saved brief history• source-linked snapshots

$ ls ./history

2 saved briefs

01ProductGPT-5.6 Sol is the documented Codex default family

August 17, 2026 · 10:00 PM ET · Evening brief

Current model docs recommend 5.6 Sol, Terra, and Luna; GPT-5.3-Codex is already deprecated for ChatGPT-signed Codex, and GPT-5.4 retires from that surface on August 31, 2026.

deeper context / thread read

[Codex models]Recommended set is 5.6 Sol (flagship coding/computer-use), Terra (everyday, lower cost vs 5.5), and Luna (fast/cheap repeatable work); default Power uses gpt-5.6-sol at medium reasoning.
[Codex models]Ultra mode is documented as subagent-backed parallel work, not just a higher single-agent reasoning slider; Max is extra time on one task.
[Codex models]gpt-5.2 and gpt-5.3-codex are already deprecated in Codex with ChatGPT sign-in; gpt-5.4 and gpt-5.4-mini retire from that path on August 31, 2026. API-key Codex is called out as unaffected by the 5.4 retirement.

[Codex models] original source ↗

02ProductGPT-5.3-Codex-Spark stays a Pro-only research preview

August 17, 2026 · 10:00 PM ET · Evening brief

Spark remains a separate near-instant text-only Codex model with its own limits, not the same thing as fast mode on Sol.

deeper context / thread read

[Codex models]Model card: text-only research preview for near-instant coding iteration, available to ChatGPT Pro users via `codex -m gpt-5.3-codex-spark`.
[Codex speed]Speed docs distinguish Spark from fast mode: Spark is its own model choice and usage pool; preview is Pro-only.
[Codex pricing]Pricing docs still list Spark as a Pro research preview and say it is not in the API at launch.

[Codex models] original source ↗

03WorkflowCodex CLI 0.147.0 is the latest numbered changelog cut

August 17, 2026 · 10:00 PM ET · Evening brief

The official Aug 7 CLI note is 0.147.0; GitHub is also publishing 0.148.0-alpha builds, including alpha.21 on Aug 17.

deeper context / thread read

[Codex changelog]Install line in the changelog is `npm install -g @openai/codex@0.147.0`.
[Codex changelog]0.147.0 adds portable Agent Plugins, conversation sections, `--approve-for-me`, Cursor/Claude import sync, and opt-in MCP 2026-07-28.
[Codex changelog]Same note removes deprecated `codex exec --full-auto` in favor of `--sandbox workspace-write`.
[GitHub Releases]GitHub pre-release rust-v0.148.0-alpha.21 was published Aug 17; alphas are not the same as the 0.147.0 changelog entry.

[Codex changelog] original source ↗

04RemoteComputer History is opt-in on the macOS desktop app

August 17, 2026 · 10:00 PM ET · Evening brief

Aug 13 changelog: an opt-in timeline of app/site activity for ChatGPT and Codex, with explicit geo and plan limits.

deeper context / thread read

[Computer History]Feature turns activity across apps and websites into memories and a timeline; users choose sources, can pause, and can review or delete history.
[Codex changelog]Availability: ChatGPT Pro, Business, and Enterprise; Business/Enterprise admins must enable it first.
[Codex changelog]Initial availability excludes the EEA, Switzerland, and the United Kingdom.
[OpenAI Developers]OpenAI Developers framed the same launch as opt-in context so Codex can pick up recent work and suggest skills or scheduled tasks.

[Codex changelog] original source ↗

05RemoteLinux desktop preview plus Claude/Cursor import

August 17, 2026 · 10:00 PM ET · Evening brief

Aug 11: ChatGPT desktop (with Codex) is in preview on supported Linux distros, and the app/CLI can import other agents’ setup and recent work.

deeper context / thread read

[Linux desktop app]Linux preview targets Ubuntu, Debian, and Fedora on x64 and ARM64 via .deb/.rpm; sign-in covers projects, local files, and Codex.
[Import from another agent]Desktop import supports Claude Code, Claude Cowork, and Cursor (instructions, settings, skills, plugins, projects, recent work), with optional auto-updates in Settings > Import.
[Codex changelog]CLI can import supported Claude Code and Cursor setup/chats with `/import`.
[OpenAI Developers]OpenAI Developers: import is available in the ChatGPT desktop app and can keep imported work in sync.

[Codex changelog] original source ↗

06RemoteStaff: Unified Exec is now default on Windows

August 17, 2026 · 10:00 PM ET · Evening brief

Derrick Choi said Windows Codex now uses the same long-lived command model as macOS/Linux. This is a staff post, not the Aug changelog row.

deeper context / thread read

[Derrick Choi]Choi: Unified Exec is now the default for Codex on Windows.
[Derrick Choi]Stated capability: start a dev server, keep it running, watch new output, send later input, and run test watchers in the background.

