I bought an 18-month Gemini subscription for $1
I was curious what was even inside, though the Gemini models are trash imo
Keep seeing the same thing with ChatGPT, Cursor and Gemini is a king of abuse
People build entire stores around abusing promo campaigns and enterprise subscriptions, then resell access for almost nothing. This has become a way to make money for a lot of people
Weirdly, it also gives AI access to people who can’t get it for their own reasons
Companies keep trying to fight the abuse but the market for these subscriptions is still incredibly large
The only one I can single out is Claude. I rarely see anyone offering it significantly cheaper than a normal subscription. Maybe Anthropic is handling the abuse better for now
Scam Altman at your service
Same task. Same conditions.
> GPT Astra worked nonstop for nearly 30 hours and delivered roughly 20% of the implementation.
> Sol 5.6 orchestrating two or three 5.5 subagents used half the tokens, completed the entire task and finished roughly three times faster.
I genuinely don’t know what the hell is going on here. Maybe I missed a setting. Maybe GPT Astra just burns your money. I ran it as an experiment to see a scam on a global scale, not as proof that I’d already found one.
The whole internet is shocked by the 3D models of houses and terrain it can generate. But who is actually going to verify the real performance?
I definitely wouldn’t switch everything to GPT-6 Astra
The September Intelligence Index from Artificial Analysis gives a neutral snapshot instead of vendor marketing:
> GPT-6 Astra sits at Index 61. It leads in computer use and agentic automation, with additional strength in mathematics and cybersecurity. Terminal-Bench 4.0 shows the gap with 57.9% for Astra, 55.8% for Fable and 19% for Gemini.
Astra costs $10/$50 per million tokens. The headline price is high. Greater efficiency can still make the completed task cheaper if the alternatives need more tokens or attempts to finish it
> Claude Fable 5.1 leads overall intelligence at Index 66. It also scores 70 on the Coding Agent Index versus 67 for Astra. Both charge $10/$50 but Fable’s cache costs four times less at $0.25 versus $1. That difference gets noticeable when an agent runs for a long time
> Claude Opus 5 is another strong choice. Index 63 with input tokens at $5, half Astra’s input price
> Then there is Gemini 3.8 Flash. Index 59 and $0.75/$3.75 pricing through the end of the year. Hard to ignore when cost decides the job
So I’d choose Astra when the agent must monitor services, click through interfaces, complete forms, or operate websites. Software operation is its one meaningful edge in this comparison.
For general intelligence, research and long reasoning chains, Claude Fable 5.1 or Claude Opus 5 makes more sense. Both rank higher on the aggregate and Opus is cheaper. If the goal is simply running a large volume of requests cheaply, I’d still choose Claude.
> For actually putting an agent behind a screen – Claude Cowork + Computer Use is the easiest place to begin
Claude first tries connectors such as Slack, Gmail and Notion. If none can do the job, it falls back to the browser and starts interacting with the interface
> Cua at cua.ai is good when handing an agent your computer still feels terrifying
It runs inside a sandboxed virtual machine and cannot damage your primary computer. Much easier for the first cautious experiments
> OpenClaw/Hermes goes in another direction, with a local assistant running on a VPS 24/7, controlled through Telegram. Give it scheduled autonomous work and let it keep going in the background
No need to wait for OpenAI’s flagship. Claude Cowork, Cua and OpenClaw/Hermes can take a real task right now
Is this the end?
Replying to @ChaseLochmiller @OpenAI
GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years.
AGI has arrived. Congratulations @OpenAI team.
400K GPUs coming online next.
6
124
I definitely wouldn’t switch everything to GPT-6 Astra
The September Intelligence Index from Artificial Analysis gives a neutral snapshot instead of vendor marketing:
> GPT-6 Astra sits at Index 61. It leads in computer use and agentic automation, with additional strength in mathematics and cybersecurity. Terminal-Bench 4.0 shows the gap with 57.9% for Astra, 55.8% for Fable and 19% for Gemini.
