What does Claude Fable 5.2 cost, and does it keep my prompts?

Anthropic's post, dated November 9, 2026, lists $8 per million input tokens and $40 per million output tokens for Fable 5.2 and Mythos 5.2, and $16 and $80 for fast mode. Anthropic says eligible customers can continue with zero data retention (Anthropic's help center calls that option "available for a limited time") before they put its safety monitoring (EFS) in place; for other customers it has "30-day retention for safety monitoring". Anthropic's help center adds that data is deleted after 30 days except content flagged by its safety systems or legally required to be kept. GitHub's year-end zero-data-retention exemption is GitHub's statement about Copilot customers, not an Anthropic rule.

Does a small decision model cut AI costs by 98 percent?

The Strands post, dated November 9, 2026, says one Amazon ticket-labeling workload moved to a decision model that "performed just as well at 98% less cost". That is the authors' claim about one internal workload at list prices that leave out the Lambda, and the post gives no accuracy figures. The pages we read do not say the result holds for other tasks.

What does Meta charge for Muse Ember?

Meta's post, dated November 9, 2026, lists $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200K tokens; a footnote gives higher rates for longer prompts. Its cost-per-task chart says "estimates at proposed pricing" in the caption and "list pricing" in the footer. The page prints no stage label.

Is Microsoft's live-speech patent a product?

No. US 2026/0341309 A1 is a published patent application, not a product or an announcement. It describes software that swaps spoken words in live video. We found no Microsoft statement, product page or demo.

Signal / Noise Nov 10: Claude Fable 5.2 Retention

Q: How much does the AWS AgentCore Web Search API cost?

AWS says queries are billed at $7 per 1,000 queries, the same rate as Web Search through AgentCore Gateway, with no gateway charges. That is $0.007 per query by our arithmetic. AWS's post, dated November 10, 2026, prints "general availability" and lists six Regions.

Five small objects stand in a row on a dark brushed-metal isometric platform with a glowing red edge: a metal coin stack, a tiny sorting machine with a red glass hopper, a magnifying lens on a stand, a chip with glowing red circuit traces, and a blank paper sheet with a red seal and a metal stamp.

The short version

Anthropic lists Claude Fable 5.2 at $8 per million input tokens. It says eligible customers can keep zero data retention (its help center calls that option "available for a limited time") and other customers get 30-day retention for safety monitoring. Also: a small ticket-sorting model, AWS web search at $7 per 1,000 queries, Meta's Muse Ember and a Microsoft patent application, not a product.

In plain words

Think of AI models as helpers that each come with a price per job and house rules for what happens to your notes. Four items this time are prices and rules: a new model from Anthropic, a small ticket sorter described on the Strands Agents blog, a web search from AWS and a small model from Meta. The fifth is a patent filing, which is an invention on paper, not a product. For each we say what it is in one plain sentence, what the source's own pages print, and what we could not check.

Four of this edition's five items come with a cost figure, and the fifth is a patent filing. Signal means you can act on it this week. Watch means worth knowing; one cheap check at most. The items are dated November 9 and 10, 2026 (the patent application was published on October 29, and we say so). Every number belongs to whoever published it. For the previous edition, see the November 9 edition.

Signal: What does Claude Fable 5.2 cost, and how long are your prompts kept?

Anthropic, the company behind Claude, released a new model, Claude Fable 5.2, on November 9, 2026. It published a price, a faster mode and rules on how long your prompts are kept. Prices below are per million tokens, and a token is a chunk of text, often part of a word.

Anthropic's post gives one price that covers both Fable 5.2 and a second model, Mythos 5.2: "$8 per million input tokens and $40 per million output tokens". An optional fast mode "speeds up token generation by up to 2.5 times" and costs "$16 per million input tokens and $80 per million output tokens". Our arithmetic: fast mode is 2 times the standard price in both directions (16 / 8 and 80 / 40). Anthropic's pricing page lists Fable 5.1 at $10 and $50, so the new standard price is 20 percent lower (8 / 10 and 40 / 50).

Anthropic's Covered Models help page lists Claude Fable 5.2 as "Generally available". The launch post says "Claude Fable 5.2 is available today on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure." GitHub's changelog says the model is "now generally available in GitHub Copilot"; that describes Copilot only.

What happens to your prompts?

Zero data retention, or ZDR, is an arrangement under which, in Anthropic's API docs, Anthropic "does not store customer prompts or responses at rest after the API response is returned"; the docs also list products it does not cover. Several sources speak to it; we keep them apart.

Anthropic's post says: "Eligible customers can continue using Claude Fable 5.2 and Claude Mythos 5.2 with zero data retention before implementing safety monitoring with EFS. For other customers, we have 30-day retention for safety monitoring." EFS is Enterprise Frontier Safeguards. The post says that with it, "automated safety monitoring is run on prompts and responses that are stored in a customer’s infrastructure, making it compatible with zero data retention (ZDR)". It is "now in private beta", and enterprise customers can ask for access through their account team or a form.

