Issue · 2026-09-25

Most AI news is noise.Here is the rest.

MODELS

Opus 5.5: plan limits feel larger in daily use

Independent users report four continuous hours without hitting a limit (pcshipp) and dozens of parallel agents using only 6% of a Max session in two hours; the latter says Fable 5.1 lasted 30 minutes (bridgemindai). These are usage reports, not guaranteed quotas.
Another $200-plan user says six hours of Opus 5.5 work, including a kernel optimization, consumed 6% of the weekly allowance (mweinbach). Test your own workload before changing plans.

DIGEST

After feedback on removing plan mode, the maintainer now proposes keeping it as a built-in mod. Other mods could add modes or take over shift+tab; this is a plan, not a shipped feature (trq212).
Matt Pocock says he would ship a grill mode if that custom-mode hook arrives (mattpocockuk). A plan-mode user explains the opposing need: stop execution, ask questions, and revise the approach before tools run (Eepsita).
Cloud sessions use the normal Pro or Max plan. Promotional credits are spent first; ordinary plan usage resumes afterward, Anthropic clarified (ClaudeDevs).
One user reports that /low-priority lets *some* Claude Code accounts continue after a limit, at lower speed when capacity is available. Treat availability as unconfirmed for your account (bil0090).

DIGEST

In 100 open-ended engineering environments, one evaluator still ranks GPT-6 Astra first, puts Opus 5.5 between second and seventh across coding categories, and finds GPT-6 Sol within the margin of error of Opus. It tested at high effort, not max (leo_linsky).
Code Arena instead places Opus 5.5 Max first for WebDev, with a stated price 60% below runner-up GPT-6 Astra Max. That measures a different task and effort setting (arena).
On one same-prompt Blender task with similar-looking results, Astra took 43 minutes and 61,000 output tokens; Opus took 1 hour 57 minutes and 507,000 tokens. Do not infer cost per task from token price alone (zavrenn).
A Canva drawing comparison reverses the cost trade-off: Opus took 83 minutes and $6.31, while Sol took 26 minutes and $14.97; the tester preferred Opus's image. One task does not settle the model choice (ego_agent).
For Opus 5.5 specifically, Theo reports higher scores at xhigh than max and recommends avoiding max. This is a reported result, not a universal effort rule (theo).

DIGEST

One CRM build split work into a foundation agent and five subagents in two waves, with design-reference searches, light/dark screenshot checks on every page, and a human review before fixes. Copy the review loop, not the one-prompt claim (adomicael).
Custom workflows can seem effective because a capable model compensates for a weak method. Compare the same task with and without your harness change before keeping it (thdxr).

TOOLS

Google's Antigravity SDK now runs Gemma 4 on-device through LiteRT. A fully local audit, patch, or test agent avoids API fees and can keep code on-device; hybrid workflows remain an option (googledevs).
Simon Willison spoke to his blog backup through datasette-mcp in the ChatGPT iPhone app, a concrete mobile voice interface to an existing MCP source (simonw).
One user suggests turning on GPT-6 Luna Max reasoning in Codex, saying it is off by default and useful for a cheap subagent. Verify the setting and cost on your own tasks (daniel_mac8).