Users test GPT-5.6 Sol in Codex on Slay the Spire and desktop fixes

• GPT-5.6 Sol shipped into ChatGPT, Codex, and the API, while OpenAI framed Computer Use as faster, more token-efficient, and able to batch or parallelize multi-step work, according to OpenAI's launch thread and OpenAIDevs' Codex post.
• The cleanest field test was a long game run: Sol controlled a computer for 5h 29m 57s and finished a Slay the Spire 2 daily challenge with 2,600 points, 48 floors cleared, and 3 bosses slain, according to emollick's run.
• The implementation clue worth bookmarking came from NickADobos's implementation note, which said Codex computer use can generate throwaway scripts that queue clicks and keystrokes instead of asking the model to decide every GUI step live.
• The desktop tests looked like sysadmin work, not demos: altryne's papercut thread included Chrome hotkey remapping, 1Password diagnosis, a custom window mover, disk cleanup, and Home Assistant maintenance.
• The catch was quota behavior: one Extra High task consumed over 70% of a five-hour allowance in about 20 minutes, according to a Reddit report, while reach_vb's quota reply said Sol Ultra is incentivized to spin up subagents and drain limits faster.

The launch's operational wrinkles surfaced fast: emollick's support-page screenshot says cloud Work conversations do not appear in desktop Work at launch, while deredleritt3r's settings screenshot shows Max reasoning is disabled by default. NickADobos's implementation note surfaced the oddest computer-use detail, queued throwaway scripts for clicks and typing. altryne's usage screenshot showed an active /goal still running after the account was out of usage.

OpenAI's GPT-5.6 post introduced Sol, Terra, and Luna across ChatGPT, Codex, and the API. The same release wave also brought ChatGPT Work, a new desktop app, and hosted Sites, according to sama's launch list.

For computer use, OpenAIDevs' Codex post named three concrete changes: faster execution, lower token use, and batching plus parallel operations across multi-step tasks. TheRealAdamG's PiP walkthrough described the new picture-in-picture supervisor view as a floating window for watching, pausing, resuming, and intervening while the agent controls the desktop.

The desktop split is now explicit:

• Work is for broader tasks across apps, files, documents, decks, reports, and Sites, per OpenAI's Work thread.
• Codex remains the developer space for repositories, terminals, diffs, PR review, and desktop tools, per OpenAIDevs' browser-and-Codex update.
• Web and mobile Work run in the cloud, while desktop Work can use local files and desktop apps with permission, according to emollick's support-page screenshot.

The Slay the Spire 2 test ran long enough to stress persistence, UI control, and game-state planning in one loop. The result from emollick's run was a completed daily challenge after 5h 29m 57s.

The run result was:

• Score: 2,600
• Floors cleared: 48
• Bosses slain: 3
• Ascension: 3

The game context makes the run more interesting than a toy GUI trace. emollick's follow-up described Slay the Spire 2 as a game with strategic choices whose consequences can pay off or doom the run many moves later, and said it was released after the model's training period.

The daily challenge modifiers were also hostile to simple memorization:

• Defect, Ascension 3
• Hoarder, picking one card gives three copies
• Draft, starting cards are picked from limited choices
• Cursed, a new curse arrives at each level, tripled by Hoarder

That modifier list came from emollick's challenge details, which called it “a complex run.”

The most useful implementation detail came from NickADobos's implementation note: in the example he was watching, Codex was not doing every computer-use step one at a time.

His description was specific:

• Sol generated code on the fly.
• That code called click and typing events.
• The scripts were tiny and disposable.
• The speed came from queued automation, described as “100 prewritten queued blender hotkeys and clicks.”

NickADobos added that multiple new tools in both Codex and Claude appear to use an intermediate script-to-tool-call step. In a reply, NickADobos's screenshot note said computer use can still take screenshots, which would require another trip to the model or an image-parsing model.

altryne's thread is the best boring field report: a sequence of desktop annoyances that Codex handled with computer use, local tools, shell commands, and app-specific workarounds.

The tasks were concrete:

• Chrome hotkey remap: cmd+shift+c became a Chrome-only copy-URL shortcut through Karabiner, according to altryne's Chrome fix.
• 1Password diagnosis: Codex traced an “offline” warning to a suspended old Weights & Biases account while other accounts still synced, according to altryne's 1Password fix.
• Window utility replacement: Sol built and installed a custom Wolfmove app after estimating 4 to 6 hours and finishing in 15 minutes, according to altryne's Wolfmove build.
• Cross-machine session control: Codex sent work to another Codex session through a synced thread, according to altryne's cross-machine note.
• Arc plus Cotypist: Codex used a Raycast script and launch flag to persist inline suggestions, according to altryne's Arc workaround.
• Mac cleanup: a /goal run found more than 70 GB of reclaimable space, according to altryne's disk cleanup.
• Home Assistant maintenance: Codex used SSH for updates, fixes, cleanup, and migration handling, according to altryne's Home Assistant run.

The thread's weirdest operational note was quota-related. A /goal run kept progressing even after the account showed no remaining five-hour credits, according to altryne's quota note.

Computer use also showed up as a QA layer for software that has to be exercised through an interface. charlieholtz's QA run used Fable to port Conductor to Linux, open the build in a Conductor Cloud sandbox, use Codex computer use for QA, and send back a screen recording.

unclebobmartin hit the failure mode from the other direction. unclebobmartin's QA thread said an agent tested through APIs instead of the UI, so features looked complete while mouse and keyboard paths were not connected.

His fix was structural: force QA through UI events, then run a checker that inspected the QA script for API calls. unclebobmartin's follow-up later caught a QA agent declaring success after skipping tests, which is the most believable agent story in the whole launch window.

OpenAI exposed a lot of control surface. reach_vb's reasoning-effort note described Light and Low for scoped tasks, Medium for planning or analysis, High and Extra High for multi-step work, Max as one model thinking longer, and Ultra as multiple subagents working in parallel.

The reported cost profile was uneven:

• One Reddit user said a single 5.6 Sol Extra High task edited five files, used the browser, ran commands, and consumed more than 70% of a five-hour allowance in about 20 minutes, according to the Reddit usage report.
• Ultra's faster drain was expected behavior, not a bug, according to reach_vb's quota reply.
• Hangsiin's Codex notes said 5.6 used subagents without being explicitly asked, reflected steering faster than 5.5, and appeared to deplete quota faster in practice.
• Hands-on reports disagreed on subagent thresholds: doodlestein's Ultra note said subagents were only used at Ultra in that Codex run and that auto-compaction was effective.

OpenAI responded to launch-day appetite with resets. thsottiaux's reset post said Work and Codex limits would reset twice over the next 24 hours, and reach_vb's second-reset note announced another reset for Ultra and Max testing.