The 30-day rule is not new with 5.2. Anthropic's help center says its policy for these "covered models" "went into effect on June 9, 2026". It says the 30 days apply "on every platform where these models are offered" and that the change applies to organizations that have set up zero data retention. Deletion after 30 days has exceptions: the help center names data that was "flagged by our automated trust and safety systems" and data Anthropic is "legally required to keep".

GitHub's changelog is a separate source and speaks only about Copilot. It says the model "requires data retention by default" and that "Retained data is not used to train Anthropic’s models". It adds that eligible enterprises can use the model with zero data retention "through the end of the calendar year under a time-bound exemption"; after that, "continued use of the models will require EFS". That deadline is GitHub's statement about Copilot customers, not a general Anthropic rule.

The pages we read do not say what makes a customer "eligible" or what EFS costs. Anthropic's Covered Models page calls the ZDR option "available for a limited time" and says Anthropic "may modify or withdraw the arrangement". Where retained data sits depends on the route (Bedrock, Google Cloud or Azure Foundry); the help center describes each, and we did not read the cloud providers' own retention terms.

What it means for you: in our reading, before anyone on your team sends client or staff data to this model, decide who may accept 30-day retention, and ask your counsel whether any contract or law applies (we are not lawyers).

Our take: Signal. In our reading, the terms are a decision you can take this week.

Signal: Can a small "decision model" like Strands Decider sort tickets for far less than a general AI model?

A decision model weighs a question against answers fixed in advance and returns probabilities instead of writing text; here, it picks a ticket category. Hugging Face's explainer, dated November 10, 2026, puts it this way: "A decision model takes some input context (state), evaluates a question with a set of predefined possible answers (primitives), and returns a structured decision along with probabilities."

A post on the Strands Agents blog, dated November 9, 2026, shows one case. Its footer names Amazon Web Services, and it says "Amazon and AWS teams relied on Claude Sonnet 4.6 to label tickets in an internal system". Two Amazon engineers then swapped it for Strands Decider, "a decision model we launched a month ago, that performed just as well at 98% less cost". The tickets could be "customer escalations, capacity requests, automated alarms" or "anything else across the company", and had to "fit within 20 category types".

What the post measured

The post's table shows an average time per call of 4,215 ms for Sonnet 4.6 and 897 ms for Strands Decider E4B (ms is milliseconds, thousandths of a second). It shows 6,724 tokens per ticket against 2,010. On cost, "at published us-east-1 on-demand prices", it gives about $22,200 against about $500 per million tickets and calls that "about 44 times cheaper".

Two cautions come from the post itself. These are list prices that "leave out the Lambda", a separate AWS service the workload also uses. And the Decider figure depends on how busy each server is: the authors say full use "would be $250 per million tickets", and they add headroom for bursts to reach about $500. Our check: the post's token counts per ticket, priced at the list rates it prints, come to about $0.0222 per ticket, matching its $22,200 per million tickets. The $250 full-use figure also checks out ($1.408 an hour divided by the post's 5,600 tickets an hour). We did not reproduce the $500 figure, which rests on the authors' allowance for bursts.

The post says the change "will save millions of dollars of token costs in a year". That is a projection, and the post does not itemize it.

What is not shown

This is one internal workload, reported by its own authors. The post gives no accuracy figures, only words: the model "performed just as well", the authors say it "got clearly better results" than their first Decider build on a test set, and they report "E4B beating production Sonnet on a separate set of our tickets". The post mentions no independent audit. The authors also report that on long tickets the model "drops the oldest comments to fit" its 4,096-token window. The authors name the limits: "If your labels need explanation text, if the catalog changes weekly, or if you can’t fit the useful context into a few thousand tokens, a decision model probably can’t replace the LLM’s job." The Hugging Face post has none of these figures (we searched it); we use it only for the idea. We tested neither model.

What it means for you: in our reading, if you pay an AI model to sort tickets or forms into a fixed list, ask your vendor whether a smaller model built for that one job exists and how its accuracy was measured on your data.

Our take: Signal. In our reading, the authors' own list of when it does not fit is as useful as the headline number.

Signal: How much does the AWS AgentCore Web Search API cost?

A web search API lets software ask for web results and pay per search. AWS's What's New post, dated November 10, 2026, announces "the general availability of the Web Search API on Amazon Bedrock AgentCore", AWS's service for running AI agents. "General availability" is AWS's printed label.

