Vanja Petreski

A steel vault door stands ajar in a dark data center; blue light spills through the gap, where streaks of light race along a curving track between rows of glowing server racks

ChatGPT Pro Max: The $500 Price Leaked, the Limits Didn't

OpenAI hasn't announced it, but a $500-a-month Pro Max plan has been spotted in ChatGPT's subscription screen, pitched as "Fastest Work and Codex". What the card shows, what OpenAI's own rate card already says about speed and allowances, and what to check at DevDay.

Three glowing machine cores on a wet steel floor at night: a blue star, an amber sun and a silver crescent moon, with a balance scale standing between the star and the sun

GPT-6 Sol and Luna: Cheaper Than Astra, Not Better

A fifth of Astra's price for Sol, a hundredth for Luna. My rule for the Codex seat, written before the charts: switch only to a model that beats Astra at maximum settings and costs less. Sol costs less. On every chart OpenAI published, it doesn't beat Astra.

Three glowing machine cores on a steel rack in a night workshop; the violet one is mid-swap, a heavy core lifted out on chains above a smaller, brighter core seated in its socket

Opus 5.5: Fable's Level at 40% of the Price

Anthropic shipped Opus 5.5 on September 22: by its own numbers, Fable 5.1's level on most work, at $4/$20 per million tokens against Fable's $10/$50, with medium as the new default effort. What the release measures, where the catch is, and how my Claude seat left Fable the same afternoon.

Three glowing machine cores on a rain-soaked workbench at night, the centre one being hoisted into place, lit in violet and blue

Grok 4.7: Same Price, Longer Runs, One 402

xAI shipped Grok 4.7 today: a larger base model, 500k context, xhigh still the ceiling, same $2/$6 price, and a few contradictions in its own pages. What the model card measures, and how every Grok pin on my desk moved to 4.7 the same afternoon, balance wall included.

Hooded author portrait in a rain-soaked machine workshop lit in violet and blue

The Solver Was Always the Hard Part

A year as founding principal AI-native engineer at a stealth healthcare-scheduling startup: therapy schedules for nursing facilities, OR-Tools and dynamic programming, HIPAA without the theater, Terraform on Google Cloud, and three AI agents under one rulebook. A case study.

Jev: Keep the Power. Cut the Overhead.

Jev turns natural language into choices, scores and probabilities. A practical guide to the model, the harness, the demos—and what it could change for my three-agent setup.

A closed laptop on a dark desk while a cloud chassis of screens keeps working — AI

Grok Bot Is a Computer That Stays On

Named teammates on a persistent cloud computer. SuperGrok Heavy includes it. How it works, what it costs, and what to hire it for.

Two compact industrial cores docking into a dark terminal chassis in a hangar — AI

DeepSeek, GLM, and OpenCode

OpenCode is an open-source coding agent. DeepSeek and GLM supply models; subscriptions start at $10–$18/month with usage allowances, separate from any billed overflow.

GPT Image 2.5: Pin Sunburst

Two GPT Image 2.5 models, same token rates. Flare is the volume default. Sunburst is the precision one. Pin that when the picture is the product.

Claude Mods

Anthropic's proposed TypeScript middleware for Claude Code: how the continuation model works, the proposed API in full, worked examples, and the open questions on availability and cost.

Pace the Frontier. Name the Speed Limit.

A model can get promoted without receiving a single new weight — just a token, a runtime, and permission to spawn workers. Amodei is right that time is worth buying.

Brain Rot Is the Default Setting

AI assistance can reduce practice when it replaces the thinking a task requires. Research points to task-specific risks and ways to preserve unaided skills.