Astra costs $10/$50 per million tokens. The headline price is high. Greater efficiency can still make the completed task cheaper if the alternatives need more tokens or attempts to finish it
> Claude Fable 5.1 leads overall intelligence at Index 66. It also scores 70 on the Coding Agent Index versus 67 for Astra. Both charge $10/$50 but Fable’s cache costs four times less at $0.25 versus $1. That difference gets noticeable when an agent runs for a long time
> Claude Opus 5 is another strong choice. Index 63 with input tokens at $5, half Astra’s input price
> Then there is Gemini 3.8 Flash. Index 59 and $0.75/$3.75 pricing through the end of the year. Hard to ignore when cost decides the job
So I’d choose Astra when the agent must monitor services, click through interfaces, complete forms, or operate websites. Software operation is its one meaningful edge in this comparison.
For general intelligence, research and long reasoning chains, Claude Fable 5.1 or Claude Opus 5 makes more sense. Both rank higher on the aggregate and Opus is cheaper. If the goal is simply running a large volume of requests cheaply, I’d still choose Claude.
> For actually putting an agent behind a screen – Claude Cowork + Computer Use is the easiest place to begin
Claude first tries connectors such as Slack, Gmail and Notion. If none can do the job, it falls back to the browser and starts interacting with the interface
> Cua at cua.ai is good when handing an agent your computer still feels terrifying
It runs inside a sandboxed virtual machine and cannot damage your primary computer. Much easier for the first cautious experiments
> OpenClaw/Hermes goes in another direction, with a local assistant running on a VPS 24/7, controlled through Telegram. Give it scheduled autonomous work and let it keep going in the background
No need to wait for OpenAI’s flagship. Claude Cowork, Cua and OpenClaw/Hermes can take a real task right now
The marathon continues
OpenAI showed off its new model today – GPT-6 Astra. Subscribers are promised access in the next few days. Also there gonna be a lot of usage resets
Benchmarks, blah blah. Only nerds and investors jerk off to those numbers anyway and half of them are made up
Let’s wait a week or two and see Claude fuck Astra up
Although Claude probably won’t even need to. Right before the next release the current model will start acting dumb as usual so the new one looks unbelievably smart
The marathon continues
OpenAI showed off its new model today – GPT-6 Astra. Subscribers are promised access in the next few days. Also there gonna be a lot of usage resets
Benchmarks, blah blah. Only nerds and investors jerk off to those numbers anyway and half of them are made up
Let’s wait a week or two and see Claude fuck Astra up
Although Claude probably won’t even need to. Right before the next release the current model will start acting dumb as usual so the new one looks unbelievably smart
I just don’t believe Gemini models are actually useful right now
Gemini 3.8 Flash has great benchmarks. Their updates still feel like Google pretending to work. They stopped trying to compete with Claude and went off in their own direction.
And Gemini still has no visible adoption beyond abuse and people buying 18-month subscriptions for $1.
Anyway, another Gemini release shows up and I just skip it now
I bought an 18-month Gemini subscription for $1
I was curious what was even inside, though the Gemini models are trash imo
Keep seeing the same thing with ChatGPT, Cursor and Gemini is a king of abuse
People build entire stores around abusing promo campaigns and enterprise subscriptions, then resell access for almost nothing. This has become a way to make money for a lot of people
Weirdly, it also gives AI access to people who can’t get it for their own reasons
Companies keep trying to fight the abuse but the market for these subscriptions is still incredibly large
The only one I can single out is Claude. I rarely see anyone offering it significantly cheaper than a normal subscription. Maybe Anthropic is handling the abuse better for now
2
197
BTW
Shared cloud sessions might be OpenClaw’s most important new feature. You can join an agent’s work or hand off a task without losing the context it already has.
Everyone expects agents to move into the cloud. OpenClaw already looks ready for agents working across several people instead of sitting inside one person’s local setup.
The change inside OpenClaw says even more. The entire project team moved from local harnesses to a shared agent. Now they use team.openclaw.ai, which sees what everyone is working on and coordinates between them.
Once AI agents hit the limit of working with one user at a time, the work moves to communication between whole teams.
