Frander

AI systems & visual generation.

Testing AI tools | Agents | Prompts | Image/Video | Models | Creative experiments.

[Frander @Frandeeer]

CLAUDE IDENTITY CHECKS are coming.

Anthropic is rolling out verification for some users and certain capabilities.

In practice, this can mean:

passport
government ID
live selfie
third-party verification

This is not just a small UX update.

This is the beginning of AI access becoming gated.

Today: “verify for some features.”
Tomorrow: “verify to use advanced models.”
Next: “your region, document, or identity decides what AI you can access.”

People are still arguing about prompts.

The real game is moving to access control.

[Frander @Frandeeer]

Fiona Fung says Claude Code made engineers lonelier.

That sounds like a soft problem.

It is not.

AI agents are changing coding from a team activity into something closer to managing private workers.

You write the goal.
The agent edits files.
It runs commands.
It fixes mistakes.
You review the output.

Very efficient.

But also weirdly isolating.

The interesting part is that Anthropic noticed this inside the Claude Code team itself.

Their response was not “use less AI.”

It was:
bring back shared maker time,
programming lunches,
hackathons,
and watching how other people use agents.

That is probably the next real skill.

Not just prompting an agent.

Learning how other people manage agents.

[Frander @Frandeeer]

Fiona Fung says Claude Code made engineers lonelier.

That sounds like a soft problem.

It is not.

AI agents are changing coding from a team activity into something closer to managing private workers.

You write the goal.
The agent edits files.
It runs commands.
It fixes mistakes.
You review the output.

Very efficient.

But also weirdly isolating.

The interesting part is that Anthropic noticed this inside the Claude Code team itself.

Their response was not “use less AI.”

It was:
bring back shared maker time,
programming lunches,
hackathons,
and watching how other people use agents.

That is probably the next real skill.

Not just prompting an agent.

Learning how other people manage agents.

[Dep @0xDepressionn]

https://t.co/5RvUJw0WVd

[Frander @Frandeeer]

@JeffKazzee That's right, my friend, the code is great, but it's a bit weird at times to be honest.

[Frander @Frandeeer]

Nate Herk ran GLM-5.2 inside Claude Code all day.

That is the real story.

Not “Chinese model beats everything.”
Not “Claude is dead.”

The bigger shift is that Claude Code is becoming a harness.

You can keep the workflow:
plan,
edit files,
run commands,
debug,
iterate.

But swap the model underneath.

https://t.co/rbQAIYhkpC says GLM-5.2 supports 1M context and is built for long-horizon coding agents.

Its API price is also hard to ignore:
$1.40 input / $4.40 output per 1M tokens.

So the question changes.

It is no longer:
“Which model is the best?”

It is:
“Which tasks actually deserve the most expensive model?”

For critical production work, I still want the most reliable frontier model.

But for agents, experiments, repo exploration, refactors, research, and long coding sessions, GLM-5.2 inside Claude Code is a very different kind of pressure.

The model layer is becoming replaceable.

The agent workflow is becoming the product.

[Design Arena @Designarena]

https://t.co/JSn0lDCNkB

[Frander @Frandeeer]

@Rossst_03 You're welcome, happy to help.

[Frander @Frandeeer]

Nate Herk ran GLM-5.2 inside Claude Code all day.

That is the real story.

Not “Chinese model beats everything.”
Not “Claude is dead.”

The bigger shift is that Claude Code is becoming a harness.

You can keep the workflow:
plan,
edit files,
run commands,
debug,
iterate.

But swap the model underneath.

https://t.co/rbQAIYhkpC says GLM-5.2 supports 1M context and is built for long-horizon coding agents.

Its API price is also hard to ignore:
$1.40 input / $4.40 output per 1M tokens.

So the question changes.

It is no longer:
“Which model is the best?”

It is:
“Which tasks actually deserve the most expensive model?”

For critical production work, I still want the most reliable frontier model.

But for agents, experiments, repo exploration, refactors, research, and long coding sessions, GLM-5.2 inside Claude Code is a very different kind of pressure.

The model layer is becoming replaceable.

The agent workflow is becoming the product.

[Frander @Frandeeer]

GLM-5.2 is the Chinese AI model Silicon Valley is suddenly watching.

Not as a toy.
As a real open-source model for coding agents.

https://t.co/rbQAIYhkpC built it for long-horizon engineering work:
large codebases, agent loops, multi-step tasks, and 1M-token context.

That is the part founders care about.

If an open model is good enough for real coding work, it changes the economics of agents:

cheaper experiments,
longer runs,
more self-hosted stacks,
less dependence on one closed provider.

The signal is simple:

open models are moving from “interesting” to actually usable.

[Frander @Frandeeer]

GLM-5.2 is the Chinese AI model Silicon Valley is suddenly watching.

Not as a toy.
As a real open-source model for coding agents.

https://t.co/rbQAIYhkpC built it for long-horizon engineering work:
large codebases, agent loops, multi-step tasks, and 1M-token context.

