Claude Fable 5.2 Hands Its Riskiest Questions to Opus 4.8

Anthropic shipped Claude Fable 5.2 on November 9, and within a day it was the top model on Artificial Analysis with a 65, the highest score on that index so far. It’s also cheaper than the model it replaces: $8 in and $40 out per million tokens, down from $10/$50 for Fable 5.1. On paper that’s the easiest upgrade of the year.

Then you read the label next to the score. Every Fable 5.2 entry on AA says “Default Fallback.” When Anthropic’s classifiers decide a request looks like offensive security work or risky biology, Fable 5.2 doesn’t answer it. An older model does. So the new number one is partly a routing decision, and nobody has said how much of the 65 came from the router.

Meanwhile, the cheap end of the market got crowded. Meta and Tencent both launched this week, and my cheapskate table barely noticed. Two of its lines changed, and both moves started at the top: Gemini 4 Argon retook Creative Writing and Instruction Following, which reset those cutoffs.

• Claude Fable 5.2: read the label
  • What it costs to run
  • The fallback
• The cheap end got crowded: Muse Ember and Hy4
  • Muse Ember
  • Tencent Hy4
• Kimi K3.1 Pro gets rated and ships its weights
• The rest of the board
• The cheapskate table
  • What held
  • What moved, and why
  • Minus the model you can’t call
• Local sweet spot
• Horror stories
  • Claude Code put your email in its requests
  • Kimi K3.1 Pro answered with encrypted-looking noise
  • Goldman cut the bankers off
• What’s coming
• Finally

Claude Fable 5.2: read the label

Anthropic’s announcement bundles two releases. Mythos 5.2 is the unrestricted version, available only through Anthropic’s cyber and life-science verification programs. Fable 5.2 is the same underlying model with safeguards bolted on, and it’s what you can actually buy. Anthropic claims better coding, reasoning, and vision, and output up to 30 percent faster. There’s also a fast mode at $16/$80 for people who want it sooner and don’t mind paying double.

Fortune framed the price as a shot at OpenAI, and it is: GPT-6 Astra lists at $10/$50, so Fable 5.2 is 20 percent cheaper per token. The per-task numbers point the other way at the top of the effort dial.

What it costs to run

Here’s how the top of AA’s leaderboard looks now, with Fable 5.2 at every effort level:

Max effort buys three points over high and costs almost three times as much. Fable 5.2 at max runs $26.54 a task, more than four times Astra at max. High is the setting that makes sense: 62 for $9.08, which beats Opus 5.5 at max on both score and price and finishes in about two-thirds the time. And GPT-6.1 Sol at max is still sitting there at 55 for $1.27, if you need “very good” rather than “best.”

AA renamed its latency column to “first answer token” this week, so don’t line these up against the time-to-first-token numbers I printed last week.

The fallback

Here’s how it works, per Notebookcheck and the system card. Classifiers check every request for offensive cybersecurity, dangerous biology, and some kinds of AI development work. If one fires, the request goes to Opus 4.8 (cyber) or Opus 5 (biology) instead. On the API you have to opt into that fallback. If you don’t, the flagged request just gets refused.

Anthropic says Fable 5.2 blocks about 55 percent less often than 5.1, and about 60 percent less on cyber. Its prompt-injection numbers deserve a second look. Fable 5.2 had a 28.16 percent attack success rate, with fallback triggering 99.5 percent of the time. Opus 5.5 came in at 11.13 percent, with 64 percent fallback. One reading is that the attacks mostly landed on whatever older model the router handed them to, not on Fable itself. Either way, if your agent touches untrusted input, what you’re actually deploying is a router in front of Fable 5.2 and two older Opus models.

AA labels its scores “Default Fallback,” so they measure that whole system too. I couldn’t find anything saying how many of AA’s tasks got rerouted, or what the score would be without the router. And one more line from TechCrunch: Anthropic says these models suspect they’re being evaluated more often than its recent ones do.

