Anthropic cut the price of cache reads on its newest models by 75% this week, and nearly every writeup repeated the company’s “roughly 25%, up to 45%” savings estimate without checking which workloads actually collect it. I ran the arithmetic off the published price table, and two builders on identical models can see a 39% cut and a 0.3% cut on the very same change.
The Only Line That Moved on Anthropic’s Price List
Anthropic shipped Claude Fable 5.1 and Mythos 5.1 on September 1. What makes the change legible is that Anthropic publishes the whole rate table, so you can see what moved and, more usefully, what did not. Base input stays at $10 per million tokens. Output stays at $50. The five minute cache write stays at $12.50 and the one hour cache write stays at $20. Exactly one line changed: cache reads fell from $1.00 per million tokens to $0.25, which the docs describe as 0.025x the base input price, against the 0.1x that every other Claude model charges.
Here is the part the coverage skipped. That cut lands entirely on the cached-input line, so the size of your saving is set by the ratio of cache reads to output tokens in your own workload, not by the discount. Two worked examples at the published rates:
| Workload | Fable 5 | Fable 5.1 | You save |
|---|---|---|---|
| Agent loop: 200k cached context re-read 40 times, 100k output | $15.50 | $9.50 | 39% |
| Long-form generation: 20k cached context re-read 5 times, 500k output | $25.35 | $25.28 | 0.3% |
Same models, same 75% cut, same week. The agent loop saves 39% because cache reads were most of its bill. The generation job saves 0.3% because output tokens were most of its bill, and output did not move at all.
So: if you run long agent loops that re-read a big context many times, this is a genuine cut and you should take it. If you generate long documents from short prompts, ignore the headline and change nothing about your budget. One more thing worth checking before you retune anything: the one hour cache write is still 2x base input at $20 per million, so a cache that gets written and never re-read still loses you money at the new prices exactly as it did at the old ones. This is the second pricing surprise from Anthropic in three days, after the Sonnet 5 increase that was scheduled for September 1 and then quietly cancelled.
Matters. Go measure your cache-read-to-output ratio before you budget a cent against this.
Sony and Warner Want $150,000 Per Song From the Company You Build On
The publishing arms of Sony Music and Warner Music sued Anthropic late Friday night in federal court in Northern California, and TechCrunch’s account of the filing describes a “brazen campaign” of intellectual property theft. Sony Music Publishing and Warner Chappell allege Anthropic trained Claude on tens of thousands of copyrighted compositions, pulling lyrics and sheet music from the Beatles, Taylor Swift, Michael Jackson and hundreds of other writers. They name Anthropic, CEO Dario Amodei and co-founder Benjamin Mann, and they are asking for up to $150,000 per infringed work plus $25,000 for each instance of stripped copyright management information.
The number that makes this serious is not the one in the complaint, it is the one already on Anthropic’s books. The company settled an authors’ class action for $1.5 billion, a figure a judge approved this July, and it settled on precisely the theory the publishers are running here: Judge William Alsup held that training on lawfully purchased books was transformative fair use, but that building a library out of pirated copies was not. The music publishers read that split and filed into the half Anthropic lost.
Our read: this is not a repeat of the authors’ case, it is a test of whether one adverse ruling becomes a template every rights holder can file against. If it does, the training-data liability sitting under every frontier lab stops being a tail risk and starts being a line item, and the labs with the deepest pockets are the ones most worth suing. That is a pricing story eventually, not just a legal one. It also lands on a company already spending the week explaining itself in Washington, where the Pentagon’s 180-day clock on Claude was running out.
Breaks your stack, slowly. Nothing changes in your code this week. But if you are betting a product on a single model vendor, vendor legal exposure is now a real column in that spreadsheet.
Microsoft Took Down Its Own Auth Layer, and Copilot Went With It
A core authentication configuration failed at 3:08pm UTC on August 31 and cascaded across Microsoft 365, and TechCrunch reported the outage was still dragging on the following day. Exchange Online, Teams, SharePoint, OneDrive and Microsoft 365 Copilot all went down or degraded. Copilot prompts that needed Microsoft 365 data, which is to say most of the useful ones, simply failed.
