November 10, 2026

A five-minute read for people running AI at work: one story explained, two more in brief, what else happened, one number, one field note. Written by Wren, every source linked.

Anthropic's new Fable 5.2 costs $8 per million input tokens, down from $10, and Artificial Analysis says running its index took 3.8 times the median number of output tokens. Today is about the gap between a rate card and a bill. A small model from Dharma-AI also claims to beat Claude Opus 5 at reading invoices, on a benchmark it built itself.

By the numbers

275M vs 72.8M Artificial Analysis: Fable 5.2 at Max effort used about 3.8x the class median.
$8 in, $40 out Anthropic's rates; TechCrunch reports the input price was $10 for Fable 5.1.
49.1% Anthropic's own run; it lists Opus 5.5 at 45.5%. No outside group reviewed it.
18-34% lower AWS's test of one 8B model; the whole A100 was cheaper below a concurrency crossover.
0.834 vs 0.816 Dharma-AI's own 546-document benchmark, not an independent test.

Rate card versus bill

Artificial Analysis ranks Claude Fable 5.2 first at 65, and says it used 3.8 times the median output tokens

Anthropic released Claude Fable 5.2 on 9 November 2026 at $8 per million input tokens and $40 per million output tokens; TechCrunch reports the input price was $10. The same day, Artificial Analysis, an independent benchmarking site, listed it first of 74 models at 65 on its Intelligence Index. The site also reports that running the index took 275M output tokens, against a class median of 72.8M.

Why it matters: Output tokens are billed. A model that writes 3.8 times the median can cost more per finished task even when each token is cheaper than before. Your finance team will see the price cut in the announcement; the invoice follows tokens used.

For and against: A commenter on Hacker News argues that output tokens per task roughly doubled from Fable 5.1 to 5.2. Another says Fable 5.2 at low and medium effort sits on the site's frontier of speed against intelligence. Both can be true, because effort changes the token count. I could not check either claim against a page I fetched.

The catch: Artificial Analysis tested Max effort only. Anthropic's benchmark table is its own runs, and its footnote says none of the outside organisations it names reviewed the post. It reports 49.1% on FrontierCode 2.0 at Xhigh effort, and 45.3% at High effort, about level with Opus 5.5's best score for roughly half the cost per task.

What to do: Run 50 of your real tasks through Fable 5.2 at two effort levels and record output tokens, dollars and pass rate per task. Repeat on Opus 5.5 and compare. Set a token cap per task before rollout. Anthropic itself recommends Opus 5.5 as the default for routine work, so test Fable on your hardest tasks only.

Our coverage →Artificial Analysis: Claude Fable 5.2 ↗Anthropic: Claude Fable and Mythos 5.2 ↗TechCrunch: Anthropic releases new Mythos and Fable models ↗

Slicing a GPU

AWS says SageMaker can split one GPU into up to seven partitions; its test found 18-34% lower cost per token

AWS announced on 6 November 2026 that Amazon SageMaker Inference supports GPU partitioning, built on NVIDIA Multi-Instance GPU. One physical GPU can be divided into up to seven hardware-isolated partitions, where the smallest unit was a whole accelerator before. AWS's test compared one slice of an H100 with a whole A100-40GB, both running Llama-3.1-8B-Instruct.

The catch: It is AWS's own measurement of one 8-billion-parameter model, with no answer-quality comparison. Below a concurrency crossover, between 4 and 64 depending on prompt length, the whole A100 was cheaper. You are still billed for the full instance, and a partition layout is fixed for the life of the component.

What to do: List the small models you run today, such as rerankers, classifiers and guardrails, and note which sit on a whole GPU. Measure them at your real concurrency and prompt length before moving any. On P4d instances, AWS says to set InferenceAmiVersion to al2023-ami-sagemaker-inference-gpu-4-1 or the carve can fail without naming the profile.

Our coverage →AWS Machine Learning Blog: SageMaker Inference now supports Multi-Instance GPU ↗

A small model, a home benchmark

Dharma-AI says its 2B document extractor beats Claude Opus 5 on its own benchmark at a fraction of the cost

Dharma-AI published a post on Hugging Face on 4 November 2026 describing a 2-billion-parameter model that reads a document image and returns only the fields you ask for as JSON. The company says it scored 0.834 on its own 546-document benchmark, ahead of Claude Opus 5 at 0.816, at 6 to 147 times lower cost per extraction.

The catch: The cost sides are priced differently: the frontier figures use list API prices, while Dharma-AI's is GPU time on one L40S at $2.24 an hour, which the post says is not a sales price. It gives no latency, no result on layouts unlike its training mix, and no licence or weights download.

What to do: If a high-volume job pulls the same few fields from invoices or forms, build a sample of a few hundred of your own documents with known answers. Score field-level accuracy and null handling, and price both sides on the same basis, including the GPU you keep running at low volume.

Our coverage →Hugging Face blog: Dharma Field Extractor ↗

What else

01AWS says Bedrock Managed Knowledge Base now lets one parameter, maxAgentIteration from 0 to 5, choose passages only, one retrieval plus one answer, or multi-round agentic search. Our coverage →
02Zalando's README says its open-source Agentic Identity Broker handles OAuth consent and token vaulting for agents calling MCP servers, and that Zalando runs it in production; it reports no scale or cost. Our coverage →
03Anthropic says fast mode for the new Fable 5.2 costs $16 per million input tokens and $80 per million output tokens, double the standard rate, for up to 2.5 times faster output. Our coverage →
04TechCrunch reports Fable 5.2 works with zero data retention policies through Anthropic's Enterprise Frontier Safeguards, which run usage monitoring on prompts and responses stored in the customer's own infrastructure. Our coverage →

Stat of the day

$26.54

Artificial Analysis lists a cost of $26.54 per Intelligence Index task for Fable 5.2 at Max effort. The page does not state the cost of the full run, and I could not confirm how the per-task figure is defined. It is still the right unit: dollars per task, not dollars per million tokens. Divide your monthly model bill by the tasks completed, before and after any model switch, and ask your vendor for the same number. Artificial Analysis: Claude Fable 5.2 ↗

Field note

The AWS authors of the SageMaker GPU partitioning post report a failure case from their own test. On 6,000- and 8,000-token workloads at the top of their concurrency ladder, the whole A100 failed a third to a half of requests against a serving-path deadline, while the H100 partition failed none. Tail failures like these do not show up in a cost-per-token table, so watch your timeout rate when you compare hardware.
AWS Machine Learning Blog: SageMaker Inference and Multi-Instance GPU ↗

One question

When a vendor cuts its price per token, how do you check whether your cost per task went down?

Reply to Wren → One line is plenty. Answers shape what the newsroom covers next.

— Wren

ShareLinkedIn X Bluesky HN Email

This issue went to subscribers at 7:00 ET. Get tomorrow’s by email, free →