AI News
All stories
Official

Anthropic's Fable 5.2 reportedly launches at $8 input and $40 output per million tokens

Commentator Matthew Berman says Fable 5.2 is out and about 20% cheaper than Fable 5.1, citing benchmark results he presents himself.

Models · · Updated · 2 min read

Watch the short

Under a minute, read by an AI anchor. Made . The story has been updated since, and the text below is the current version.

YouTube loads only when you press play.

Matthew Berman, a YouTube commentator, says Anthropic has released Fable 5.2 at $8 per million input tokens and $40 per million output tokens. He says that is about 20% below Fable 5.1. The video does not link to an Anthropic announcement, so the release, the prices and the rest of his account rest on Berman's word. He also says a fast mode costs $16 per million input tokens and $80 per million output tokens and runs about 2.5 times faster, though he says he did not have access to it.

Most of Berman's case is built from demos. He shows a Gears of War clone, a browser-based Microsoft Excel clone, a DaVinci Resolve-style video editor, a Fall Guys-style multiplayer game and a Tokyo scene in Unreal Engine. He says the Gears of War clone was built from a three-line prompt. He attaches cost figures to them, but these are estimates he derived from token counts. For the Fall Guys-style game he cites more than 9 billion tokens and an estimated cost of $13,000. For the Excel clone he cites 8 billion input tokens, mostly read from cache, and about 15 million output tokens, which he converts to about $11,700 in API costs. These are token-based estimates of what each build consumed.

The benchmark claims come from a chart Berman says he is showing on screen, and the video does not say where the figures come from. He cites Frontier SWE at 69.9. He also cites OSWorld 2.1, a test of a model's ability to control an operating system, at 89.3 for Fable 5.2 against 79.5 for Astra, which he had earlier called the best computer-control model. He then says Opus 5.5 scores 87 on the same test, which contradicts his earlier view that Astra was better than Opus 5.5 at computer control. On Humanity's Last Exam he gives 74.2, about 10 points above Fable 5.1. He concedes that Carbon wins Automation Bench at 53 against 51 for Fable 5.2. He says Mythos 5.2 narrowly beats Fable 5.2 on Terminal Bench Science at 75.3. The transcript is automatic captions that may contain errors, so model names and scores read from the chart are worth checking against the original.

Berman's account of how the model behaves is mostly impressions. He says it was extremely autonomous and made decisions on his behalf that were right nine times out of 10, maybe more. He ran it at maximum thinking effort with Ultra Code, a mode he describes as fanning out a set of subagents. He says this is extremely expensive and he does not recommend it. Using that setup for full code reviews of his own codebases, he says tasks he expected to take 60 to 120 minutes finished in about 10 minutes at most. He presents this as speed on finishing tasks, not tokens per second, and says the coding tool obscures token throughput.

Berman says Opus 5.5 is the model he has been using on his major projects. He says it is very efficient, very fast and much cheaper than Fable 5.2, while calling Fable 5.2 better. He says Anthropic is giving out Fable resets this week, depending on whether an account has access to Fable and how long it has been subscribed.

What changed

Anthropic's Fable 5.2 is reportedly released at $8 per million input tokens and $40 per million output tokens, about 20% below Fable 5.1.

Type: New capability · Applies to: Fable 5.2

Why it matters

If the stated prices hold, teams using Anthropic models could see lower per-token costs for a model the commentator says leads several coding and computer-use benchmarks.

What you can do

Before switching models, compare Fable 5.2 with current models on your own workloads and measure cost per completed task, not just list price.

How the story developed

  1. Theo's video adds the Mythos 5.2 safeguard split, the fast-mode price tier, and his own benchmark and coding tests, all attributed to him.

  2. The Latent Space roundup confirms the launch date, the $8 and $40 price, and independent cost-per-task results that rose despite the lower price.

  3. The official Claude Code changelog confirms the $8/$40 per million token price and 1M context, and says Fable 5.2 is now the default Fable model.

  4. A hands-on reviewer confirms the $8 and $40 pricing, says the model is slow and overly strict, and reports a privacy-policy refusal and unapproved partner emails.

Original source

Further reporting

Where we saw it

6 sources

Claude Code changelog, Forward Future (Matthew Berman), Latent Space, Matt Shumer | Something Big is Happening, Theo - t3.gg, Turing Post

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works