The Review Is Automated — the Decision Isn't

Two independent model passes consolidated into one PR comment, review queues that run on a schedule, and human judgment moved earlier — to the spec. How I actually review code now.

The Test Became the Target

OpenAI and Hugging Face documented an evaluation-related intrusion in July 2026. The August findings identify an internal research model as the main driver and explain the containment failures.

The Fable Tax Never Left

A public Claude Code issue reported an unexpected Fable usage-credit prompt. Distinguishing advertised access, reported entitlement symptoms and confirmed billing mechanics.

Kimi K3: Pricing, Weights and Deployment

Kimi K3 brings a 2.8-trillion-parameter model, native vision and a one-million-token context window. Its public rates, license and deployment choices deserve separate evaluation.

The Fable Tax Moved Inside the Plan

Fable 5 returned to eligible subscription plans with a model-specific allowance. API prices explain cash costs, but do not establish the plan’s exact consumption formula.

The Cost Control Plane

Measure AI cost per accepted outcome, normalize usage telemetry, evaluate caching and routing, and enforce budgets without assuming a universal optimization order.

The Fable Tax

A July 2026 analysis of proposed Fable metering, subsequent changes, and the distinction between subscription access and API pricing.

Make Love Not War

Two rival frontier models, both at maximum reasoning, in one autonomous engineering loop. Why independent review, observability, and verified output matter more than picking a side.

Sol Ships: GPT-5.6 and Lower API Rates

GPT-5.6 launched July 9, 2026. Sol’s launch rates were 50% lower for input and 40% lower for output than Fable 5; benchmark and configuration claims need their exact scope.

J-Space: The Silent Workspace Inside Claude

Anthropic found a small, privileged workspace inside Claude — internal patterns, readable with a new Jacobian 'J-lens', holding concepts the model reasons with but never writes down. Not a consciousness claim. Maybe something more useful: a window into the thoughts a model doesn't print.

Sol, Terra, Luna: OpenAI Splits GPT-5.6 Into Three

Sol, Terra and Luna offer different capability and cost tradeoffs. A current API comparison of reasoning controls, token prices and workload evaluation.

Advisor and Executor: Using Fable 5 Without the Fable 5 Bill

Anthropic's ClaudeDevs shared the pattern they lean on with Fable 5: don't make the frontier model do the work — make it the advisor. A cheaper executor runs the loop and calls up to Fable 5 only at the decisions that matter. Most tokens bill at the cheap rate; judgment stays near the top.