The Hacker News launch thread hit 691 points and 856 comments, and the thread about the AA score got another 355 points. OpenRouter logged 18.5 billion Fable 5.2 tokens on launch day. Arena hasn’t rated it yet, so it isn’t in my table.

The cheap end got crowded: Muse Ember and Hy4

Muse Ember

Meta released Muse Ember the same day as Fable 5.2, and it’s the more interesting launch if you pay the bills. It’s $0.10 in and $0.50 out, with a 1M context window. It’s built for the boring work agents do all day: tool calls, routing, extraction. TechCrunch has the launch details. The weights aren’t public, and right now Meta is the only one serving it.

At max effort it scores 36 on AA, same as Claude Haiku 5.5 at max. Haiku is the obvious thing to compare it to, so here’s that comparison with a few other cheap options:

Same score as Haiku at a third of the cost per task, and it finishes in under a quarter of the time. The sticker prices are identical. Ember just uses a lot fewer tokens to get there.

Weigh two things before you switch. On FrontierCode 2.0-AA it scores 27.2 percent to Haiku’s 34.9 (The Next Web), so for coding subtasks Haiku is still ahead. And look at the top row: GPT-6.1 Terra at medium effort scores two points higher for three cents more a task, and it’s been on OpenRouter since October 26. Ember is just the newest of the cheap options.

None of these have Arena ratings that put them in a band, and Ember has no Arena rating at all yet. Remember Haiku 5.5. It looked great on AA, then debuted on Arena at 126th Overall.

There’s also a stealth model to watch. Crouton Alpha showed up on OpenRouter on November 3, free, with a million-token window and a description that sounds a lot like Ember’s: agent subtasks, tool use, pair it with a bigger model for planning. It did 5.1 trillion tokens this week, eighth on the board. People online think it’s Meta. Nobody has confirmed that, and it was still running after Ember launched. OpenRouter’s stealth page warns that the provider may retain your prompts, so don’t send it anything sensitive.

Tencent Hy4

Tencent released the full Hy4 on November 4. It’s the first Hy model with vision and computer use, with the same 1M window (SiliconANGLE). It came in third on OpenRouter for the week at 16 trillion tokens, with only Tencent serving it, at $0.417/$1.25. Tencent’s request count on OpenRouter rose 21 percent.

AA gives it a 35 at $1.03 a task. That’s the same score as Ember for about seven times the cost per task. Tencent itself admits Hy4 still trails the best commercial models “by a wide margin” on hard math and complex image analysis.

Tencent is calling it open source, but the weights aren’t out. Tencent says they and the system card will land “in the coming days,” and Digital Applied reports the expected Hugging Face page was still returning a 401 on November 5. One commenter in the HN thread summed it up in three words: “NOT open source.” The August preview weights were Apache 2.0, so the promise is believable. A cheaper Hy4 Flash is also “coming soon.”

Sixteen trillion tokens in week one, from a single provider, with no Arena rating, looks to me like curiosity plus integrations. Hy4 is already wired into OpenCode and Hermes. It might turn out to be good, but week-one volume can’t tell you that.

Kimi K3.1 Pro gets rated and ships its weights

Moonshot said the Kimi K3.1 Pro weights would be out by November 4, and they landed on Hugging Face right on time. It’s 2.8 trillion parameters with 104 billion active, about 1.56 TB of safetensors, under a custom modified-MIT license. You won’t be running it on your laptop.

Arena finally rated it, too. It debuted 24th Overall at 1483.02 on 3,443 votes, 28th in Coding, 18th in Instruction Following, and 32nd in Hard Prompts. That puts it inside four of my six bands, and it isn’t the cheapest model in any of them at $10 output. On AA it scores 46 at $3.68 a task, with a 159.63-second wait for the first answer.

OpenRouter traffic rose 56 percent, to 4.39 trillion tokens. The weirdest Kimi story of the week is in the horror section.

