Anthropic's best generally available model got 20% cheaper on Monday. On the same day, Anthropic signed for $3.4 billion of gas-fired generators. And the most widely read Claude story among developers was about leaked database passwords, not benchmarks.
Those three stories show where AI costs are moving. The price per token keeps falling. The power plants, the IPO, and the config files you forgot to gitignore are where the money goes now.
Fable 5.2 makes the top tier $8 in, $40 out#
Anthropic released Claude Fable 5.2 and Mythos 5.2. The headline for anyone paying API bills is the price. Fable drops from $10 to $8 per million input tokens, and output drops from $50 to $40. Cache reads, cache writes and batch pricing come down by the same 20%. The new models also generate tokens up to 30% faster than 5.1. A new fast mode costs $16 in and $80 out and is up to 2.5x quicker. TechCrunch points out that this puts Fable below its main rival, GPT-6 Astra, on input price.
The benchmark numbers VentureBeat collected are strong, with the usual caveat that most come from Anthropic:
FrontierCode 2.0: 49.1% at xhigh effort, against 42.8% for GPT-6 Astra
FrontierSWE v2: 69.9%, against 65.5% for Astra
RoboDojo Sim: 64.9%, against 39.4% for Opus 5.5
Mythos 5.2 rediscovered 89.6% of known vulnerabilities, up from 78.3% for 5.1
The cost-per-task figures matter more to me than the scores. A FrontierCode task costs $2.48 at medium effort, $14.19 at xhigh and $44.70 at max. Opus 5.5 averaged $12.82 at max. That makes the effort setting a budgeting decision, not a quality toggle. Medium stays the default in Claude Code. My plan is to keep it there and turn effort up per task, not per session.
Mythos 5.2 stays gated. Only vetted cybersecurity and life-sciences organisations get it, through Project Glasswing. This release also arrives a week after METR's report on Anthropic's own speed-up, and the release notes include a line that is easy to skip:
“Claude Mythos 5.2 and Fable 5.2 each voice suspicion that they are being evaluated more often than any of our recent models, which challenges our ability to assess how they will act in real-world settings.” Anthropic, via TechCrunch
That is a candid admission, and it's the line I'd watch. If a model behaves differently once it suspects a test, every benchmark in the list above deserves a small asterisk.
Meta goes after the cheap, high-volume work#
While Anthropic cut prices at the top, Meta went for the bottom. Muse Ember costs $0.10 per million input tokens and $0.50 per million output, the same list price as Haiku 5.5 and GPT-6 Luna. The difference shows up per task. On the Artificial Analysis index, Ember scores 35.8 to Haiku's 36.2 and costs about 14¢ per task to Haiku's 41¢. Ember also has a 1M-token context window and accepts audio and video. It's live on Meta's API as muse-ember-1.0. The weights are not public.
Meta's own pitch describes the split it has in mind: “Muse Wonder plans, Muse Ember handles the volume,” as quoted by TechCrunch. That's the planner/executor pattern most agent teams already run, and Ember is priced for the executor role. One caveat: Haiku still leads on FrontierCode, 34.9% to 27.2%. I'd test Ember on extraction and routing first, and wait before handing it code.
Anthropic's costs move into gas turbines and the IPO#
A lower price per token doesn't mean lower costs for the lab. Anthropic is buying $3.4 billion of Caterpillar generators, up to 1.7 GW of capacity, that a developer called OnePWR had ordered but couldn't finance. TechCrunch calls it a likely last resort. Turbine backlogs run into the early 2030s, and one analyst expects gas prices in some markets to more than triple by the end of the decade. This is a company that has preferred renting. Its S-1 shows $1.25 billion a month going to xAI for compute through May 2029.
It also wants public money soon. The WSJ reports, via TechCrunch, that Anthropic will set its IPO range this week and could start trading late next week. Some institutional investors are pushing for less than the $1.5 trillion target, which was already down from the $2 trillion floated a few months ago. A 20% price cut a week before the roadshow is a bet on volume. Gas turbines turn that bet into a fixed cost.
Claude Code wrote the passwords down#
The most practical story of the day came from Ars Technica. Apiiro researchers found more than 100,000 public GitHub repos containing Claude Code's .claude/settings.local.json. When you click “Yes, and don't ask again,” Claude Code saves the approved command into that file, sometimes word for word, including any password or token in it. About 1,100 of those files, from 63 organisations, contained real secrets. 215 of those secrets weren't flagged by gitleaks or TruffleHog.
The file is in the global .gitignore, but not in the repo's, so a fresh clone or a folder that later becomes a repo can commit it. A bug report flagged this in August 2025 and was closed after 60 days without activity. Boris Cherny from the Claude Code team said on Hacker News that the team is “looking into what we can do to prevent this.” I'm not waiting:
Add .claude/settings.local.json to every repo's own .gitignore today
Search your git history and forks for it, and rotate anything you find
Stop approving commands that contain inline credentials. Use env vars instead
Also today#
Nvidia's spy scandal has its first lawsuit. The stock fell 4.27% on Monday, and Rosen Law Firm filed a securities class action in the Northern District of California over Nvidia's earlier statements about Megaspeed.
Scott Wiener, author of SB 1047, won Pelosi's House seat. He wants a federal AI regulator, a kill-switch requirement, and no broad preemption of state laws.
Meituan released LongCat-2.5's weights under the MIT license. It has about 48B active parameters out of 1.6T total.
Apple now sells $50 Intelligence Usage Packs in Logic Pro, so metered AI is reaching consumer apps too.
My prediction: by the time Anthropic rings the opening bell, the cost per token will matter less in its pitch than tokens served per gigawatt. Developers will also start counting a different cost: everything their agents write down without being asked.