On this page · 11 sections
Summary. OpenAI will remove GPT-5.4 and GPT-5.4 mini from Codex on 31 August 2026 for anyone signed in with ChatGPT, with gpt-5.4 replaced by gpt-5.6-terra and gpt-5.4-mini by gpt-5.6-luna. Read the rate card and the forced migration is a price cut: GPT-5.4 is $2.50 per 1M input tokens and $15.00 output, Terra is $2.00 and $12.00, a 20% reduction on both. GPT-5.4 mini is $0.75 and $4.50, Luna is $0.20 and $1.20, a 73% reduction. On the same night, GitHub's promotional Copilot credit allowance expires and Copilot Business drops from 3,000 to 1,900 AI credits per user per month. Two AI code reviewers, one date, opposite directions.
The comparison that matters is not which model is smarter. Both vendors sell access to the same frontier models at nearly identical list prices, and where the models overlap the numbers are the same to the cent: GPT-5.6 Terra is $2.00 and $12.00 per 1M tokens on OpenAI's own pricing page and $2.00 and $12.00 inside GitHub Copilot's AI credit rate card. GPT-5.3-Codex is $1.75 and $14.00 in both places. What differs is everything around the model: whether you can see which model ran, whether a batch tier exists, whether the review burns a second meter, and how much of the price you control. That is where a code-review workflow gets expensive or stays cheap.
This is a decision guide for engineering leads running both, or choosing one, before 31 August.
The 31 August cutoff, exactly as published
OpenAI's release note, dated 31 July 2026, is specific about scope: "On August 31, 2026, GPT-5.4 and GPT-5.4 mini will no longer be available in Codex for users signed in with ChatGPT. GPT-5.4 and GPT-5.4 mini will remain available on the OpenAI API and Codex sessions authenticated with an API key."
So it is not a model deprecation. It is a surface removal, and the boundary is your authentication method. A team running Codex through ChatGPT sign-in loses those models on 31 August. A team running Codex sessions with an API key keeps them. If half your engineers authenticate one way and half the other, you will get a support ticket that looks like a bug and is not.
OpenAI names the replacements directly: replace gpt-5.4 with gpt-5.6-terra and gpt-5.4-mini with gpt-5.6-luna. The migration checklist in the same note is the part worth acting on: "Before the cutoff, update workspace defaults, saved model settings, managed configurations, custom agents, and scheduled tasks that use either model." Scheduled tasks are the ones teams forget, because nobody is watching when they run.
| Retiring model | Standard rate, per 1M tokens | Replacement | Replacement rate | Change |
|---|---|---|---|---|
| gpt-5.4 | $2.50 in, $15.00 out | gpt-5.6-terra | $2.00 in, $12.00 out | -20% input, -20% output |
| gpt-5.4-mini | $0.75 in, $4.50 out | gpt-5.6-luna | $0.20 in, $1.20 out | -73% input, -73% output |
| gpt-5.4, cached input | $0.25 | gpt-5.6-terra, cached | $0.20 | -20% |
| gpt-5.4-mini, cached input | $0.075 | gpt-5.6-luna, cached | $0.02 | -73% |
| gpt-5.4, long context | $5.00 in, $22.50 out | gpt-5.6-terra, long context | $4.00 in, $18.00 out | -20% input, -20% output |
| gpt-5.4, cache writes | Not published | gpt-5.6-terra, cache writes | $2.50 | New cost line |
Two things in that table are easy to miss. The GPT-5.6 family introduces a cache-writes charge that the 5.4 family did not have: $0.25 for Luna, $2.50 for Terra, $6.25 for Sol per 1M tokens. A workflow that writes cache constantly and reads it rarely gets a new line item it did not have before, even while the headline rate falls. And OpenAI's pricing page shows short-context and long-context columns without publishing the token threshold that separates them, so you cannot compute the crossover point from the page alone.
