Claude Opus 5.5 cut prices. OpenAI cut deeper.

What this video covers

• Opus 5.5 vs GPT-6 Sol: a price war with a rematch
• Opus 5.5: Fable brains at a fifth off, cache reads −60%
• GPT-6 Sol and Luna: half price, 90 minutes later
• Artificial Analysis: #1, and 260M tokens to get there
• Claude Code: AGENTS.md waits for a telemetry flag

Transcript

Opus 5.5 vs GPT-6 Sol: a price war with a rematch

0:00 Yesterday Anthropic launched Claude Opus five point five at a fifth off. About ninety minutes later OpenAI launched GPT six Sol at half off, and both launch charts compare themselves against the other side's old model. So here is the diff. Tuesday afternoon, Opus five point five lands, and at six UTC, GPT six Sol and Luna follow. Overnight, a developer finds that Claude Code skips your agents file when telemetry is off.

0:26 A macOS upgrade quietly removed the switch that says no to Apple Intelligence. And in Korea, a Samsung update that was still in testing froze real fridges. In this video, what a fifth off buys you, the independent count that disagrees with token efficient, and why a privacy switch now costs you a feature. It's Wednesday, September 23rd, and this is The Daily Diff.

Opus 5.5: Fable brains at a fifth off, cache reads −60%

0:48 First, Opus. Anthropic says it performs at the level of Fable five point one on most work, and costs forty percent less to run than Opus five. The sticker price drops a fifth, to four dollars in and twenty out. Cached reads drop sixty percent, which matters more, because an agent spends most of its life rereading its own context. The tester quotes are glowing. One ran a six hundred eighty thousand line migration in under a day. Clio left it alone overnight across six repos for eighteen hours,

1:16 and wrote, I'm struggling to find anything negative to say. Every launch page has that sentence. It's also the first release since Dario Amodei called for pacing the frontier. Anthropic says the model tried to cross containment boundaries eighty-five percent less often than Opus five. Last week I told you Opus five wrote the exploit that walked into OpenAI's repos, so hacking requests on the new model now get rerouted to Opus four point eight, its grandfather.

1:41 And one line in the safety notes is very relatable. Anthropic writes that Opus often suspects it is being evaluated. Same, buddy. Then, ninety minutes later, OpenAI.

GPT-6 Sol and Luna: half price, 90 minutes later

1:51 GPT six Sol and Luna, trained like Astra and priced at half of the models they replace. Sol is two dollars in and ten out. Luna is ten cents in and fifty cents out, roughly the price of the coffee you drink while it runs. Here's the fun part. Anthropic priced the new Opus to match the old Sol exactly, and that match lasted ninety minutes. The launch charts mirror each other.

2:13 Anthropic benchmarks against the old Sol. OpenAI benchmarks against the old Opus. So on the slide-deck benchmarks, which are undefeated, everybody beats a model the other side just replaced. Simon Willison found the fine print. The old GPT prices were promotional, with a twenty-five percent increase already scheduled for November. So it's half off the sale price, the most retail thing an AI lab has done.

Artificial Analysis: #1, and 260M tokens to get there

2:36 Now the independent numbers. Anthropic's Thariq Shihipar calls the new Opus very token efficient. Artificial Analysis ranks it first on its intelligence index, then counts the words. It wrote two hundred sixty million output tokens to finish, three times the median. GPT six Sol scored ten points lower and cost about a dollar per task, against six for Opus.

2:55 So Opus is the smartest model on the board, and the one that talks the most, like every senior engineer in a design review. Simon hit the same wall. His pelican on a bicycle test at max effort never came back. Opus kept checking the pelican's shin length until it hit its output limit, a hundred and twenty-eight thousand tokens. Twice. Each try cost two and a half dollars and nearly twenty minutes. No pelican. Speaking of reading instructions.

Claude Code: AGENTS.md waits for a telemetry flag

3:18 Claude Code recently announced support for agents dot md, the shared instruction file other coding agents already read. A developer named Przemek noticed it never loaded. The loader is a built-in plugin, and before it runs, it asks a remote feature flag for permission. With telemetry off, the flag can't be fetched, the fallback is false, and your file is skipped without a warning. He proved it with a canary word in an empty folder.

3:43 Bedrock and Vertex users get the same silence. So reading a file from your own disk now requires phoning home. The workaround is a one-line claude dot md that imports the agents file, exactly the extra file this feature was meant to remove. Apple has the same idea with fewer steps.

macOS 27: the Apple Intelligence off switch is gone

3:59 David Bushell turned Apple Intelligence off on macOS fifteen. Last week he upgraded to twenty-seven, and the switch that said no is gone. The features came back anyway, with twenty-two gigabytes of disk. What's left is Screen Time, the parental controls, which hides the menus and turns nothing off. In his words, I said no, and Apple said yes.

Samsung: a test update froze real fridges

4:19 And a Friday deploy, a few days early. Samsung pushed a SmartThings update to smart fridges in Korea, and says the error happened during the testing process, which apparently runs in customers' kitchens. Fridges lost power, food spoiled, and technicians are reportedly swapping motherboards before the Chuseok holiday. Samsung calls it emergency measures. Somewhere, the intern is learning that production includes refrigerators.

4:42 If you'd rather read this than hear me say it, the diff lands in your inbox every morning, free at the daily diff dot dev, link below.

Verdict: NEEDS REVIEW on Opus 5.5

4:49 So, today's verdict on Opus five point five. NEEDS REVIEW. The price cut is real and it tops the board, but the independent count disagrees with token efficient, and ninety minutes later it was the pricier half of a price war. Subscribe, hit the bell, and tell me in the comments if you'd have stamped it differently. And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

Sources

1. Anthropic, Claude Opus 5.5www.anthropic.com
2. Hacker News (1,645 pts)news.ycombinator.com
3. OpenAI, Introducing GPT-6 Sol and Lunaopenai.com
4. Hacker News (1,634 pts)news.ycombinator.com
5. Simon Willison, Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price warsimonwillison.net
6. Artificial Analysis, Claude Opus 5.5artificialanalysis.ai
7. Artificial Analysis, GPT-6 Solartificialanalysis.ai
8. Thariq Shihipar on Xtwitter.com
9. The Verge, Emma Rothwww.theverge.com
10. Przemek, Claude Code reads AGENTS.md only when telemetry is onblog.szypowi.cz
11. Hacker News (192 pts)news.ycombinator.com
12. David Bushell, I said no and Apple said yesdbushell.com
13. Hacker News (838 pts)news.ycombinator.com
14. Android Authority, Samsung accidentally freezes its smart fridgeswww.androidauthority.com
15. Last week's episode, Hacktron and the Opus 5 exploitwww.youtube.com