I expected humans to lead teams of agents. Instead, agents are starting to coordinate teams of humans.
Who is still using OpenClaw?
OpenClaw 2.0 is officially out. The local AI agent that runs on your own machine and talks to you through Telegram now has 16,000 pull requests, 933 contributors and 569 first-time contributors behind it.
The full release is at docs.openclaw.ai. In short, OpenClaw is much easier to install and much less likely to leak something you never meant to show it.
It now takes two npm commands. During setup it automatically detects your ChatGPT or Claude subscription, existing API keys and local models including Ollama, LM Studio and llama.cpp. It also tests the model connection before saving the configuration. Small detail, except the old flow could happily save a broken setup and leave you to figure out why nothing worked.
The chat itself got a full rebuild. Sessions sit in the center now. There is a command palette, transcript search, archiving and the option to fork a session from any message. You can branch off halfway through a conversation without destroying the useful version you already had.
Mobile sync works through a QR code. Full and limited access modes are clearly labeled. OpenClaw asks for permissions individually so you can see what you are actually giving it. Sessions and transcripts have also moved to SQLite, which should make the whole thing less fragile.
The best change is how it handles secrets.
When the agent needs a key or password, you enter it through a separate masked window. The value never enters the conversation or the model context. Previously those keys could settle into logs that might later be sent to a third-party API. A local agent with access to your machine should never treat passwords like another chat message. OpenClaw 2.0 finally handles this properly.
Why use any of this when ChatGPT already works in a browser?
Because a browser chat does not live on your machine. A local agent can work with your local files, run scripts and connect to your servers. ChatGPT can tell you what command to run. OpenClaw can run it, inspect the result and continue working.
Still, OpenClaw and Hermes are two flavors of shit compared with a well-configured custom solution. A custom setup gives you better control over models, permissions, tools and where your data goes.
For 97% of people, these out-of-the-box agents are more than sufficient.
Now we just have to wait for Hermes to get better.
1
1
187
Who is still using OpenClaw?
OpenClaw 2.0 is officially out. The local AI agent that runs on your own machine and talks to you through Telegram now has 16,000 pull requests, 933 contributors and 569 first-time contributors behind it.
The full release is at docs.openclaw.ai. In short, OpenClaw is much easier to install and much less likely to leak something you never meant to show it.
It now takes two npm commands. During setup it automatically detects your ChatGPT or Claude subscription, existing API keys and local models including Ollama, LM Studio and llama.cpp. It also tests the model connection before saving the configuration. Small detail, except the old flow could happily save a broken setup and leave you to figure out why nothing worked.
The chat itself got a full rebuild. Sessions sit in the center now. There is a command palette, transcript search, archiving and the option to fork a session from any message. You can branch off halfway through a conversation without destroying the useful version you already had.
Mobile sync works through a QR code. Full and limited access modes are clearly labeled. OpenClaw asks for permissions individually so you can see what you are actually giving it. Sessions and transcripts have also moved to SQLite, which should make the whole thing less fragile.
The best change is how it handles secrets.
When the agent needs a key or password, you enter it through a separate masked window. The value never enters the conversation or the model context. Previously those keys could settle into logs that might later be sent to a third-party API. A local agent with access to your machine should never treat passwords like another chat message. OpenClaw 2.0 finally handles this properly.
Why use any of this when ChatGPT already works in a browser?
Because a browser chat does not live on your machine. A local agent can work with your local files, run scripts and connect to your servers. ChatGPT can tell you what command to run. OpenClaw can run it, inspect the result and continue working.
Still, OpenClaw and Hermes are two flavors of shit compared with a well-configured custom solution. A custom setup gives you better control over models, permissions, tools and where your data goes.
For 97% of people, these out-of-the-box agents are more than sufficient.
Now we just have to wait for Hermes to get better.
AI coding agents are dopamine slot machines
You sit there late at night, coding with an agent. Hours disappear. Your eyes are red, you’re exhausted and somehow you still can’t stop. The next attempt might fix it. Then it finally runs. You get the click of "it worked" and keep going instead of sleeping.
