Issue · 2026-09-22

Most AI news is noise.Here is the rest.

ORCHESTRATION

Jev is becoming the cheap decision layer under other people's harnesses

Measured: a Jev + Astra harness beat Minecraft's Ender Dragon in 8m43s for $0.97 in total API spend, $0.01 of it Jev, with the harness open-sourced (rronak_).
90% off repetitive work by moving steps out of LLM calls and into code: 100,000 compliance alerts cost over $290K on Opus 5 alone (_aj).
LangChain benchmarked Jev against LLM judges on accuracy, repeatability, latency and cost, as a cheaper way to grade agents (LangChain).
What people are building on the API: a browser agent (jev-ultrafast) and context compression (fast-jev-compaction) (charliejhills).
Read before adopting: a reverse-engineering of Jev's architecture and where it falls down (_raghavdixit_), and simonw's notes on decision models as a new model category (simonw).
It is API-only and not open source; Nimble is the open version claimed to perform as well (aisearchio).
Skills stay out of the context window: mark them user-invocable only and let Jev pick from the list per request (dani_avila7).

MODELS

Leak: claude-opus-5-5 in stealth testing under the codename claude-wafer-eap, was planned as Opus 5.2, ~20% cheaper than Opus 5 and reportedly above Astra and Fable 5.1 (Mr_Salio).
The table being circulated: $5/$25 per Mtok down to $4/$20, cache read $0.50 down to $0.20, cache write $5 (TokenGremlin).
The line that actually decides the bill is the $0.20 cache read next to the $5 cache write — keep working in one context instead of rebuilding it every turn (itsvlady).
Counterpoint: a cheaper sticker is an illusion if the model burns more tokens to reach the same result (TokenGremlin).

MODELS

Reported in final launch preparation, with the Terra tier discontinued and Anthropic answering with Opus, Sonnet and Fable 5.5 (synthwavedd).
Do not plan around Tuesday: no confirmed launch date exists, including among well-informed leakers (TokenGremlin).
Evidence over rumour: some gpt-5.6-luna requests already return as gpt-6-luna in Codex backend responses, which broke the cost tracking that noticed it (LuminaBench).

DIGEST

The launch claim: a notable improvement over Grok 4.6 at the same price and speed (SpaceXAI).
On Artificial Analysis it comes out more expensive than Astra (theo).
First numbers: 46.3% on CursorBench 4.0 against Fable 5.x (Ryanhlx).
How to read day one: ignore takes that lean on a single public benchmark in either direction (kunchenguid).

ORCHESTRATION

claude plugin eval init writes the test cases from your description of a good result; every case also runs without the plugin, so the delta between the two scores is the verdict (addyosmani, doublenickk).
Limitation before you trust the gate: it cannot test a CLAUDE.md rule, because its baseline arm unloads a plugin (dylan_merigaud).
The front-matter switch is spelled disable-model-invocation: true, not the symmetric model-invocable: false (ariefr23).
OpenCode 2.0 can modify everything about itself in real time, including during running sessions (thdxr).