Skip to content
Technology

Fable 5.2's Savings Outpace Its Price Cut

Endor Labs found Fable 5.2's real coding bill fell 77 percent, far past Anthropic's 20 percent price cut, because it uses fewer tokens.

Edited by Ryan Elliott Dennis4 min read · 975 words · 6 sources
A vintage calculator beside its original sales receipt, the paper trail that shows what a gadget actually cost
“Anthropic advertised a price cut of 20 percent on its newest model. An outside lab ran the same coding tasks and found the real bill fell 77 percent. Can a sticker price ever tell a buyer what an AI job will actually cost?”

Anthropic cut Claude Fable 5.2's price on Nov. 9. The sticker fell 20 percent, from $10 to $8 per million input tokens, and from $50 to $40 on output1. Anthropic shipped the cut as it readied for a historic IPO that could land within the week.

Dianne Penn, the company's head of product management for research, said the release matches Dario Amodei's September call for a slower pace. Fable 5.2 reached every user the same day. Mythos 5.2 carries the same underlying model under a different name. It stayed limited to vetted organizations in Anthropic's life sciences and cyber verification programs1.

Endor Labs wanted a harder number. The security research firm ran the same coding tasks on Fable 5.1 and Fable 5.22. It tracked every token along the way.

Fewer Tokens Explain the Real Drop

Endor Labs' median task on Fable 5.2 cost 54 cents. The same task on Fable 5.1 cost $2.332. That gap is a 77 percent drop, close to four times the 20 percent Anthropic advertised.

Fable 5.2's Real Cost Drop Nearly Quadruples Its List Price Cut

Percent drop from Fable 5.1 to Fable 5.2: Anthropic's list price per token against Endor Labs' measured cost per finished coding task

The first bar is the price cut Anthropic advertised; the second is the drop Endor Labs measured by running matching coding tasks on both models.

Data through Nov. 16, 2026.

Sources: CNBC [1]; Endor Labs [2]. Chart by Kelly Dennis.

The numbers behind this chart
ItemValue
Anthropic's list-price cut20%
Endor Labs' measured cost drop77%
A UniCom 141P printing calculator, which sold for 695 dollars when it launched in 1972
A UniCom 141P printing calculator, sold for $695 when it launched in 1972. Its price tag left how many calculations it would take to balance a ledger an open question.Swtpc6800 Michael Holley / Wikimedia Commons, Public domain

The difference traces to one number: tokens. A token is a small chunk of text, close to a syllable, and a model pays for every one it writes. Fable 5.2 wrote 68 percent fewer output tokens than Fable 5.1 across Endor Labs' full run2.

Quality held steady too. Luca Compagna, the Endor Labs researcher who ran the benchmark, found the model's security score improved along with its speed. "On our tasks the thought got cheaper too: the cost per counted secure pass fell from $10.19 to $2.84," he wrote2.

A Secure Coding Pass Costs $7.35 Less on Fable 5.2

Endor Labs' cost per counted secure pass, Fable 5.1 against Fable 5.2, on matching coding tasks

Each row is the same measure, the cost of one secure coding pass; the blue dot is Fable 5.1's cost, the magenta dot is Fable 5.2's, and the rod is the gap between them.

Data through Nov. 16, 2026.

Source: Endor Labs [2]. Chart by Kelly Dennis.

The numbers behind this chart
ItemFable 5.1Fable 5.2
Cost per secure pass$10.19$2.84
A visitor signs in at Google's Mountain View headquarters
A visitor signs in at the Googleplex, Google's Mountain View headquarters. Google priced Gemini 4 Hydrogen below its last Flash model on Nov. 12.Erik Möller / Wikimedia Commons, Public domain

Google's Cut Runs Into the Same Question

Google shipped a similar bet three days later. Gemini 4 Hydrogen launched Nov. 12, priced at $1.50 and $7.50 per million tokens, a flat rate across its full context window5. That rate treats a short question and a sprawling one the same way.

Koray Kavukcuoglu, Google DeepMind's chief AI architect, wrote the model's own introduction. "Today, we're introducing Gemini 4 Hydrogen, bringing cutting-edge intelligence to everyone — with speed and efficiency for everyday coding and agentic workflows," he wrote5.

Artificial Analysis measured the same tradeoff Endor Labs found in Fable 5.26. Hydrogen's price fell, and its token use on complex work fell with it. Its skill at coding alone in a real terminal lagged behind. The model scored 54 percent on Terminal-Bench, a timed test of command-line work, while Claude Sonnet 5.5 scored 64.

Dan@DanDr1s

🚨 Fable 5.2 is looking seriously impressive.

It beats Opus 5.5, GPT-6 Astra and Gemini 4 Carbon on 13 of 14 benchmarks in this comparison.

Some standout results:

• 89.6%: CVE Rediscovery
• 89.3%: Computer Use
• 85.1%: Protein Design
• 82.1%: Agentic Search
• 76.3%:… pic.twitter.com/4d9RhxI65V

A chart claiming Fable 5.2 beats rival models on 13 of 14 benchmarks
Open on X (opens a new tab)

A benchmark chart like this one is the sticker price. Endor Labs' task-level run, below, is closer to the receipt.Post by Dan (@DanDr1s) on X, Nov. 9, 2026

The Skeptics Trust the Receipt Over the Sticker

Some analysts read a list price with more suspicion than Endor Labs does. A sticker, in their view, answers a different question than a bill.