Psychologists have known this mechanism as variable ratio reinforcement since the 1950s. The reward arrives after an unpredictable number of attempts. Casinos run on it. Instagram and TikTok run on it. Now coding with AI agents does too.
Productivity and the dopamine loop are two independent things that can exist in parallel.
The tool can genuinely help you finish more work. The loop can also keep you glued to the screen. Instagram occasionally gives you useful news, a good link or an answer you needed. The useful post doesn’t make the feed any less addictive.
Traditional coding gave you two rewards. You figured out how to solve the problem then watched your solution run. With AI coding, the model often does the figuring out. You mostly wait, run the result and see whether it works. Less satisfaction from understanding the solution. More frequent little hits from execution.
Getting hooked does not mean being ineffective. You can build faster, finish real work and still burn yourself out waiting for the next successful run.
Remember to breathe fresh air and touch grass between building multibillion-dollar products.
Perplexity and Nvidia just put the entire agent stack on your machine, with zero token costs for local work
Portable Computer is a local version of Perplexity Computer: the orchestrator, planner, tool router, search index and subagents
All run on your hardware using compact Qwen 3 and Nvidia Nemotron models
Every step handled locally costs zero tokens. When the local model can’t finish something, it asks permission before using a larger cloud model. Projects, files and pinned notes persist across sessions, while device data stays local, including sensitive information like participants’ emails.
The catch is the hardware. Portable Computer launches on Nvidia DGX Spark with Linux. Support for RTX 3090, RTX 4090 and RTX 5090 GPUs with at least 24 GB of VRAM is promised soon. You’ll also need a Pro or Max subscription.
A decent update but the zero-token claim sounds funny when entry requires a separate, extremely expensive machine and a paid subscription on top
Perplexity and Nvidia just put the entire agent stack on your machine, with zero token costs for local work
Portable Computer is a local version of Perplexity Computer: the orchestrator, planner, tool router, search index and subagents
All run on your hardware using compact Qwen 3 and Nvidia Nemotron models
Every step handled locally costs zero tokens. When the local model can’t finish something, it asks permission before using a larger cloud model. Projects, files and pinned notes persist across sessions, while device data stays local, including sensitive information like participants’ emails.
The catch is the hardware. Portable Computer launches on Nvidia DGX Spark with Linux. Support for RTX 3090, RTX 4090 and RTX 5090 GPUs with at least 24 GB of VRAM is promised soon. You’ll also need a Pro or Max subscription.
A decent update but the zero-token claim sounds funny when entry requires a separate, extremely expensive machine and a paid subscription on top
When Codex searches the web, does it actually visit every result in real time?
The flow most people imagine:
> Codex sends the query to Google or Bing
> gets a list of matches
> a crawler immediately visits every website.
That’s not how the available pieces fit together.
Fresh web information does not come from the model’s weights. Codex invokes a separate web search tool. In local Codex, the default cached mode searches OpenAI’s existing index. It does not synchronously open every matching page.
There are also indexed and live modes. live can open a current page while Codex is working on a task. Opening that specific page still does not mean every result was crawled in real time.
The crawler names describe different jobs:
> OAI-SearchBot collects pages for ChatGPT Search
< ChatGPT-User may access a page after a specific user request
> GPTBot concerns whether content may be used for training
A site can allow OAI-SearchBot so its pages can appear in search while blocking GPTBot to opt out of potential training use. Search access and training access can be configured separately.
The remaining question is where the search results come from. OpenAI does not publicly identify the exact external search engine behind Codex or explain precisely how the cached index is populated.
