The Diff

Archives
Log in
Subscribe
November 10, 2026

No. 038: Anthropic says Fable 5.2 tops its coding benchmark - and still recommends Opus 5.5. Meta says Muse Ember matches Haiku 5.5’s best score at a third of the cost.

Anthropic released Claude Fable 5.2 and Mythos 5.2 on Nov 9 - same model, two doors - at $8/$40 per million tokens, and says Opus 5.5 is still the daily driver. Plus: Meta’s smallest model, Muse Ember, at $0.10/$0.50. Plus the quiet ones nobody talks about.
View in browser: The Diff No. 038
The Diff  No. 038
Tuesday, 10 November 2026  ·  AI dev news, only what is verified
Anthropic announcement card for Claude Fable 5.2 and Mythos 5.2
Anthropic released Claude Fable 5.2 and Mythos 5.2 on Nov 9 - same model, two doors - at $8/$40 per million tokens, and says Opus 5.5 is still the daily driver. Plus: Meta’s smallest model, Muse Ember, at $0.10/$0.50.

Anthropic says Fable 5.2 reaches 49.1% on FrontierCode 2.0 at Xhigh effort, "the highest score of any released model we’ve tested" - its own benchmark, its own wording. Artificial Analysis, an outside index, puts the model at 65 and first of 74, and also at the most expensive per task and among the slowest.

Also today: Meta says its new Muse Ember matches Claude Haiku 5.5’s best score at about a third of the cost - and its own table shows Ember behind Haiku on the coding eval.

A top-of-the-chart model you probably should not start with, and a cheap one built for subagents.

IN THIS ISSUE
🚀 Anthropic says Fable 5.2 tops its coding benchmark - and tells most people to start with Opus 5.5 anyway
01
🚀 Anthropic says Fable 5.2 tops its coding benchmark - and tells most people to start with Opus 5.5 anyway
Anthropic, Nov 9: Claude Fable 5.2 and Mythos 5.2 are one model; Fable is open to all users at $8/$40 per million tokens, Mythos goes through application programs. Artificial Analysis ranks it first on intelligence and last on cost.
02
🔥 Meta says Muse Ember matches Haiku 5.5’s best score at a third of the cost - its coding score is lower
Meta, Nov 9: Muse Ember is its smallest model, $0.10 per million input tokens and $0.50 per million output tokens on the Meta Model API. Meta’s table: same Artificial Analysis Index as Haiku 5.5 (36), lower on FrontierCode 2.0-AA (27.2 vs 34.9).
🔥 VIRAL
The loud ones on Hacker News. Numbers are a morning snapshot.
Anthropic announcement card for Claude Fable 5.2 and Mythos 5.2
01  Anthropic + Artificial Analysis, Nov 9-10 · HN: 314 points at 06:27 IST

🚀 Anthropic says Fable 5.2 tops its coding benchmark - and tells most people to start with Opus 5.5 anyway

Anthropic, Nov 9: Claude Fable 5.2 and Mythos 5.2 are one model; Fable is open to all users at $8/$40 per million tokens, Mythos goes through application programs. Artificial Analysis ranks it first on intelligence and last on cost.
SCENE

Anthropic’s own account of its flagship demo: one engineer gave Fable 5.2 the task of designing a prototype operating system, starting from an empty folder. The model coordinated over 1,000 subagents, working around the clock for eight days, and produced a system Anthropic says boots on a laptop, with drivers, file storage, windows, a login screen and an installer. Their experiment, their description.

THE THING

What Anthropic published: Claude Fable and Mythos are "the same underlying model." Fable carries cybersecurity and biology safeguards and is available to all users; organizations in cyber or biomedical work can apply for Mythos through its verification programs. Both cost $8 per million input tokens and $40 per million output tokens (lower than 5.1, Anthropic says), with an optional fast mode at $16/$80 that it says is up to 2.5 times faster. In Anthropic’s own table, Fable 5.2 scores 49.1% on FrontierCode 2.0 at Xhigh effort against 45.5% for Opus 5.5, and at High effort roughly matches Opus 5.5 (45.3% vs 45.5%) for what Anthropic calls roughly half the cost per task. Anthropic’s own guidance: "For most coding and knowledge tasks, we recommend that users start with Claude Opus 5.5, our strongest daily driver," moving up to Fable or Mythos only for "the most ambitious, open-ended technical work." The outside check is Artificial Analysis: Intelligence Index 65, ranked #1 of 74 models, but also #73 of 73 on cost at $26.54 per Intelligence Index task, 46.4 minutes per task against a 12.6-minute median, and 275M output tokens against a 72.8M median - its words: "particularly expensive" and "very verbose." That is one run of one index; it is AA’s measurement, with Fable at Max effort.

✂️ STEAL IT

Steal the two-number habit: read the vendor’s best score and the independent cost-per-task side by side. Anthropic’s headline is the 49.1%; the number that decides your bill is AA’s $26.54 per task. If you run agents in loops, route only your hardest tasks to Fable as advisor or orchestrator - roles Anthropic itself describes - and leave the rest on a cheaper tier.

Watch  VERDICT

Skim Anthropic’s benchmark section and the AA page, and rerun one real task of yours on Opus 5.5 and Fable 5.2 before you switch anything. It fails you on latency: AA measured about 216 seconds to first answer token at Max effort, and the "ambitious work" claims (the hypervisor, the Fermat formalization) are Anthropic’s own, not independently verified.