Sanchit Vir Gogia, chief analyst at Greyhound Research, reads Fable 5.2's cut as a narrower claim than Anthropic's headline suggests. "Anthropic has cut the price of a token, not the price of a thought," he said3. He wants buyers to track spending against finished work, apart from the rate printed on the invoice.

Adam Pease of Aragon Research reaches a similar caution from the cost column. "Fable 5.2 still costs twice as much as Opus 5.5 ($4/$20), and fast mode doubles the price again," he wrote4. A team that reaches for Fable 5.2 on habit alone pays full price anyway.

Anthropic's Own Advice Undercuts Its Headline

The company's own guidance agrees with the skeptics more than its press release does. Penn told CNBC that Fable 5.2 suits "power users or power use cases," hard problems in engineering, mathematics and robotics1. Opus 5.5, a cheaper model Anthropic released in September, handles the daily work most buyers actually run.

That routing advice matters more to a buyer's bill than the price cut does. A 20 percent cut applies to a model built for a narrow set of hard jobs. It saves little for a business that sends routine work there anyway. The real decision, Anthropic's own routing advice shows, happens before the invoice, at the moment a team picks a model.

What to do

Check your own task cost before trusting a list price. Run a handful of real jobs on the old model and the new one. Then weigh the finished bill against the rate card, since one tells a different story than the other.

Match the model to the job the way Anthropic's own guidance does. Send routine work to Opus 5.5 or a similarly priced model. Save Fable 5.2 for the hard problems it was built for, and skip it everywhere else.

Ask your vendor for a task-level number, the way Endor Labs published one. A rate card shows what changed on paper. The finished bill shows what changed for your business.

By the numbers

  • 20 percent: Anthropic's advertised cut to Fable 5.2's list price, from $10 to $8 per million input tokens1.
  • Nearly four times: how much further the real per-task cost fell, 77 percent, against that list-price cut2.
  • Fable 5.2's median task cost 54 cents, Endor Labs found, down from Fable 5.1's $2.332.
  • 68 percent: fewer output tokens Fable 5.2 wrote than Fable 5.1 across Endor Labs' full run2.
  • $10.19 to $2.84: the cost per secure coding pass Endor Labs measured, Fable 5.1 against Fable 5.22.
  • Gemini 4 Hydrogen prices at $1.50 and $7.50 per million input and output tokens5.
  • Ten points: the gap between Hydrogen's 54 percent Terminal-Bench score and Claude Sonnet 5.5's 646.
  • Twice: what Fable 5.2 still costs against Anthropic's own Opus 5.5, by Aragon Research's count4.

What to watch

Watch whether more vendors publish a task-level cost, the way Endor Labs did, apart from a rate card alone. See whether Google answers Artificial Analysis's terminal score with a fix, or lets Hydrogen's speed carry the story regardless. The clearest test still belongs to the buyer: run the job, then check the receipt.

Sources

  1. CNBC, "Anthropic releases Claude Fable 5.2 and Claude Mythos 5.2 ahead of IPO," CNBC, Nov. 9, 2026, https://www.cnbc.com/2026/11/09/anthropic-claude-fable-5-2-mythos-5-2.html
  2. Luca Compagna, "Fable 5.2: cheaper and faster, not more secure, and much more memorization," Endor Labs, Nov. 16, 2026, https://www.endorlabs.com/learn/fable-5-2-cheaper-and-faster-not-more-secure-and-much-more-memorization
  3. Prasanth Aby Thomas, "Anthropic unveils Claude Fable 5.2, Mythos 5.2 with coding gains and lower prices," InfoWorld, Nov. 10, 2026, https://www.infoworld.com/article/4246681/anthropic-unveils-claude-fable-5-2-mythos-5-2-with-coding-gains-and-lower-prices.html
  4. Adam Pease, "Fable 5.2 Price Cut Rewards Model Routing Discipline," Aragon Research, Nov. 9, 2026, https://aragonresearch.com/fable-5-2-price-cut-rewards-model-routing-discipline/
  5. Koray Kavukcuoglu, "Introducing Gemini 4 Hydrogen for Agentic Workflows," Google, Nov. 16, 2026, https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/
  6. Artificial Analysis, "Gemini 4 Hydrogen Models - Intelligence, Performance & Price Comparison," Artificial Analysis, Nov. 12, 2026, https://artificialanalysis.ai/models/gemini-4-hydrogen

Header photo: Joe Haupt from USA / Wikimedia Commons, CC BY-SA 2.0

Cite this piece

Kelly Dennis, "Fable 5.2's Savings Outpace Its Price Cut," Ai Lately, Nov 17, 2026, https://ailately.com/articles/fable-5-2-savings-outpace-price-cut

Tags: Claude Fable 5.2 · AI pricing · Gemini 4 Hydrogen · inference cost · token economics

Related, lately

More

Keyboard

j k
Move through a list
Enter
Open the selected piece
/
Search, and browse the columns
g then r
Right Now
g then a
Articles
g then s
The Signal
g then o
Opinion
g then b
Analysis
g then p
People
?
This sheet
Esc
Close

AI Right Now is free with an account.The Citation Audit is free with an account.Skills are free with an account.Tutorials are free with an account.You have read your 3 free articles.

Every article and comic, AI Right Now, the Citation Audit, Skills and Tutorials.

Join Ai Lately for free

Join with Google or your email. Already a member? Sign in

Press anywhere else for the home page.