So the common claim that Codex definitely always routes searches through Bing is speculation unless OpenAI discloses the provider and how its cached index is built
1
3
127
Your agent never has to forget what happened in your project
Agents recover context by rereading project knowledge. Architecture, plans, personal preferences, what changed yesterday, where the work stopped. How fast they catch up depends on how you store all of it
I use an LLM wiki for EVERY project
Each project has its own wiki, so I can switch between Claude, Codex and Kimi without explaining the whole project again. They all get the same architecture, plans, personal context and latest work state
You can throw plans, ideas and changelogs into random .md files. I did something more structured: the wiki tells the agent what to read, where to write and how to find previous decisions
To set it up, create a project wiki and ask the agent to initialize the required .md files plus the rules for using them. You can also add a skill that handles the workflow
Then the loop runs naturally. The agent does the work, records what changed, searches the wiki for related decisions during future tasks and uses them as context. The project keeps its memory even when you switch models
The prompt will be added below the post
Karpathy described the method and Claude already understands the pattern well enough to build it from one prompt
2
131
Initialize a local LLM wiki as living documentation inside the repository. Give it the project’s purpose, current state, architecture and operating principles, then keep decisions, change history and experiments there too. The goal is to understand both how the system works today and how it ended up this way.
Leave agents one short rule: whenever the code changes, update the wiki. Check every claim against source files and migrations first.
Do not present historical guesses as facts.
16
Scraping browser pages used to be the annoying part of my work
Playwright covers a lot of it but when you sstart to open any services, cloudflare shows up and everything stops.
And sometimes the agent doesn't get what it's supposed to do at all, it just keeps trying things.
Added some of these and it got noticeably easier.
Hosted, nothing to set up:
– Firecrawl – turns any page into clean markdown for the model, JS sites and PDFs included
– Jina AI Reader – put r.jina.ai in front of the URL and get the text back, free tier is generous
– Apify – marketplace of 28 thousand ready-made scrapers, for anything popular somebody already wrote it
– Thunderbit – Chrome extension, pulls a page into Sheets or Notion in two clicks
– RSS.app – turns sites and socials into an RSS feed with filters and autoposting into Telegram.
Open libraries, the ones the agent installs for itself
– Crawl4AI – hands the site to the model as markdown
– Browser Use – drives the browser like a person, logs in, clicks, takes what you asked for
– Crawlee – proxy rotation, retries and queues, for when there are really a lot of pages
– Scrapy – the industrial python classic for crawling millions of pages
– Scrapling – adjusts to layout changes on its own so it breaks noticeably less often
– AutoScraper – show it one example of the data you want and it works out the pattern and pulls the rest
– Defuddle – rips out ads, menus and sidebars, it is the engine behind Obsidian Web Clipper
– MarkItDown – Microsoft's, converts basically any file to markdown, PDF and Office and images and audio
When the site still fights back:
– curl_cffi – requests with a real browser fingerprint, simplest way to stop looking like a bot
– curl-impersonate – same trick at the curl level, it is the engine sitting under curl_cffi
– CloakBrowser – patched Chromium instead of Playwright, gets through antibot and captcha
– invisible_playwright – same thing on Firefox
– scrcpy – mirrors your Android screen onto the computer, for when the data only lives inside an app and there is nothing left to parse
5
2
243
This feature took Gemini two years
I pay for the service and they finally let me remove visible watermarks from generated images
The toggle is buried in profile settings and the watermark is still enabled by default
SynthID remains and the file metadata still labels the image as AI-generated
Most importantly, Gemini is still shit
Gemini 3.7 Flash, GLM 5.3 and free Manus access are the three AI updates worth looking at this week. Manus is the one you can use right now
> Google says Gemini 3.7 Flash is more accurate than Gemini 3.6 Flash
That would fix the old model’s annoying problems with dates and outdated knowledge. I haven't tested it yet and I'm not sure I will
But $0.75 per million tokens is absurdly cheap. If Google actually fixed the accuracy issues, this could be one of the best-value models available.
> GLM 5.3 looks like a code-focused evolution of GLM 5.2
They kept the base and trained it further for coding, vulnerability research and similar tasks. It also uses fewer tokens on long tasks. I think GLM 5.3 could become a strong competitor to current top models
> Then there is Manus, the Chinese agent that handles tasks and builds automations using fine-tuned Qwen and Claude models
Someone (won't tell who exactly) banned foreign investment in Manus. Now Manus has to move from overseas infrastructure to domestic servers. Existing users need to back up their account data and upload it again after the migration
The migration creates a rare little window for everyone else: Manus is completely free to use during the process.
