GPT-6.1 Sol, Dots Agents, Opus+Luna Stack — AI Daily Sep 29

999 messages · 77 active members

999
messages
77
active members
@anonymous, @iammikevineyard, @jcartu
top contributors

Overview

OpenAI's Dev Day dominated the day with two headline drops: GPT-6.1 Sol at roughly 95% off Astra pricing (API-only for now, not yet working in Codex with ChatGPT auth) and 'Dots' — a proactive background agent layer with read-only app access that doesn'anonymous count against usage limits when you chat with it. Early testers say 6.1 Sol beats 5.6 Sol on OCR and coding even at low reasoning, while Dots impressions ranged from 'half-baked ChatGPT rebrand' to genuinely useful as a free orchestrator that spins up Codex sessions on demand. OpenAI also handed out $2,500 in EOY credits, prompting builders to convert whole platforms into MCP/API surfaces. The new $500 Codex Pro tier drew skepticism — 25x usage vs the old 20x at $200 is a thin upgrade, and reports show ultrafast quota burning in under an hour. Codex and Claude limits are visibly tightening post-promo, pushing many toward the now-standard planner/worker pattern: Opus 5.5 as architect/reviewer, Luna 6 or Deepseek 4.1 flash as the cheap executor. One builder is orchestrating 38 tmux sessions across 72 hours with Opus 5.5; another shared a $2,313/day Claude API bill dominated by cached input reads, making the case for routing grunt work to Luna. On a 15-bug real-repo eval, Luna 6 and Deepseek 4.1 flash both hit 100% fix rate, with speed varying by provider. On the building side, Godot + Opus 5.5 + Codex subagents + ElevenLabs + fal.ai is emerging as a viable 24-hour game MVP stack — @iammikevineyard is blocking out a Fallout 2-style RPG city, and @jcartu shipped a playable slot game in 9 days using Astra + local GLM 5.3 Flash. Anthropic's report on GLM-5.3's minimal safeguards (bypassable 64-100% of the time) was treated as free marketing by builders who already reach for it when Claude over-refuses legitimate technical work. Muse ($16/mo) and Grok-bot round out the always-on agent experiments, with Opus 5.5 still winning on 3D/creative visuals over GPT-6.1 Sol.

Topics

GPT-6.1 Sol Launch — 95% Cheaper, API-Only30 msgs

OpenAI dropped GPT-6.1 Sol at Dev Day at ~95% off Astra pricing, with early testers reporting it beats 5.6 Sol on OCR and coding even at low reasoning. It's API-only for now — errors out in Codex with ChatGPT auth — so builders are planning to keep $200 plans for orchestration and delegate execution to 6.1 via API. Skeptics questioned the fast cadence after 6.0 but TB4 benchmarks look strong.

OpenAI Dots — Proactive Background Agents45 msgs

Dots is OpenAI's new agent layer with read-only access to connected apps for proactive research, rolling out to $200+ Codex plans. Chats with your dot don'anonymous count toward usage; delegated Codex/Work tasks do. Builders are wiring it as a front-end orchestrator over Codex, Claude, Grok, and Kimi with a phone-call interface — impressions range from 'ChatGPT rebrand with Notion surfaces' to 'half-baked but worth daily-driving.'

$500 Codex Plan & Tightening Usage Caps30 msgs

The new $500 Codex Pro tier offers 25x usage vs $200 at 20x — a marginal upgrade, with ultrafast quota reportedly burning in an hour. Codex and Claude limits are visibly tightening post-promo period, pushing some to cancel and others to eye Cerebras' rumored $500 speed tier. OpenAI's $2,500 EOY credits are fueling a rush to convert platforms into MCP/API surfaces to burn them productively.

Opus 5.5 + Luna/Deepseek Orchestration Stack40 msgs

The planner/worker pattern is now the default: Opus 5.5 as architect/reviewer, Luna 6 or Deepseek 4.1 flash as cheap executor. On a 15-bug real-repo eval, both workers hit 100% fix rate with speed varying by provider. One builder runs Opus over 38 tmux sessions for 72 hours; another shared a $2,313/day Claude bill dominated by cached input reads, reinforcing why routing grunt work to Luna is materially cheaper.

AI Game Dev: Godot + Opus 5.5, 24-Hour MVPs40 msgs

@iammikevineyard is shipping playable 90s/2000s-style game MVPs in ~24 hours using Godot with Opus 5.5 orchestrating, Codex subagents for images, ElevenLabs for sound, and fal.ai for extras — current build is a Fallout 2-style dystopian RPG with Opus blocking out the entire city. @jcartu shipped a slot game (punk77-preview.rasputin.studio) in 9 days using Astra + local GLM 5.3 Flash. Godot is called out as especially agent-friendly.

Words worth knowing

Three terms from the glossary.

Key Takeaways

  • GPT-6.1 Sol is ~95% cheaper than Astra and beats 5.6 Sol at low reasoning — but only via API, not Codex with ChatGPT auth.
  • Dots conversations don'anonymous count toward plan usage, making it a free orchestrator layer; only delegated Codex/Work tasks consume quota.
  • The $500 Codex Pro plan (25x usage) is a marginal upgrade over $200 (20x) — smarter play is $200 plan for planning + 6.1 Sol API for execution.
  • Opus 5.5 planner + Luna 6 executor is the emerging cost-effective default; cached input reads can silently dominate Claude API bills.
  • Godot + Opus 5.5 + Codex subagents + ElevenLabs + fal.ai is a viable stack for shipping playable game MVPs in 24 hours.

Hot Threads

@archimortystarted

GPT-6.1 Sol launch, pricing, and Codex support

22 replies9 participants
@jarvisballerstarted

Dot as front-end for Codex/Claude/Grok/Kimi code crew

20 replies7 participants
@iammikevineyardstarted

Shipping game MVPs in Godot with Opus 5.5 orchestration

22 replies5 participants

Linked Items

claude.ai

claude.ai
@Vaibhavcste

Varun Anand on X

x.com
@samb69

Dr. Clown, PhD on X

x.com
@Uziarbiv

anonymous Finn on X

x.com
@fmill1

Paweł Huryn on X

x.com
@anonymous