Anthropic: Claude Fable 5.2 and Mythos 5.2  ·  Artificial Analysis: Fable 5.2  ·  HN: the launch  ·  HN: the AA score

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeperAlready knew it
02  Meta research blog, Nov 9 · HN: 266 points at 06:27 IST

🔥 Meta says Muse Ember matches Haiku 5.5’s best score at a third of the cost - its coding score is lower

Meta, Nov 9: Muse Ember is its smallest model, $0.10 per million input tokens and $0.50 per million output tokens on the Meta Model API. Meta’s table: same Artificial Analysis Index as Haiku 5.5 (36), lower on FrontierCode 2.0-AA (27.2 vs 34.9).
SCENE

Cline founder Saoud Rizwan, in a quote Meta published with the launch: Muse Ember is "purpose-built for, and excels at, subagent work like coding and document analysis," for teams cutting costs "by running it as the subagent in their harness." A partner’s statement, chosen and posted by Meta.

THE THING

Meta’s claim is that Ember matches the best score of Anthropic’s Haiku 5.5 at about a third of the cost; the numbers in Meta’s own comparison are Artificial Analysis Intelligence Index 36 for both, against 28 for Luna. On coding the same table is less flattering: FrontierCode 2.0-AA 27.2 for Ember against 34.9 for Haiku. Meta also published partner quotes: Box on contract and candidate screening, Cloudflare saying Ember ran "up to 30% faster at the tail" than the comparable model it tested, You.com saying it completed every coding quest with the lowest latency of any model it tested. All of that is Meta-posted partner testimony, not independent evaluation.

✂️ STEAL IT

Steal the subagent split: if your harness already has an orchestrator and workers, the worker is where a $0.10/$0.50 model earns its place. Run it on your cheapest, most repetitive subtask (grep-and-summarize, file triage) and compare to what you pay now.

Try today  VERDICT

Pick one repetitive subagent job and run it on Ember next to your current worker. It fails you if the task needs coding quality: Meta’s own table has Ember below Haiku 5.5 on FrontierCode, and an outside lab has not yet published its numbers.

Meta: Muse Ember  ·  HN discussion

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeperAlready knew it
📊 THE ATTENTION GAP
HN points at 06:27 IST. Attention, not quality.
Anthropic’s Fable 5.2 and Mythos 5.2 launch
314 HN
Fable 5.2 scores 65 on Artificial Analysis
312 HN
Meta’s Muse Ember
266 HN
ALSO ON THE RADAR

Anthropic’s Fermat claim is Anthropic’s. The launch post says that in prerelease testing the model "formalized the proof of Fermat’s Last Theorem in 30 hours" and "solved two open problems in the physics of black holes." Both are Anthropic’s own statements in its announcement; this issue has not seen an independent check of either. The post →

🧪 ONE THING TO TRY TONIGHT
Same task, two models, one cost number.
Thirty minutes. Pick one real task you do weekly - a bug fix, a refactor, a doc summary. Run it on Opus 5.5 and on Fable 5.2, and write down three things for each: did it finish, how long did it take, and what did it cost in tokens. Then run the cheapest repetitive step of your agent loop on Muse Ember. Anthropic recommends starting on Opus 5.5 and Artificial Analysis measures Fable 5.2 as slow and expensive; find out on your own workload whether the extra score is worth it.
What this issue could not see. The Anthropic announcement and Meta’s Muse Ember post were read in full, along with Artificial Analysis’s Fable 5.2 page, and the HN counts were taken at 06:27 IST. Every benchmark, partner and tester quote about Fable 5.2, Mythos 5.2 and Muse Ember is the vendor’s own, picked by the vendor; the only outside number in this issue is Artificial Analysis’s, and it covers Fable at Max effort only. We have not seen an outside evaluation of Muse Ember. Nothing was installed or run. Nvidia: the story moved again on Nov 9, but what was read was a U.S. presidential statement relayed by AP, not an agency or company confirmation, and no developer consequence beyond No. 037’s was established, so it is not covered today. Reddit remains absent by policy; X blocks anonymous access and was not consulted. VIDEO OF THE DAY is omitted for the fifth day running: nothing fresh, relevant and verifiable on today’s stories turned up - we omit rather than pad. Deduped against every published issue, No. 001-037, via the kit shipped-index and the live home archive: Fable 5.2, Mythos 5.2 and Muse Ember have not been covered before.
Talk to the editor 💬

Reply to this email with anything: what to chase, what to drop, what you want more of. Every reply shapes the next issue.

More of this kind of issueLess of this
Images: credited in each report; source screenshots and announcement artwork, not our tests.
Don't miss what's next. Subscribe to The Diff:
← Newer No. 039: Anthropic says it is prototyping secret scanning for Claude Code - and that auto mode is the real answer. Fable 5.2's report card: #1 on the index, $26.54 per task, one dev's $34 bill. Older → No. 037: Goldman pulls Claude Code from its bankers and traders - Bloomberg says the October guidance cut off everyone but engineers. Nvidia’s board investigates Jensen Huang.
Powered by Buttondown, the easiest way to start and grow your newsletter.