Search was not new on AgentCore. The post says web search "has been available as an MCP tool through AgentCore Gateway" (MCP is a common standard for connecting agents to tools). The new InvokeWebSearch API "gives applications, custom agent loops, and data pipelines search results with no resources to create". Results come from "the Amazon-operated web index", and each includes "a text extract and, when available, the page title, URL, and publication time".

The price: "Web Search API queries are billed at $7 per 1,000 queries, the same rate as Web Search through AgentCore Gateway, with no gateway charges." Our arithmetic: $7 / 1,000 = $0.007 per query. AWS's pricing page repeats the $7 rate.

AWS's developer guide adds one condition: the acceptable-use terms "include the requirement to retain and display the source citations and links that come with each result". The guide says each account can by default send "up to 10 InvokeWebSearch requests per second in each Region". The API is offered in six Regions: US East (N. Virginia), US East (Ohio), US West (Oregon), Asia Pacific (Sydney), Asia Pacific (Tokyo) and Europe (Ireland).

The three pages we read (the post, the pricing page and the developer guide page) do not state search quality or index size, and the Web Search section of the pricing page lists no free allowance. The $7 is the search line only, not a total for an agent's task. Web Answer, a different AgentCore feature with its own price, is in our November 5 edition; we do not compare the two.

What it means for you: in our reading, if an agent you pay for searches the web, ask how many searches one task makes, because AWS bills per query.

Our take: Signal, small. In our reading, it is a price list whose conditions are printed on the same pages.

Watch: What does Meta say its small model Muse Ember costs?

Muse Ember is a small AI model from Meta, which Meta says is built for the repetitive steps an AI agent performs all day. Meta's post is dated November 9, 2026 and calls it "our smallest and most efficient model designed for complex reasoning execution and the work agents repeat all day".

The post says it "is available today on the Meta Model API, at $0.10 per million input tokens and $0.50 per million output tokens". Meta's Availability section lists "$0.004 per million cached input tokens". A footnote sets the limit: "Pricing shown is for prompts up to 200K tokens. For longer prompts, Muse Ember costs $0.30 per million input tokens, $0.012 per million cached input tokens and $1.50 per million output tokens." The page prints no stage label.

Two labels sit on the cost-per-task chart: its caption says "Muse Ember costs are estimates at proposed pricing", and its footer says "Muse Ember cost at list pricing". The rates above are in the post's text and its Availability section. The page does not say how "proposed" and "list" relate.

The post also carries Meta's benchmark results, competitor comparisons and partner quotes; we did not check them and do not repeat them. Meta's earlier Muse models are in our October 19 edition.

What it means for you: in our reading, when a small model is quoted at one low price, ask which prompt length that price covers.

Our take: Watch. In our reading, a price is useful only together with the prompt length it covers.

Watch: What does Microsoft's live-speech patent application say? (A patent application, not a product)

A patent application asks for exclusive rights to an invention; it is not a product. US 2026/0341309 A1, published October 29, 2026 by Microsoft Technology Licensing, LLC, describes software that scans live video, replaces spoken words and, in some cases, redraws the speaker's lips.

Claim 10 lists words it may replace or remove: "filler words, personal words, personal phrases, colloquial terms, mumbled words, mispronounced words, misused words, misspoken words, offensive words, insensitive words, or inappropriate words".

We read the text on Justia, a patent-search site, because we could not read the USPTO copy. In it, the published text contains no consent step and no requirement to tell viewers. Claim 14 says the original recordings "are stored as ground truth for future reference". One passage describes a speaker choosing to use the system; the text does not make approval a requirement.

What it means for you: in our reading, ask live-video vendors whether edits are labelled.

Do this week

Three quick checks for an owner or engineering lead; this is not legal or buying advice.

Before anyone sends client or staff data to Claude Fable 5.2, decide who may accept 30-day retention of prompts and answers, and ask your counsel whether any contract or law applies.
If you pay an AI model to sort tickets, forms or emails into a fixed list, ask your vendor which model does it, what it costs per thousand items and how accuracy was measured on your data.
If you buy live-video AI tools, ask whether edits are labelled for viewers and whether the original stays available.

Book an AI discovery call

How we know, and what we could not check

We read every source on November 10, 2026 and tested nothing: Anthropic's post, pricing page, developer docs and two help center articles, GitHub's changelog, the Strands and Hugging Face posts, AWS's post, pricing page and developer guide, Meta's post, and the patent application in Justia's copy. Not read: the Claude Fable 5.2 system card, the cloud providers' own retention terms, the Strands Decider code and serving notes, Meta's evaluation report, Artificial Analysis, and USPTO's own viewer. For platform buyers, see our comparison of enterprise agent platforms; for another model's terms and pricing, our GPT-6.1 Terra teardown. See also our AI agents work.

Signal / Noise (Nov 10, 2026): Claude Fable 5.2 Terms, a Small Model Sorting Tickets, AWS Search Prices and a Microsoft Patent Application