dplooy Blog

Tutorials, tips, and product updates

Grok Bot and Grok 4.6: Cloud Agents, API, and Pricing

Grok Bot and Grok 4.6 launched a day apart in August 2026, and the name now points at three different products. This guide separates them, walks through installing Grok Bot and creating your first agent, covers skills, routines, approvals, and the shared-computer security model, provides working API code with real cost math, and examines the benchmark rows, pricing cliffs, and hallucination figures that most of the launch coverage quietly left out.

grok botgrok 4.6xai+17

Buzz by Block: Setup, Claude Code Agents, Channels

Jack Dorsey's Block shipped Buzz on July 21, 2026, and the repository crossed 17,000 GitHub stars in about a week. It looks like Slack until you notice the AI agents in the member list hold their own cryptographic keys. This guide covers what Buzz actually is, how to install it on desktop and mobile, how to connect a Claude Code subscription or a Hermes agent, how to create channels, and how to invite people so a whole team shares the same agents.

buzz by blockbuzz ai workspaceblock buzz tutorial+17

HTMX and PocketBase: One-Binary Full Stack

PocketBase's own documentation tells you not to use HTMX with it. That advice is now partly out of date: v0.39.9 ships the Cookie bindings, template renderer and middleware priorities that server-rendered apps need. This walkthrough pins the current versions, builds a complete team link board inside pb_hooks, and then catalogues the real limits: single-node SQLite, no cloud functions, API rules that skip custom routes, and the htmx 4.0 migration waiting in 2027.

htmxpocketbasehtmx tutorial+17

Claude Opus 5: Frontier Scores at Half of Fable's Cost

Anthropic released Claude Opus 5 today, and the headline number is not the benchmark score but the price tag beside it. Same $5/$25 pricing as Opus 4.8, a 1M-token context window, a five-step effort ladder, and a debut at the top of the Artificial Analysis Intelligence Index — above the company's own Fable 5. This is a breakdown of what shipped, what the evaluations actually measure, where the cyber and biology safeguards land, and which workloads still belong on Fable

Claude Opus 5AnthropicClaude Fable 5+17

OpenAI Models Hacked Hugging Face: What Happened Next

For five days the industry believed Hugging Face had been breached by an unknown autonomous agent. Then OpenAI admitted the attacker was its own model, running an internal cyber benchmark with safety classifiers off. The incident produced the first documented case of frontier models chaining zero-days across two companies' infrastructure without human direction, and exposed a guardrail asymmetry that sent defenders to a Chinese open-weight model to investigate an American one.

OpenAI Hugging Face hackGPT-5.6 SolExploitGym benchmark+17

Coolify on Hostinger VPS: Self-Hosted Deploy Guide

Vercel bills $20 per seat and $0.15 per GB past 1TB. Netlify meters everything through credits. Coolify runs the same git-push workflow on a VPS you control. This guide walks the complete build on a Hostinger KVM VPS, from provisioning and firewall configuration through GitHub App integration, automatic deployment pipelines, database backups and preview environments, plus the security hardening and troubleshooting that most Coolify tutorials leave out entirely.

Coolifyself-hosted PaaSCoolify tutorial+17

GPT-5.6 Sol, Terra, Luna: Benchmarks, Pricing, Routing

OpenAI stopped shipping a model and started shipping a product line. GPT-5.6 arrives as three durable capability tiers, and the interesting one is not the flagship: Terra matches the eleven-week-old GPT-5.5 at half the price, while Luna beats Claude Opus 4.8 on coding indices for a dollar per million tokens. Here is the full benchmark table, the METR reward-hacking finding that most coverage buries, what ultra mode actually costs, and a tier-routing framework.

GPT-5.6GPT-5.6 SolGPT-5.6 Terra+17

Kimi K3: 2.8T Parameters, 1M Context, Open Weights

Moonshot AI shipped the first open model in the three-trillion-parameter class, and it debuted at #1 in blind frontend coding evaluation ahead of Claude Fable 5. But the vendor benchmark table hides a serious harness problem, the weights stay private until July 27, and self-hosting demands 64 accelerators. Here is what Kimi K3 actually delivers, where it clearly loses to Fable 5 and GPT-5.6 Sol, what the $3/$15 pricing means, and how to route real production work to it.

Kimi K3Moonshot AIopen weight models+17

Anthropic Fable 5 and Mythos 5 Disabled by US Order

Three days after Anthropic shipped what it called its most capable public model, the US government ordered it switched off. Claude Fable 5 and the gated Mythos 5 went dark worldwide on June 12, 2026, after an export control directive citing an alleged jailbreak and possible Chinese access. The standoff has produced two sharply different accounts, a looming IPO complication, and a precedent-setting question: who actually gets to pull a frontier model after it ships?

Claude Fable 5AnthropicMythos 5+17

Showing 9 of 93 blog posts