The lesson is not that Microsoft had an outage. Everyone has outages. It is that Copilot did not fail on its own merits: it failed because it is downstream of an auth layer shared with mail and file storage, so one config error took out the assistant and the data it reads in a single stroke. Deeply integrated AI inherits the blast radius of everything it integrates with, and that tradeoff never shows up in the procurement deck.
Breaks your stack. If an AI feature you sell depends on a third-party assistant, your status page needs to know about their auth layer.
Google’s “Up to 88% Fewer Tokens” Is a Real Change Wearing a Marketing Number
Google introduced agentic video understanding across Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. Google’s own announcement claims it reduces analysis costs by up to 66% and token consumption by up to 88% while improving accuracy by up to 7%.
Three “up to” figures in one sentence is the tell, and on LongVideoBench a best case is not a median. But the mechanism underneath is genuinely different and worth your attention. Instead of ingesting video at a fixed frame rate, default 1 FPS, the model dynamically searches, scans and inspects target segments across frames, audio and transcripts. It decides what to watch and at what resolution rather than swallowing everything at a constant sample rate.
That is the right architecture for long-form video, where static sampling has always forced a bad choice between an enormous token bill and a preprocessing step that throws away the moment you needed. Take the mechanism seriously and treat all three percentages as a ceiling you will not hit.
Marketing wrapper, real engineering. Benchmark it on your own footage before you rewrite a pipeline.
OpenAI Put ChatGPT Inside Epic, Read-Only, Across 325 Million Patient Records
Healthcare organizations can now connect Epic electronic health record environments to ChatGPT for Healthcare, and TechCrunch’s writeup of the launch puts Epic’s reach at more than 325 million patients. Clinicians can pull appointment notes, lab results, medications and specialist documentation, then ask questions across them. A companion Healthcare Public Data plugin wires in ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed and PubMed. OpenAI says physicians rated 99.1% of responses safe across 27 clinical use cases.
Two design decisions here are worth stealing regardless of your industry. The integration is strictly read-only, so the model cannot write back into the record, which caps the worst-case failure at a bad answer rather than a corrupted chart. And the scope is retrieval and summarization over data the clinician already has access to, not diagnosis. That is a deliberately small surface, and small surfaces are how you ship AI into regulated workflows at all.
The 99.1% deserves a harder look than it is getting. At the volume of a hospital system, the residual 0.9% is not a rounding error, it is a daily occurrence, and OpenAI is defending two lawsuits over health guidance already. The read-only constraint is doing more safety work here than the benchmark is.
Matters. The pattern to copy is read-only plus narrow scope, not the accuracy number.
Instagram Will Bury AI Influencers That Will Not Admit They Are AI
Instagram is replacing its optional “AI creator” tag with a required “AI-generated profile” label for accounts where the person on screen is entirely synthetic, and Engadget reports that accounts skipping the label lose reach. Unlabeled synthetic accounts stop being surfaced in Reels and Explore. They can still post to existing followers, they just stop going viral. The rule targets fully artificial personas, not real creators using AI tools, which is the distinction that makes it workable.
This is enforcement by distribution rather than by removal, and it is the most honest thing a platform can do here: Meta is not claiming it can detect synthetic media, it is making non-disclosure expensive. Anyone running an AI-assisted content operation should read the line carefully, because “real creator using AI tools” is where most legitimate businesses sit and that side of it is untouched.
Marketing, and overdue. Disclosure is cheap. Rebuilding an audience after a demotion is not.
What I Am Watching
The cache-read cut is the only item here that changes a number in your budget this week, and for half of you it changes it by almost nothing, which is the whole reason to check rather than assume. The music publishers’ case is the one to track over months, because it decides whether the authors’ settlement was a one-off or a template. And the Microsoft outage is the free lesson: the more deeply your AI is integrated, the more things can take it down that have nothing to do with AI.