Use Manus for free until August 23
3
168
Gemini 3.7 Flash, GLM 5.3 and free Manus access are the three AI updates worth looking at this week. Manus is the one you can use right now
> Google says Gemini 3.7 Flash is more accurate than Gemini 3.6 Flash
That would fix the old model’s annoying problems with dates and outdated knowledge. I haven't tested it yet and I'm not sure I will
But $0.75 per million tokens is absurdly cheap. If Google actually fixed the accuracy issues, this could be one of the best-value models available.
> GLM 5.3 looks like a code-focused evolution of GLM 5.2
They kept the base and trained it further for coding, vulnerability research and similar tasks. It also uses fewer tokens on long tasks. I think GLM 5.3 could become a strong competitor to current top models
> Then there is Manus, the Chinese agent that handles tasks and builds automations using fine-tuned Qwen and Claude models
Someone (won't tell who exactly) banned foreign investment in Manus. Now Manus has to move from overseas infrastructure to domestic servers. Existing users need to back up their account data and upload it again after the migration
The migration creates a rare little window for everyone else: Manus is completely free to use during the process.
Use Manus for free until August 23
A team of AI agents that assign roles in a shared chat and work together on tasks from their own cloud computers
Made by xAI – Grok Bot
One agent manages the job. The others handle their own pieces, talk through the shared chat, and keep going while you’re offline
Each bot gets a cloud computer and can click through websites and apps without API and MCP. You show it how you do something, it picks up the habit, then only bothers you when it needs approval
For now, only subscribers to SuperGrok Heavy, Cursor Ultra and Cursor Teams Premium can try it – the most expensive plans
3
134
I just reserved my Cloudflare Wallet tag: ktmudak.cloudflare.pay.
Reserve yours now at cloudflare.pay
i think its good idea
2
91
Meta unveiled Muse Spark 1.2 and Muse Code as a new direct competitor.
It handles the full repository loop by reading the code, drafting a plan, making changes, running tests and recovering after a crash. It can launch parallel subagents for architecture, tests, documentation and alternative approaches, all sharing the same session context.
Muse Code also keeps a local log of prompts, tool calls and edits, so you can resume from any point. Input can be text, images, PDFs or video. Output is text only.
Meta says it trained Muse Spark on long agent tasks like setting up a repo from scratch, debugging, researching and optimizing while keeping the goal intact and compressing context along the way.
The model is closed. No weights and no local deployment. Muse Code is still in beta and only works on macOS and Linux. You can also use it through the Meta Model API.
Context – 1,048,576 tokens.
> On Terminal-Bench 2.1, Claude Opus 5 scores 86.7%, Muse Spark 1.2 gets 82.9% and GPT-5.6 Terra gets 81.8%.
> DeepSWE 1.1 looks worse for Muse, with Claude Opus 5 at 65.0%, GPT-5.6 Terra at 64.8% and Muse Spark 1.2 at 59.3%.
There’s no independent verification of Meta’s numbers though, which smells a bit like bullshit.
API pricing is $1.25 per 1M input tokens, $4.25 per 1M output tokens and $0.15 for cached input.
There’s also a contributor tier advertised as twice as cheap, with $0.10 input and $0.20 output. Meta gets to use your data for training there, so it doesn’t fit corporate or confidential code.
research.meta.ai
Muse Code launched its public beta with a long feature list and no large-repository verdict
Muse Code can plan changes, write code, validate results and run parallel agents in isolated branches.
Sounds good. A messy repository full of old patterns and hidden dependencies is where those claims either hold up or fall apart.
No detailed large-repo success or failure stories have surfaced yet. People need to show what Muse Code changed, what it broke and what they had to fix before anyone treats it as proven.
So yeah, Muse Code looks promising. Large repositories still get an unproven for now
1
223



