[Derrick Choi] original source ↗

07RemoteCodex Remote on iOS is the week’s staff emphasis

August 17, 2026 · 10:00 PM ET · Evening brief

Dimillian and Atty Eleti amplified a start-from-iOS Remote write-up. Widgets, system integrations, and Files mounts in that thread are user wishes, not shipped claims.

deeper context / thread read

[Thomas Ricouard]Dimillian said Codex Remote drew Viticci back to the platform and that the team will take Remote further.
[Atty Eleti]Eleti called Codex Remote a glimpse of the future while quoting the same iOS start-from-phone / voice / thread-dispatch description.
[Federico Viticci]The quoted write-up describes full start-from-iOS, near desktop parity, and voice dispatch; it also lists widgets, Reminders/Calendar, and Files-app mounts as still-open wishes.

[Thomas Ricouard] original source ↗

08CommunitySottiaux asked for obvious Codex/API gaps still in reach

August 17, 2026 · 10:00 PM ET · Evening brief

A high-volume Aug 17 staff prompt is intake, not a roadmap. Recurring concrete asks in the thread include usage, subagent status accuracy, ChatGPT-app performance, and more visible profile controls.

deeper context / thread read

[Thibault Sottiaux]Sottiaux asked what Codex, the API, or models should already do that is 100% within reach but still missing.
[X reply]One concrete request: the subagents bar can show agents as working when they are not until the user opens the subagent.
[X reply]Another concrete request: unexpected usage burn despite a very high reported cache rate.
[X reply]Other concrete requests in-thread: ChatGPT-app stutter/perf on high-end PCs, and making custom-profile permissions more visible in-app instead of mostly config files.

[Thibault Sottiaux] original source ↗

09WorkflowiOS 1.2026.209 tightens Remote task reconnect

August 17, 2026 · 10:00 PM ET · Evening brief

Aug 7 mobile changelog is bug-fix heavy: reconnection, first-prompt preservation, voice vs background-conversations, and approval recovery.

deeper context / thread read

[Codex changelog]Task reconnection now shows progress while keeping existing tasks usable during connection checks.
[Codex changelog]New tasks keep their first prompt and attachments while starting, avoiding empty tasks that could not be reopened.
[Codex changelog]Voice conversations follow the Background Conversations setting; approvals and user-input requests are documented as recovering after reconnect.

[Codex changelog] original source ↗

10BuildersDaybreak splits into Blue and Red cyber access

August 17, 2026 · 10:00 PM ET · Evening brief

Aug 10: gated Trusted Access tiers, not default coding. Blue uses general models such as 5.6 Sol; Red is separately approved 5.6 Cyber-class work.

deeper context / thread read

[Codex changelog]Daybreak Blue is for most defensive work: vuln discovery, secure review, detection, IR, malware analysis, patch validation, via general-purpose models such as GPT-5.6 Sol.
[Codex changelog]Daybreak Red is separately approved for purpose-trained models such as GPT-5.6 Cyber (reproduction, exploit validation, pentest, red team). Blue does not grant Red.
[Cyber safety]Access is scoped to the approved identity, workspace or API org/project, model, and product surface via Trusted Access for Cyber.

[Codex changelog] original source ↗

11Signal, not releaseStaff keep pointing at talk-to-Codex and auto-compaction

August 17, 2026 · 10:00 PM ET · Evening brief

These are staff practice notes and UX tips, not versioned launches.

deeper context / thread read

[Gabriel Chua]Chua published a note arguing that talking through a rough idea with Codex is a useful way to find the actual task; he said the article started as a road-trip voice session.
[Gabriel Chua]Chua also documented Shift+Esc to mark all Codex chats as read.
[Romain Huet]Huet, in a reply, called Codex auto-compaction pretty magical—endorsement of existing behavior, not a dated ship announcement.

[Gabriel Chua] original source ↗

12Product1M-token context is available in Codex for ChatGPT accounts

August 17, 2026 · 9:41 PM ET · Evening brief

Tibo Sottiaux says the GPT-5.6 Sol long-context setting no longer requires an API key for Codex users signed in with ChatGPT. The normal, smaller default remains the tuned choice for most work.

[Tibo Sottiaux on X] original source ↗

13WorkflowLuna is the practical bounded subagent lane

August 17, 2026 · 9:41 PM ET · Evening brief

Pierre Vannier framed Luna as the fast model to delegate scoped tasks to while Sol and Terra handle the main thread. It is a useful signal for how Codex multi-agent work is being positioned.

[Pierre Vannier on X] original source ↗

14RemoteCodex Remote and iOS continue to expand

August 17, 2026 · 9:41 PM ET · Evening brief

The Codex team is still iterating on remote workflows and mobile access, making quick follow-ups and long-running work less tied to one machine.

[Damien on X] original source ↗

15BuildersOpenAI published the GPT-5.6 builders guide

August 17, 2026 · 9:41 PM ET · Evening brief

The official developer update focuses on building effectively with GPT-5.6 and is the canonical reference for API-facing changes.

[OpenAI Developers] original source ↗

16Signal, not releaseAstra is acknowledged, but no release date is public

August 17, 2026 · 9:41 PM ET · Evening brief

A popular Codex checklist mentioned that Codex will have Astra. Treat timing claims elsewhere as speculation until OpenAI publishes a concrete announcement.

[Tibo Sottiaux on X] original source ↗