I may have potentially identified one of the reasons some of us have increased usage burn (OpenCodex users & similar specifically)

Fyi I make no claim that this is the main issue/reason - just a little something that could be contributing. TL;DR: On the day I installed OpenCodex, my Guardian approval reviewers stopped showing codex-auto-review and started using ordinary models, instead matching the main task—including Astra. My recorded reviewer usage since then totals $1,024.85 at current API rates , versus $702.26 at GPT-5.4 rates or $56.18 at Luna rates , assuming identical tokens and caching. These are API cost equivalents. Setting auto_review_model = "gpt-5.6-luna" successfully switched new reviews to Luna / low. The timing suggests an OpenCodex routing change, but I haven’t proved the cause or equivalent review quality. I audited my Codex rollout history and found a clear change on 26 July 2026, the day I installed OpenCodex . This concerns Guardian’s automatic approval reviews for tool actions , not /review code reviews. Every identifiable Guardian session created between 9 (day Sol/Terra/Luna released) and 25 July used codex-auto-review . On 26 July , new reviewers started using Sol, Terra, Luna and other provider models. The first ordinary-model reviewer appeared at 07:16 BST , after my installation that morning. I haven’t found any newer sessions using codex-auto-review . Later examples included Astra reviewing actions requested by Astra and DeepSeek reviewing actions requested by DeepSeek . What codex-auto-review actually is OpenAI explicitly documents Auto-review as GPT-5.4 Thinking with low reasoning : OpenAI’s Auto-review explanation OpenAI’s rate card PR #17505 also describes replacing hardcoded gpt-5.4 with the codex-auto-review slug. That supports the connection, although an alias alone doesn’t prove the backend model for every historical request. Current upstream Codex source prefers codex-auto-review for ChatGPT authentication and gpt-5.6-luna for API-key authentication. OpenCodex has its own auto_review_model override. I hadn’t set it. The usage comparison I use Pro Lite and Plus accounts through OpenCodex. These are USD API cost equivalents. The table below reprices the same recorded tokens at current model rates: Period Guardian sessions Recorded models’ API equivalent Same tokens at GPT-5.4 rates Same tokens at Luna rates 4–10 September 124 $248.65 $88.16 $7.05 26 July–10 September 1,049 $1,024.85 $702.26 $56.18 For the wider period, that’s approximately 31.5% lower at GPT-5.4 rates or 94.5% lower at Luna rates . Here’s which reviewers contributed: Reviewer Sessions, 26 Jul–10 Sep API equivalent Sol 582 $801.53 Astra 80 $203.11 Luna 268 $10.90 Terra 118 $9.25 GPT-5.5 1 $0.06 Total 1,049 $1,024.85 All 80 Astra sessions were in the September window. Their tokens alone work out at $203.11 using Astra rates, $51.00 using GPT-5.4 rates, or $4.08 using Luna rates . For comparison, before the switch I found 614 codex-auto-review sessions created during 9–26 July . The 601 with recorded usage contained 42.49M uncached input, 394.93M cached input and 740K output tokens : approximately $216.05 at GPT-5.4 rates versus $17.28 at Luna rates . How I counted Only sessions explicitly marked as Guardian reviewers. A session can contain multiple approval decisions. Each session’s final cumulative usage counted once; reasoning tokens weren’t added again on top of output. Prices are published standard API rates as of 10 September 2026 , including applicable long-context adjustments—not reconstructed historical prices. Alternative-model columns hold token counts, caching and request sizes constant. They don’t predict actual token usage or equivalent decision quality. Dates use UTC. September 10 is partial; the wider audit ends just before my Luna config edit at 05:53 UTC . The periods overlap. The wider priced subset excludes 65 other-provider/internal/unsupported-model sessions. Provider-prefixed entries naming the same model are grouped together. Totals are calculated before rounding. Pricing sources: Astra , Sol , Terra , Luna , GPT-5.5 , GPT-5.4 . What I changed I added this at the top level of the Codex config.toml used by OpenCodex : toml auto_review_model = "gpt-5.6-luna" Afterward, I verified four new Guardian sessions using Luna / low reasoning , including one requested by a DeepSeek thread. They produced 16 completed approval assessments . That confirms the routing change, not equivalent review quality. Most reviewers were already running at low reasoning, so the cost comparison is primarily about model choice. My working hypothesis: switching to OpenCodex changed which dedicated reviewer Codex could select, causing reviews to fall back to the requesting thread’s model. The timing and sampled rollouts fit, but I haven’t isolated the cause from Codex updates or other configuration changes. I’m leaving Luna pinned and keeping an eye on this. If you use OpenCodex, check what model your Guardian sessions are actually using . If anyone has rollout history spanning their OpenCodex installation, I’d be interested to see whether the same switch appears.