Anthropic cut its best model by 20 percent days before its IPO, and Meta answered with a 10-cent model
Hi everyone,
Anthropic cut the price of its best model by 20 percent on Monday, made it 30 percent faster, and did it seven days before the biggest stock market listing in history.
That is not a gift. That is a pitch.
Claude Fable 5.2 came out on November 9, with a twin called Mythos 5.2 that most of us will never touch. Fable is the one you can actually use, through the Claude API, Claude Code, and the big cloud platforms. It reads a million tokens at once, writes up to 128,000 tokens in one go, and costs $8 in and $40 out per million tokens, down from $10 and $50.
The four numbers that matter:
- 30 percent faster output than Fable 5.1, mostly on coding work
- $8 in, $40 out, a 20 percent cut from the old price
- 49.1 percent on FrontierCode 2.0, ahead of Opus 5.5 at 45.5, GPT-6 Astra at 42.8, and Gemini 4 Carbon at 41.1
- First place out of 36 models on the independent Vals Index, at 71.31 percent
A cheaper flagship is not generosity. It is a company that needs the world to believe its model is the standard before it rings the bell.
The part the leaderboard hides
Anthropic plans to list on the stock market in the week of November 16. Reports say it could raise up to $100 billion, which would make it the largest listing ever, at a value over $1 trillion. Sales growth has slowed, and the reports point at the same two pressures: OpenAI on one side, cheaper Chinese models on the other. So the week before the listing, Anthropic ships its best model, faster and cheaper than the last one. The timing is doing a lot of the work here.
And the independent numbers back the model, just not the headline. Vals AI ran Fable 5.2 on the same day it came out and put it first out of 36 models. But Gemini 4 Carbon sits at 70.45 percent and Opus 5.5 at 70.14, and the error bars overlap. First place, yes. Clearly better than the rest, no. The honest reading is that three models are bunched at the top and Fable is holding the lead by a nose.
The coding scores are the real story. Fable scored 93.77 percent on VibeCode, first out of 127 models tested. Code migration came in at 83.27 percent, second out of 30. On the kind of work where a model has to keep going, use tools, and finish a long job, Fable looks genuinely ahead of the pack.
But the same test run shows Fable got worse than its own predecessor at some jobs. On Tax Agent Bench it scored 32.09, 27th out of 84, down from 50.75 for Fable 5.1. On legal work it scored 69.17, behind Carbon at 75.93. And on Terminal-Bench-Science, Anthropic's own table says 74.6 while the independent run says 55.71, with GPT-6 Astra ahead at 62.86. Same test, two labs, opposite order. That is the part nobody puts in a launch post.
Same day, the opposite answer
Meta gave the other answer on the same day. Muse Ember, from the Meta Superintelligence Lab, costs 10 cents per million input tokens. It is built for agent work: tool calls, coding subtasks, routing, pulling facts out of long documents, the boring repetitive jobs that eat tokens by the million. Meta says it matches or beats Claude Haiku and DeepSeek V4.1 Flash on smarts, while costing far less than Haiku and OpenAI's Luna.
Two answers to the same question, on the same day. Anthropic says the best model, cheaper than before. Meta says a good enough model, nearly free. I think Meta is reading the room better. The fight is moving from who has the smartest model to who has the cheapest model that still gets the job done, because agents do not run once. They run thousands of times a day.
Who should actually switch
Fable costs exactly twice what Opus 5.5 costs, $8 and $40 against $4 and $20, and Anthropic still tells most people to use Opus as the default. Fable is for the hardest reasoning and the longest agent jobs, the ones where you hand the model a repo on Monday and check back on Friday. If your work is ordinary coding and writing, Opus stays the sensible pick, and that is Anthropic's own advice, not mine.
Mythos is the stranger half of the story. It is the same model with looser safety limits on cybersecurity and biology work, and only vetted organisations can use it. The first Fable and Mythos were pulled days after their June launch, when the US Commerce Department ordered Anthropic to bar foreign nationals over national security worries. So the most capable version of this model is locked away, and the one you can buy is the one with the brakes on. That split is now part of the product, not a footnote.
So here is where I land. Fable 5.2 is the best coder you can buy today. It is also a launch timed for Wall Street, not for you. Both things can be true at once.