AI Engineer | Working with Agents I build AI explainers with code deep dives on YouTube DM for Collabs
Joined November 2020
- Tweets11.2K
- Following203
- Followers3.3K
- Likes13.1K
Pinned Tweet
🎁 New users: $10 free**
👥 Refer a friend: **$6
💸 Top up $1 → get $10 free ($11)
💸 Top up $10 → get $100 free ($110)
⚡ DeepSeek open model: $1 ≈ 30M tokens
🔥 ~$3/month = token freedom
👉 Sign up: [threerouter.com/register?aff…]
Opus 4.6 finished my task in 40 minutes.
Opus 5.5 did it in 15 and cost me the whole week.
Anyone else notice?
Opus 5.5 will let you run an entire security audit, burn all your tokens, then block the final report at the very end.
The classifier doesn't stop you upfront. It just quietly eats your session bar and refuses the output at 99%.
If you're gonna block it, block it BEFORE you spend 50% of your quota. not after.
New Gemini 4 Carbon leaks from LMArena testers and Google DeepMind insiders today
1. Huge leap in 3D scene generation & spatial geometry
2. Slower, deeper thinking
3. Tested on Google's internal "Jetski" coding platform
Google just dropped the Gemini API CLI and it kills the biggest token-burn trap in Claude Code, Codex and Cursor.
Right now your agent writes a throwaway SDK script and dumps raw base64 straight into your context.
gemini-api fixes it:
- Images, video, TTS and music go straight to disk, printing ONLY the path
- "toon" output + --jq filtering so responses barely touch context
- 100% offline --usage, --schema, --dry-run so it never guesses model IDs
Add 1 line to your AGENTS.md and you're done.
Opus 5.5 usage check
Last week: felt unlimited.
Today: 100% of a Max (5x) session gone in 33 minutes.
Now locked out for 4 hr 27 min.
Anyone else?
October 2026 SlopCodeBench just exposed something ugly about “xhigh” effort on coding agents.
They ran Opus 5.5, Fable 5.1, GPT-6.1 Sol and GPT-6 Astra on multi-stage coding tasks where models have to iterate without breaking previous tests:
• xhigh cost 2–4x more than Medium while barely improving (and sometimes hurting) strict pass rates
• GPT-6.1 Sol Medium ($178) beat Fable 5.1 ($1,550 8.7x more expensive) on strict checkpoints
• Opus 5.5 Medium led raw feature completion (145/196) but ranked worst for “slop” leaving 45 junk `_work` files (41,265 lines) in a single run
Are you overpaying for xhigh, or is Medium actually the smarter default?
Gemini 4 Argon
AI agent swarms just uncovered a 12-year-old 9.2-CVSS zero-day in core PHP (PHP2Shell / CVE-2026-109618) while offensive bot swarms race to exploit unpatched servers.
I built an interactive 3D simulation of the battle in a single HTML file:
URGENT: if you host AI apps on Railway or use LiteLLM go revoke your Claude, OpenAI, and OpenRouter API keys right now
Railway Env Breach: Attackers accessed decrypted environment variables and are actively draining Claude & OpenRouter API keys.
LiteLLM 1.107.0 Malware: Version 1.107.0 was poisoned with a credential stealer (downgrade to 1.106.2 and rotate all keys).
Check your Anthropic & OpenAI billing usage logs immediately.
An urgent notice from the Railway team
blog.railway.com/p/security-…
Tibo confirms Codex / Work reset lands tomorrow before 10am PST
Local times:
🇺🇸 San Francisco → before 10:00 AM
🇺🇸 Seattle → before 10:00 AM
🇺🇸 New York → before 1:00 PM
🇺🇸 Chicago → before 12:00 PM
🇬🇧 London → before 6:00 PM
🇮🇳 Delhi / Mumbai / Bangalore → before 10:30 PM
🇸🇬 Singapore → before 1:00 AM (next day)
🇯🇵 Tokyo → before 2:00 AM (next day)
🇦🇺 Sydney → before 4:00 AM (next day)
Fable 5.5 rumors.
Gemini 4 Carbon rumors.
A very real $500/month ChatGPT Pro tier.
AI labs are speedrunning my bankruptcy before November.
Which subscriptions are you keeping and which are getting cancelled?
Stop letting ONE AI usage bar decide when you can code.
Opus 5.5 → architecture + deep bugs
Gemini 4 Argon → multi agent set up
Codex (GPT-6.1 Sol) → verification
Spread the workload. Keep a fallback when one hits its cap.
What’s your stack?
Opus 5.5 on Max all day can torch your weekly limit.
Try this in Claude Code:
Max/xhigh → map dependencies + plan
High/Medium → implement
Switch with /effort in the SAME chat.
What’s your combo?
Gemini 4 Argon: Combined 3 biomechanical & anatomical layers into a single zero-build Three.js / WebGL simulation
I'm 4+ days into my weekly cycle on Ultra and still have 81% of my weekly limit
Meanwhile on the $20 Pro tier, the Argon cap is brutal:
- A single ~190k-context task eats 20%–23% of the entire weekly limit
- Even a short prompt burns ~9%–10% of the 5-hr bar
- Devs are hitting 0% lockout in 4 to 7 tasks
On Ultra, the Argon limit feels generous.
On $20 Pro, it's basically a 5-prompt demo.
GPT-5.6 Sol and Gemini 4 Argon just independently uncovered two 12-year-old zero-days in PHP's core C runtime ("PHP2Shell" CVE-2026-109618 & CVE-2026-109619):
1. GPT-5.6 Sol (Calif): Using standard Sol with zero custom scanners, it traced a C longjmp() OOM bug during multipart parsing that leaks live server memory and triggers RCE.
2. Gemini 4 Argon + GPT-5.6 Sol (Wiz Research): Independently audited the PHP repo and caught both CVEs including a persistent worker Use-After-Free across PHP 5.4–8.5.
OpenAI models helped researchers—operating as part of our Patch the Planet initiative—discover critical security flaws in PHP, an open-source programming language used to run an estimated 70% of websites.
Over the past two months, we led a coalition of maintainers, researchers, and security providers to prepare fixes and protections ahead of disclosure. This duration was necessary because the fix was quite intricate and required a major change to the PHP engine. Every site using PHP can now safely upgrade their version to stay secure, in addition to applying mitigations from various security vendors.
Due to the severity of these flaws and PHP’s widespread use, we consider this bug to be one of the most impactful vulnerabilities in recent memory. It also marks a new phase, where AI can find security bugs with internet-scale impact.
This is an example of the Defender’s Window in action: to our knowledge, frontier AI was able to find this vulnerability ahead of attackers, giving defenders time to prepare.
openai.com/index/coordinated…
Readers added context they thought people might want to know
OpenAI may claim a role in the coordinated response and patching process. But the initial discovery and exploit development, however, occurred several months earlier by an independent researcher.
x.com/lcfr_hax/statu…
OpenAI just announced:
GPT-6 Astra
GPT-6.1 Sol
GPT-6 Luna and Codex
are coming natively to Google Cloud Vertex AI and Google Antigravity.
To get your GCP project in the preview queue now: open Google Cloud Console → Vertex AI → Model Garden and submit the preview access form.
cloud.google.com/blog/produc…