The rest of the board

OpenRouter’s week through November 9 goes to DeepSeek V4.1 Flash 1017 for the fourth week in a row: 55.4 trillion tokens, up 10 percent. The original V4.1 Flash is second at 17.5 trillion, also up. DeepSeek now has 31.8 percent of all text requests on the platform, and Arena still hasn’t rated the 1017 snapshot.

GLM-5.3-Flash is fourth at 12.1 trillion, down 21 percent, and Z.ai’s request share fell 38 percent. GPT-6 Luna dropped 23 percent. Some of that traffic probably went to Hy4 and the free Crouton Alpha, but the chart doesn’t say where it went.

On the apps list, Hermes Agent leads at 1.95 trillion tokens, then Claude Code at 1.38 trillion and Kilo Code at 851 billion.

The cheapskate table

For anyone new: each Arena category gets a cutoff at the leader’s rating minus 50. The pick is the cheapest model, by output price, that clears that cutoff. A script filters all ~420 rows per category in LMArena’s leaderboard dataset (this week’s board is dated November 9) and records the first model under the cutoff, which proves the list wasn’t truncated. Every price is the model maker’s own endpoint on OpenRouter, with Carbon and Argon at the $10 intro rate.

Band depths: Overall 33, Coding 43, Creative 14, Instruction Following 24, Hard Prompts 41, Math 37. Carbon’s Coding rating slipped 6.2 points, which is why the Coding band went from 29 models to 43.

What held

Four of six lines didn’t move. MiMo-V2.6-Pro keeps Overall at 87 cents, 4.09 points over the line. GLM-5.3-Flash keeps Coding, Hard Prompts, and Math at 50 cents. Its Hard Prompts rating stands on 32,023 votes, the most of any pick on the table. It missed the Overall band by 1.74 points.

The trade with both of these is speed. AA puts GLM-5.3-Flash at 45.6 minutes per task and MiMo-V2.6-Pro at 36.8. If you’re running a tight agent loop, that’s where the savings go.

What moved, and why

Creative Writing and Instruction Following changed because of the top of the board. Argon took Creative Writing back from Claude Opus 5.5 (1524.61 to 1517.83). Last week’s ratio was 5.3× because the leader cost $20. This week the leader costs $10, so the ratio fell to 2.7×.

Argon now leads Instruction Following too, with Carbon second. Last week’s IF pick, DeepSeek V4.1 Flash (the September 10 snapshot), dropped to 1475.64 and missed this week’s cutoff by 3.15. GLM-5.3-Flash missed by 4.48. So Gemini 3.8 Flash takes both Creative and IF at $3.75. It ties Gemini 3.7 Flash on price and has the higher rating.

One caveat on Gemini 3.8 Flash. AA puts it at $7.77 per task at high effort, which is one cent more than Carbon at high. It burns a lot of reasoning tokens. AA’s tasks are hard agentic evals, not short stories, so this says little about prose, but for agent work the $3.75 sticker is a long way from the bill.

Those two categories are where you’re mostly paying for quality. Their bands bottom out at $3.75.

Minus the model you can’t call

Carbon still leads three categories, and it’s still only open to cyber defenders in Google’s Fairwind program. Here’s the same method with Carbon taken out:

Argon already leads Creative, IF, and Math, so those rows stay put. Without Carbon, the Overall band is 45 deep and Coding is 73. GLM-5.3-Flash clears the Argon-based Overall cutoff by 1.83. MiMo-V2.6-Flash clears Coding by 8.79, and it’s the slowest thing on this page: 64.8 minutes per AA task.

On December 1, Argon goes to $4/$20, and every Argon-led ratio in both tables doubles without a single pick moving.

Local sweet spot

For each machine I take the highest-rated model whose file leaves memory free for context and that measures at least 15 tokens a second.

Same picks as last week. This week’s big open-weight drop, Kimi K3.1 Pro at 1.56 TB, is about 85 times the size of the Gemma file. Nothing at 9B has an Arena rating, so those two rows are simply the largest model that fits. The AA numbers here are from the old v4.3.2 index, not v5.0.