That is the part founders care about.

If an open model is good enough for real coding work, it changes the economics of agents:

cheaper experiments,
longer runs,
more self-hosted stacks,
less dependence on one closed provider.

The signal is simple:

open models are moving from “interesting” to actually usable.

[MIKE @mikenevermiss]

https://t.co/FPnpzQAtb5

[Frander @Frandeeer]

@VoltexGar The bottleneck isn’t the number of agents. It’s the quality of the verification loop.

[Frander @Frandeeer]

@theo What the hell kind of settings are you using for your cloud? You're crazy, man.

[Frander @Frandeeer]

@Voxyz_ai It's actually really impressive; I took a look at it, and it looks pretty good, considering that I try to come up with cool designs

[Frander @Frandeeer]

@Shivam25mishra lol, it's really scary, especially when it shows 0 balance on the card

[Frander @Frandeeer]

@Alan_Earn unfortunately, no

[Frander @Frandeeer]

@YashHustle_22 Not likely, my friend, I think the opposite

[Frander @Frandeeer]

GLM-5.2 vs Claude Opus 4.8.

The gap is no longer “open model vs frontier model.”
It is starting to look like price vs reliability.

Z.аi claims GLM-5.2 scores 81.0 on Terminal-Bench 2.1.
Claude Opus 4.8 scores 85.0.

That is still a real gap.

But the pricing gap is much bigger:
GLM-5.2: $1.4 input / $4.4 output per 1M tokens
Opus 4.8: $5 input / $25 output per 1M tokens

So the practical question is not:
“Is GLM better than Opus?”

It is:
“How much extra reliability are you buying with Opus?”

For high-stakes agentic coding, I would still trust Opus first.

For experiments, self-hosted stacks, long-context codebase work, and cost-sensitive agents, GLM-5.2 is suddenly hard to ignore.

[Frander @Frandeeer]

GLM-5.2 vs Claude Opus 4.8.

The gap is no longer “open model vs frontier model.”
It is starting to look like price vs reliability.

Z.аi claims GLM-5.2 scores 81.0 on Terminal-Bench 2.1.
Claude Opus 4.8 scores 85.0.

That is still a real gap.

But the pricing gap is much bigger:
GLM-5.2: $1.4 input / $4.4 output per 1M tokens
Opus 4.8: $5 input / $25 output per 1M tokens

So the practical question is not:
“Is GLM better than Opus?”

It is:
“How much extra reliability are you buying with Opus?”

For high-stakes agentic coding, I would still trust Opus first.

For experiments, self-hosted stacks, long-context codebase work, and cost-sensitive agents, GLM-5.2 is suddenly hard to ignore.

[Frander @Frandeeer]

@Rossst_03 thx, bro

[Frander @Frandeeer]

WSJ just dropped a brutal investigation into Polymarket.

1,100+ staged videos.
Fake trading sites.
Fake accounts.
140M+ views.
One student alone showed 145 fake bets worth almost $410K.
The investigation itself looks real.
But the timing is way too convenient.

WSJ sits under News Corp. Fox — controlled by the same Murdoch family — now runs Kalshi data across its networks.

CNBC also signed a multi-year partnership with Kalshi.

Meanwhile Kalshi is valued at $22B, reportedly crossed $2B in annualized revenue, and is already holding early IPO talks.

Do I have proof Kalshi ordered the article?

No.

But when your competitor is preparing for public markets and the media ecosystem around it suddenly starts shaping the narrative against you, calling it random feels naive.

This looks a lot like clearing the field before an IPO.

[Frander @Frandeeer]

Watched a Codex demo and the interesting part is not “AI writes code”.

That’s already obvious.

The real shift is that Codex is starting to look like an execution layer for AI systems.

You don’t just ask it to build a random app.

You give it a goal, a structure, constraints, local context, and approval points — and it starts turning the idea into working pieces:

server logic
API calls
chatbot behavior
content agents
reports
local app structure
step-by-step implementation

The most useful part is not the code itself.

It’s the way the task gets decomposed.

One example from the video:

instead of “write me a chatbot”, the better structure is:

what the bot should do
what data it should use first
when it should call AI
how it should format answers
where human approval is needed

That’s a very different way to build.

The future of AI coding is not just faster programmers.

It’s people who can describe systems clearly enough for agents to execute them.

Bad prompt → random code.

Clear system → useful agent.
YouTube video channel - @JulianGoldieSEO

[OpenAI Developers @OpenAIDevs]

Show Codex a workflow once. Reuse it as a skill.

Record & Replay lets you show Codex a recurring task, like filing an expense report or submitting a time-off request.

Codex turns that demo into an inspectable, editable skill.

You control when recording starts and stops.

[Frander @Frandeeer]

Fable 5 might be coming back.

But the bigger story is not one model.

It’s what happens when models start working like teams.

[Kalshi Finance @Kalshi_Finance]

BREAKING: Trump says negotiations with Anthropic to restore access to its latest AI models are going fine