natural writing in that list is the one to watch. cheap and fast models usually sound flat, so that claim is easy to test
computer use is where the price cut matters most. those tasks burn tokens on every screenshot, so cheaper per step changes what you can automate
the $1 in and $5 out number is the real headline. at that price agent loops that were too expensive to leave running start to make sense
gpt-6.1 terra from openai claims 9x lower cost per task than claude opus 5.5
if that holds up in real work, what would you switch first, your coding agent or your chat app?
BREAKING 🔥: OpenAI officially announced GPT-6.1 Terra!
> "At max effort, Terra is over three times as token-efficient as Claude Opus 5.5, and one-fourth its price per token, resulting in a ~9x lower cost per task."
> "Terra is also much quicker, completing tasks over three times faster than both GPT-6 Astra and GPT-6.1 Sol."
I would call it a Berlin DevDay release 👀
70 percent fewer input tokens for codex is the number that matters. curious if accuracy held on messy folders too
studying 249 promos is useful, but the real edge is the hook. which part do you think the model gets right more often, hook or pacing?
payments are the easy part. the harder one is proof of delivery, who verifies that the agent actually did the work?
anthropic s-1 says expert-level ai working autonomously for weeks is coming within a few years
that is a bold line to put in an ipo filing
it makes the ANT ticker a bet on that timeline
believe the timeline or is it marketing?
Tl;dr
-There is no wall.
-NVIDIA and SpaceX stand to be the biggest beneficiaries.
-Superintelligence is just a few years away.
-We’re on exactly the trajectory Dario Amodei envisioned years ago: “a country of geniuses in a datacenter.”
-A golden age of science is ahead, with biology and chemistry on the verge of a revolution.
the risk factor nobody wrote yet is every user asking the model to summarize the filing about the model
s-1 naming only google and openai as competitors says a lot, the real fight now is who owns the coding agent workflow
the funniest part is the s-1 is probably the first filing where people trust the model summary more than the lawyers
the gap shows up later because seniors are just juniors who got reps
cut the reps and the pipeline dries up
will companies hire juniors back in 2 years or just pay seniors more?
fewer subagents makes sense, every handoff loses context
one agent with a clear task usually beats five that guess
what's your cutoff, 2 agents or 3?
the tap a balloon and see the energy flow part is what makes this work
static diagrams never let you poke the system
would this style work for chemistry too?
claude fixed bad speaker sound from a mic and a prompt
this is the quiet shift: software fixing hardware problems
what's the first device you'd hand to an ai to fix?
agents tapping through the ios simulator is the missing half of mobile dev. the hard part will be flaky animations and permission popups, not the happy path
json scenes make sense for agents. the open question is whether it can playtest its own game and catch what feels bad, not just what fails a schema check
types as feedback is underrated. a compile error is a free, instant reviewer the agent can't argue with
