Codex Web GPT v6.1.0

V6 brings a new model picker, optional Pro usage tracking, more control over browser chats, and the fixes developed since v5.0.8.

V6 highlights

• Clear model names and native Effort controls. Automatic models now end in (Web): GPT-5.6 Luna, GPT-5.6 Sol, GPT-5.6 Pro, and GPT-6 Pro. Sol exposes the reasoning levels your account supports through Codex's Effort selector. Instant remains a separate entry so its context budget is preserved. Existing tasks using the old model names continue to resolve to their original modes.
• A Limits tab for Pro. Opt in by checking your signed-in plan. The launcher tracks its own submitted messages for the supported Pro $100 and $200 plans, separates the two Pro models and shared allowances, and shows a yellow indicator from 75% usage. These are local estimates over rolling windows, not OpenAI's remaining quota or reset schedule; messages sent outside the launcher are not counted. Tracking is unavailable in Zero Risk.
• Smaller parts for Bigger Context. Large requests now use six ordered parts instead of three; smaller requests can still use one or two. The overall context and compaction budgets are preserved. The final part starts the task, and tools and attachments stay with that final part.
• New browser chat for each turn. An optional setting rebuilds the context of the same Codex task in a fresh ChatGPT conversation and attaches its connector again. This can help when ChatGPT stops retaining tools across follow-ups, at the cost of sending more context. Off by default; automatic mode only.
• Save chats in ChatGPT. An optional setting uses regular conversations instead of Temporary Chat. Off by default. ChatGPT memory and custom instructions may apply when enabled; it is independent of the fresh-chat setting.
• Native Linux ARM64 packages. Linux now has x64 and ARM64 AppImages and runtime bundles, with architecture-aware installation and updating.
• A new application icon. The launcher keeps its familiar ChatGPT mark inside the interface.

Earlier V6 fixes and reliability improvements

• Account-aware effort selection. Locked Pro controls are distinguished from usable reasoning levels, and the selected model family and effort are checked before sending. This addresses the Plus upgrade-modal path in #610.
• Trusted workspace recovery. Native date/time environment updates, supported steering messages, and resumed task context resolve against the current native rollout. Missing or contradictory authority still fails explicitly instead of borrowing a cached workspace.
• Compaction and continuation. Support for native responses/memento text compaction joins the existing structured handoff path. Retained-turn ownership, tool-result settlement, cancellation, and compaction errors are handled more precisely. Please retest the long-session cases in #619; this is not a claim that every continuation failure has the same cause.
• Browser submission and recovery. Composer insertion, current effort confirmation, and oversized-message diagnostics are more precise. Account cooldowns are not retried as generic browser failures, and an accepted generation is not blindly submitted again.
• Sign-in navigation and tab closure. Authentication probes share one operation instead of racing navigation on the same page. Closing a cancelled browser tab no longer leaves it stranded when the control request fails. These changes address #628 and #624; reporter testing is still needed.
• Markdown fidelity. Link destinations participate in stream consistency checks, incomplete link hydration is not committed as finished text, Codex plan markers survive conversion, and literal Windows file labels retain their backslashes. Please retest #635.
• Setup, hooks, and route ownership. Equivalent Codex TOML rewrites no longer invalidate an unchanged managed interrupt hook. Custom model-provider conflicts receive explicit guidance. Windows tunnel installation retries only its own locked files and preserves rollback errors. Troubleshooting also explains the max_threads alias, website language, and launcher connection lifecycle. Addresses #612 and #614.
• More useful diagnostics. Tool delivery, browser completion, and broker retirement have clearer receipts. These help distinguish a missing consumer, a stalled native tool, and a completed browser turn without treating missing log lines as proof of failure.

The fresh-chat setting is available for the connector-retention case in #428. ChatGPT-side safety refusals, account limits, and model claims that tools are unavailable are not presented as fixed by this release.

Thanks to everyone who supplied logs, reproductions, and test results.

Changes in v6.1.0

• Support for the updated ChatGPT interface. Sign-in checks, the message field, sending, model and effort selection, connector selection, and reading completed replies now support the new layout alongside the previous one. This addresses the recent reports where manual chat worked but setup or the browser test failed.
• Checks work with the browser hidden. Hiding ChatGPT no longer leaves the main browser page without a usable viewport. Sign-in and Limits checks can run in the background without making you open the browser first.
• Limits setup is more reliable. The launcher opens the current Billing page when needed and selects the usable profile control when ChatGPT keeps two sidebar layouts on the page. It still checks the account and subscription before enabling tracking.
• Updated MCP setup guide. The walkthrough shows the current interface. Click a video to pause or resume; a pause indicator appears in both inline and expanded views.
• Windows installation and Codex integration. The installer finds the latest version through the public release page, avoiding the GitHub API rate limit. Setup and removal also recognize equivalent rewrites of the launcher’s Codex hook without deleting unrelated settings.
• Window and Activity recovery. The launcher returns to a visible screen after a monitor is disconnected. Activity restores the last 300 valid events across rotated logs, including when a trailing log line is incomplete.
• Safer diagnostic export. Export refuses links to the original logs and writes a separate file before replacing the destination.
• Clearer Linux update guidance. An unsupported launch path is detected before downloading an update, with instructions for restoring the normal installation.
• Browser notices and instructions. The launcher dismisses the observed ChatGPT announcement banner. The model instructions also clarify that local tools run through Codex with its configured sandbox and approvals. Please retry reported tool refusals; this does not guarantee that ChatGPT-side refusals are resolved.

Thanks to @renanassisalves for PR #664, the page details, and the testing that helped reveal the extent of the interface change while I was waiting for it to reach my account. This release uses my own implementation, tested after the new interface became available to me.

Install

Quit the launcher, download the package for your system, and replace the existing application. Reopen it and restart Codex after the launcher refreshes its integration. Do not delete your launcher data or ChatGPT profile.

• Windows x64
• macOS Apple silicon · macOS Intel
• Linux x64 · Linux ARM64