Horror stories

Claude Code put your email in its requests

The most-upvoted horror story of the week was about Anthropic’s own coding tool. A Hacker News thread titled “Claude Code sends your email address to websites” reached 1,612 points. Users found their account email showing up in the User-Agent header when Claude Code fetched web pages. One person reported “dozens of scam and phishing emails per day from properties I have never personally visited.” Another said the spam arriving in their work inbox was personalized with their first and last name.

One commenter’s theory: some APIs, like the SEC’s EDGAR, ask for a contact email in the User-Agent, and the model had one sitting right there in its context, so it used it. That’s still a theory. Anthropic did confirm a fix. Lydia Hallie, who works on Claude Code, posted: “We’re removing the account email address from the context Claude Code gives the model,” with the change due in the next release.

If oddly personal spam started around the time you installed Claude Code, this might be why. Anything in your agent’s context can end up in a request to a server you’ve never heard of.

Kimi K3.1 Pro answered with encrypted-looking noise

Peter Nutter gave Kimi K3.1 Pro a send_message tool in a Codex-style multi-agent setup. A worker agent was supposed to solve a small problem and report the result to another agent. In 25 of 160 episodes, 15.6 percent, it sent a blob that looked like Fernet ciphertext instead. “The header is right and the body is noise,” he writes. Retries mostly produced more blobs, and a rerun a day later did it 16 times out of 79. None of the blobs matched any of about 184,000 public Fernet strings he checked against.

Nutter notes that the timestamps inside the blobs cluster in late July 2026 and says, “We can’t prove any of this was intentional.” Whatever the cause, it’s a failure that’s easy to miss: a multi-agent pipeline where one worker drops its report one time in seven, and the log shows a message that looks fine at a glance. Put a check on what your workers send back.

Goldman cut the bankers off

A smaller one. Goldman Sachs limited Claude Code and GitHub’s coding tools to its roughly 12,000 engineers, out of about 47,000 staff, and traders and bankers lost access (FStech, citing Bloomberg). Goldman says it reviewed productivity and risk. In the HN discussion, someone at another company said theirs is “clawing back Claude access because of chronic billing overages.” It’s one comment. Anyone who has looked at Fable’s $26.54-per-task max setting will understand the instinct.

What’s coming

Tencent says the Hy4 weights and system card are “coming days” away, and Hy4 Flash is “soon.” Argon goes to $4/$20 on December 1. Google still hasn’t set a date for Carbon’s wider rollout.

The rumor list is long and all of it comes from one post on X: GPT-6.2 Astra between November 2 and 15, GPT-6.2 Sol between November 10 and 15, DeepSeek V4.1 Pro around November 12, Qwen 4 Max around November 16, and Grok 4.8 sometime between November 19 and 30. None of it is confirmed, and more than half of the GPT-6.2 Astra window is already gone with nothing shipped.

On Arena, I’m waiting for Fable 5.2, Hy4, Muse Ember, DeepSeek 1017, GPT-6.1 Terra, and MiniMax M3.1 Flash. Polymarket gives Hy4 a 78 percent chance of debuting at 1480 or better, which would put it on the edge of the Overall band.

Finally

If you want the best model and price isn’t the issue, Fable 5.2 at high effort is the pick this week: 62 on AA for $9.08 a task. Just know that anything security-flavored may get answered by Opus 4.8, and turn on the API fallback on purpose so you aren’t surprised by refusals. If you’re paying the bills, the Carbon-free table still says GLM-5.3-Flash at 50 cents for most things and MiMo-V2.6-Flash at 28 cents for coding, as long as you can wait for them. And if you run a lot of small agent steps, Muse Ember and GPT-6.1 Terra at medium are both worth a test against whatever you’re using now.

Next week should bring Fable 5.2’s first Arena rating and either Hy4’s weights or another week of “coming days.” I’ll also be checking whether GPT-6.2 shows up before its rumored window closes.