GitHub (ninjahawk/livenerf) [5]. Chart by Kelly Dennis.
The numbers behind this chart
| Period | Claude Fable 5.2 | Claude Opus 5.5 |
|---|---|---|
| Nov. 9 | 80.8% | 61.5% |
| Nov. 10 | 88.5% | 57.7% |
| Nov. 11 | 84.6% | 65.4% |
Arena works differently. It ranks models on real coding and research sessions people run through them, apart from any fixed set of questions4. On Nov. 11 it gave Fable the top spot among agents, with a net improvement score of 30.4 percent, against 19.5 percent for Opus.
Arena.ai@arenaBig news: Claude Fable 5.2 (Max) by @AnthropicAI takes #1 in Agent Arena with an unprecedented lead: +30.4% net improvement, about 1.6x the #2 model Opus-5.5, across 4.7k real-world agentic sessions from our global community. It's also #1 in Code, Chat and Work.
A major leap… https://t.co/auTTxqACSu pic.twitter.com/vIFfNyAInZ
What Anthropic Built
Anthropic priced the new models below the ones they replace. Fable and Mythos both cost $8 per million input tokens and $40 per million output tokens3. That is a fifth less than the prior generation's $10 and $504. Anthropic says the pair also writes and sends text faster than before, up to 30 percent faster by its own account2.
Joe Haddad, an engineer at Vercel, tested the new model on real projects before the launch1. "Projects we've been pushing along with other models for weeks got finished and approved by our teammates in a few days with Claude Fable 5.2, and the humans reviewing its code had less to say," Haddad said.
Tom Pritchard, a staff engineer at Shopify, found the same thing on looser, open-ended work1. "Its answers are really well-considered, and within a few turns I can almost always say 'ok, go do that,' and trust that it'll complete it with little intervention," Pritchard said.
Matt Lee, a senior staff engineer at Datadog, pointed to a harder case: an unfamiliar system1. "You can just throw Claude Fable 5.2 at new, ambiguous systems it has never seen before, and it will figure out how to actually work within them," Lee said.
The Developer Who Still Opens Opus
Theo Browne runs the coding tool T3 Code and tests nearly every model Anthropic ships. His show reaches a large audience of working developers who decide what to install next. He spent 14 hours with the new release on his own codebase the day after launch6.
"Something I've gotten out of this release is a deeper appreciation for Opus 55, not because it is smarter, but because of the relationship between these models," Browne wrote6. On his own repository, the new model produced four false claims out of 61 tasks, at a cost near $23. Sonnet, Anthropic's cheaper model, made 0 errors on the same tasks for $4.19.
Browne Counted 4 Errors in Fable 5.2, 0 in Sonnet 5.5
False claims found in Theo Browne's own codebase test of 61 tasks, Nov. 10, 2026; lower is better
Each column counts the wrong answers Browne found when he ran both models on his own repository; Fable 5.2's four errors cost about $23, Sonnet 5.5's 0 errors cost $4.19.
Browne ran this test on his own repository alone, a sample of one developer's work, apart from a controlled benchmark.
Source: BigGo Finance [6]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Claude Fable 5.2 | 4 |
| Claude Sonnet 5.5 | 0 |
Browne still opens Opus first for daily coding work, despite the newer model's higher scores on every public chart6. His test measured something trackers skip: how often a reviewer has to catch the model's mistake.
Two Readings
Both readings can hold at once. LiveNerf and Arena measure a fixed battery of questions, repeated and scored the same way45. The newer model wins that contest every day it has run. Browne's test measures something else: a working codebase, judged by whether a reviewer has to step in.
A scoreboard rewards the model that gets the most right answers. Browne's own week rewards the model that needs the least watching. The newer model leads the first measure by a wide margin. Browne's count says the second one still favors the model it replaced, on his own repository alone6.
Trust the scoreboard for what it measures. Test the model yourself before it becomes your default.
What to do
Run your own test before you switch your default model. Pick a real task from your own repeated work, the kind a demo skips. Compare both models on it the way Browne tested his own repository.
Watch a tracker that scores the same questions every day, apart from a single launch-day chart. A model that holds its lead for a week has earned more trust than one that beat a static table once.
By the numbers
- 84.6 percent: Fable 5.2's score on LiveNerf's 78-question panel on Nov. 11, two days after launch5.
- Opus 5.5 scored 65.4 percent on the same panel that day5.
- Arena's agent ranking gives Fable 5.2 a net improvement score of 30.4 percent, against 19.5 percent for Opus 5.54.
- $8 and $40: Fable 5.2's price per million input and output tokens, a fifth below Fable 5.1's $10 and $503.
- Browne counted four false claims in 61 of Fable 5.2's answers on his own codebase test6.
- Sonnet 5.5 made 0 errors on the same 61 tasks, for $4.19 against Fable 5.2's near $236.
What to watch
Watch whether Fable's lead on LiveNerf and Arena holds through a full month of tracking. Opus's own score has already drifted a few points either way, so one strong week decides little on its own. Browne's error count is worth watching too: Anthropic patches models after launch, and a falling count would answer his complaint directly.
Sources
- Anthropic, "Introducing Claude Fable 5.2 and Claude Mythos 5.2," Nov. 9, 2026, https://www.anthropic.com/claude-fable-and-mythos-5-2
- Carl Franzen, "Anthropic releases Claude Fable 5.2 and Mythos 5.2, pairing new, record-high benchmarks with 20% API cost reduction," VentureBeat, Nov. 9, 2026, https://venturebeat.com/technology/anthropic-releases-claude-fable-5-2-and-mythos-5-2-pairing-new-record-high-benchmarks-with-20-api-cost-reduction
- Data Studios, "Claude Fable 5.2 launches at $8/$40 per 1M tokens, 20% below Fable 5.1," Nov. 9, 2026, https://www.datastudios.org/post/anthropic-claude-fable-5-2-mythos-5-2-launch-pricing-plans
- Arena.ai (@arena), post on X, Nov. 11, 2026, https://x.com/arena/status/2120568144687460434
- ninjahawk, "LiveNerf," GitHub, accessed Nov. 12, 2026, https://github.com/ninjahawk/livenerf
- BigGo Editorial Team, "Anthropic's Fable 5.2 Wins the Benchmarks. Theo Browne Still Reaches for Opus 5.5.," BigGo Finance, Nov. 10, 2026, https://finance.biggo.com/news/a090bf89b921a843
