Fast mode is available for Opus 4.8. It's the same model at roughly 2.5x the speed, and we've made it three times cheaper than before.
Turn it on with /fast in Claude Code. On the API, contact your account manager to request access or join the waitlist: https://t.co/5SLFF77Whn
Today we’re releasing DeepSWE, a new standard for agentic coding benchmarks.
On public leaderboards, top models often look relatively close in capability. DeepSWE shows where they actually diverge, reflecting the realistic experience of developers in their day-to-day work.
codex's 2x usage limit promo ends on may 31st
that's in 4 days
if your usage already feels like it evaporates too fast, get ready for june - it's going to be rough
openai is willing to give eligible enterprise customers 2 months of free codex usage
but pro users are about to lose the extra headroom that made codex feel like a workhorse
so i'm hereby creating a petition
dear @sama please make the 2x codex limits permanent for pro subscribers - you know it's the right thing to do
sign my petition by liking this post and sending it to your mom
three of the things we are most excited about:
1. AGI accelerating research
2. AGI accelerating companies
3. personal AGI accelerating everyone in achieving their goals
today it was great to announce the unit distance result.
yesterday it was great to announce that we are offering to invest $2M in openai credits into every YC company.
now we need to increase our efforts on the third!
Anyone cancelled Claude Code for Codex yet?
Feels like dev's are switching not because Codex is better.
but because it’s cheaper to actually get work done.
What’s your experience?
ngl the new kimi model is actually insane
> GPT 5.4 level coding, 7x cheaper than opus 4.7, completely open source and free to use
> also 30% cheaper than GLM 5.1
bro what 💀
The music industry had a pirate enemy.
It wasn't a person. It wasn't a company. It was a Python script.
It is called yt-dlp. 158K stars. Unlicensed. Sony couldn't sue it if they tried. There's no copyright to violate.
Then the music industry did the unthinkable.
October 23, 2020. The RIAA filed a DMCA takedown. GitHub complied. The repo and 18 forks vanished within hours.
Then one developer had an idea.
He merged the entire source code into GitHub's own DMCA takedown repository. The takedown notice now contained the tool it was trying to delete.
24 days later, GitHub reversed course. Reinstated everything. Put $1 million into a developer defense fund.
The music industry lost.
Here's the wildest part:
The community forked youtube-dl. Renamed it yt-dlp. Same code. New name. Latest release: March 17, 2026. Still impossible to kill.
One Python script vs. the entire music industry.
But DO NOT install it. Sony, Universal, and Warner deserve your $15.99/month.
100% Open Source.
(Link in the comments)
I don't believe a single one of those "I'm running 20 Claude instances in parallel" posts.
A single Claude code instance requires numerous amounts of back and forth to get it right, 20 means you're churning out complete slop that doesn't work.
I'm cancelling my Claude 5x Max subscription.
The models are ok, were better before.
I just don't like paying the company who is so unclear about things
Operates in the shadow
Isn't clear in communication on what we can do or not with their sub
When they asked for the feedback I said:
"You fucked up too many times"
I hope that they can wake up and do the right thing.
OVERRATED: running tons of agents in parallel; working on too many things at once; perpetual context-switching; opening lots of low-quality PRs that may never land.
UNDERRATED: using one or two agents at a time; focusing on the task in front of you; thinking deeply; finishing stuff; making your code works in prod.
Not a single line can be fully trusted, I don't see escape from reviewing everything at this stage. I am more concerned for PMs and TPMs than software engineers.
Claude Code just wrote 65,000 lines of slop for me 🤦♂️
> be Dan
> spend 2 hours writing app requirements
> create 68 spec files
> fire up the shell script with claude -p
> set it to 100 iterations
> run it 4 hours straight
> writes 65k LoC
Next day I open the app
> it was all done in the same context window w/ multiple compactions
> most code is not reachable
> root page crashes
> I can't even find the correct route
> after 5 min I think I found it
> the dashboard is a mockup 💀
Germany’s food safety agency called this study’s lab method “completely unsuitable for testing for microplastics.” When they repeated the experiment properly, they found 1,000 times fewer particles.
Here’s what happened. The “11.6 billion particles” number comes from a single 2019 study at McGill University in Canada. The researchers steeped plastic tea bags in hot water, then evaporated all the water so they could examine what was left under an electron microscope. The problem is that when the water dried, dissolved chemical residue from the bag’s manufacturing process hardened into tiny specks. The microscope counted those dried-out specks as plastic particles, even though they were dissolved in the liquid like sugar dissolves in coffee. They were never actually floating around as pieces of plastic in your tea.
Germany’s Federal Institute for Risk Assessment (BfR), their version of the FDA for food safety, caught this in 2020. They repeated the experiment but used a method that checks each particle one by one to confirm whether it’s actually plastic. They found 5,800 to 20,400 particles per bag. Not 11.6 billion.
A 2025 review paper co-authored by BfR researchers straight up called the original numbers “preparation artifacts,” meaning the particles only showed up because of how the lab samples were prepared, not because of the tea.
Two more things the tweet leaves out. First, the study only tested premium silky pyramid bags made from nylon and PET plastic. Those bags represent about 5% of the tea bag market. The flat rectangular tea bag most people actually use is made from wood pulp and abaca fiber, a plant from the banana family. Brands like Bigelow, Numi, Stash, Yogi, and Traditional Medicinals all use plant-based bags. Bigelow alone makes 2.2 billion tea bags a year with zero plastic.
Second, the total weight of all those “billions” of particles in the original study was 16 micrograms. That’s 16 millionths of a gram. A single grain of sand weighs about 50 times more.
No major health authority, not Germany’s BfR, not the European Food Safety Authority, not the WHO, has concluded that microplastics in food cause health problems at the levels people are currently exposed to.
The study is real. The number is probably wrong by a factor of 1,000. And it tested a product most tea drinkers don’t use.
We’ve partnered with Samsung to bring Perplexity directly into the upcoming Galaxy S26.
Every new S26 will ship with Perplexity built in as a system‑level AI, with its own wake word: “Hey Plex.”
Codex-5.3 is that mid-level engineer who finishes work on time but needs you to specify every edge case in jira or they short circuit.
Opus-4.6 is the staff engineer who 10xes revenue with killer features they invented but takes down prod every 5th Friday afternoon.
No it doesn't. Particularly with event driven microservices.
Biggest problem is- it tries to investigate and fix both by looking only for a patch work rather than trying to find deeper architectural problems.
One bad merge on this line, eventually everything becomes garbage.
install in 2 commands:
/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplace
before: claude yolo-codes 500 lines then you spend 2 hours debugging its hallucinated imports
after: claude works autonomously for HOURS without going off the rails because every step has a gate
this is the difference between AI writes code and AI engineers software
repo: https://t.co/IXw64mWvOO