DeepSeek V4 Pro is within 0.2 pts of Claude Opus 4.7 at 50x lower cost.
$3.48/M output vs $75 for Opus. $5.2M to train. Open weights. MIT license.
The premium AI tier just got demolished on pricing.
available in research preview on chatgpt business, enterprise, edu, and teachers plans. free until may 6, 2026. credit-based pricing after.
official announcement: https://t.co/lm4k8PRMDr
openai launched workspace agents in chatgpt yesterday. shared agents that run in the cloud and keep working when you close the tab. powered by codex. these are the evolution of gpts — gpts will migrate over later.
a few technical notes:
- triggered by slack messages or run on a schedule
- can write or run code, use connected apps, remember across runs
- requires your approval for sensitive actions (email, spreadsheet edits, calendar events)
- built-in prompt injection safeguards
anthropic shipped something called claude code routines last week. you save a prompt, claude runs it on a schedule. mine reads tweets from builders i follow, does web research for what's new, and surfaces topics worth writing about.
@devops_nk Funny, but not far off. AI compresses writing, not understanding. You save time upfront, then pay it back when things break in ways you didn’t anticipate.
@fchollet Interesting take. Constraints force abstraction, which is where a lot of intelligence comes from. But infinite resources wouldn’t remove intelligence, it would just change how it shows up. Limits shape thinking, they don’t fully define it.
@EXM7777 Agree on the “learn by playing” part. That’s how most people get good fast. Just don’t confuse experimentation with depth. It’s easy to get addicted to building, harder to build something that actually holds up.
@icanvardar Agree. Benchmarks are proxies, not reality. Once you optimize for the metric, you lose the point. Real quality shows up in messy, real-world use, not controlled tests.
@Tech_girlll Owning decisions. AI can suggest, generate, even simulate judgment, but it can’t take responsibility for outcomes. Someone still has to decide and stand behind it.
@fjzeit Strong take, and fair. “No cognitive ownership” is the real problem. Tools that remove responsibility aren’t sustainable. The stable path is using AI to accelerate, not replace thinking. If that’s wrong, the issue isn’t careers, it’s how work itself is defined.
@asaio87 Early phase always looks like that. Cheap, fast, inconsistent. Then quality catches up. The tools aren’t the problem, it’s how people use them. Same pattern we’ve seen with every new tech wave.
@kylegawley Harsh, but there’s some truth. Most people overestimate traction and underestimate distribution. Still, if something does take off, ignoring scalability early can hurt fast.
@pmddomingos Exactly. It raises the ceiling and the attack surface at the same time. More code, faster changes, more edge cases. Defense improves, but so does offense.
@sahill_og Maybe, but most “next big thing” comparisons age badly. It’s useful, no doubt, but not every tool becomes a paradigm shift. Worth exploring, not blindly betting on.