Where the two products actually differ
Model access is close to a commodity. The workflow around it is not.
| Dimension | Codex (OpenAI) | Copilot code review (GitHub) |
|---|---|---|
| Model selection | You choose; model IDs published | "The model is selected automatically and is not disclosed" |
| Unit price visibility | Per-token rates published per model ID | Per-token rates published, but not which model ran a given review |
| Cheap tier | Batch tier at exactly half the standard rate on every populated cell | No batch tier on the Copilot rate card |
| Speed tier | Fast mode, published rates, 2x standard for most models | Not applicable |
| Depth control | Reasoning effort levels, model-advertised order | Lite and Balanced effort levels, org default inheritable |
| Depth cost | Priced through the model you picked | No published multiplier for effort level |
| Second meter | None | GitHub Actions minutes, billed at standard Actions rates |
| Repo-level config | Review rules live in AGENTS.md |
Organization and repository defaults in settings |
| Multi-repo review | Diffs across all repositories in a multi-folder project | Per pull request |
The single biggest practical difference is model disclosure. GitHub states it plainly in the billing reference: for Copilot code review "the model is selected automatically and is not disclosed, so per-token costs may vary between reviews." You can see what a review cost after the fact in AI credits. You cannot predict what the next one will cost, and you cannot decide that security-sensitive pull requests get an expensive model while dependency bumps get a cheap one, because you are not holding that dial. What you get instead, since 7 August 2026, is the effort dial: Lite and Balanced went generally available that day, replacing the preview names Low and Medium, with Balanced routing to "a higher-reasoning model" and organization admins able to set a default that repositories inherit. Each review is labelled with the level that ran.
Codex inverts this. You pick the model ID, so you know the rate, and repo review rules live in AGENTS.md alongside contributor conventions, which means the review policy is versioned with the code. The Codex CLI changelog for 4 June 2026 records exactly that move: "Moved repo review rules and contributor conventions into AGENTS.md". The cost of that control is that nobody is choosing sensibly on your behalf.
The batch tier nobody talks about
OpenAI's pricing page carries four tabs on the flagship table: Standard, Batch, Flex and Fast mode. The Batch and Flex tables are byte-identical to each other and sit at exactly 50% of Standard on every populated cell.
| Model | Standard, per 1M | Batch, per 1M | Fast mode, per 1M | Available in Copilot credits |
|---|---|---|---|---|
| gpt-5.6-luna | $0.20 in, $1.20 out | $0.10 in, $0.60 out | $0.40 in, $2.40 out | Standard only |
| gpt-5.6-terra | $2.00 in, $12.00 out | $1.00 in, $6.00 out | $4.00 in, $24.00 out | Standard only |
| gpt-5.6-sol | $5.00 in, $30.00 out | $2.50 in, $15.00 out | $10.00 in, $60.00 out | Standard only |
| gpt-5.5 | $5.00 in, $30.00 out | $2.50 in, $15.00 out | $12.50 in, $75.00 out | Standard only |
| gpt-5.3-codex | $1.75 in, $14.00 out | Not offered | $3.50 in, $28.00 out | Standard only |
| gpt-5.4 | $2.50 in, $15.00 out | $1.25 in, $7.50 out | $5.00 in, $30.00 out | Standard only |
Code review is close to an ideal batch workload. It is not interactive, a twenty-minute turnaround on a nightly sweep of open pull requests is fine, and the input is a diff that already exists. Routing that sweep through the batch tier halves the token bill. GitHub's rate card offers no equivalent, so the same work inside Copilot code review pays standard rates whether or not anyone is waiting for the answer.
Note the Fast mode inconsistency while you are reading that table. Fast mode is exactly 2x standard for the GPT-5.6 family, GPT-5.4, GPT-5.4 mini and GPT-5.3-Codex, but GPT-5.5 Fast mode is $12.50 input against $5.00 standard, which is 2.5x. OpenAI publishes no multiplier rule anywhere on the page, only the tables, so do not assume 2x and budget from it. Related: priority processing was renamed Fast mode on 30 July 2026, and both service_tier: "priority" and service_tier: "fast" are still accepted, so existing code did not break.
The second meter on the GitHub side
Copilot code review bills twice. GitHub's billing reference says token consumption is billed in AI credits and "the agentic infrastructure that powers the review consumes GitHub Actions minutes." Mario Rodriguez, Chief Product Officer at GitHub, said the same thing when usage-based billing was announced on 27 April 2026: "Copilot code review will also consume GitHub Actions minutes, in addition to GitHub AI Credits. These minutes are billed at the same per-minute rates as other GitHub Actions workflows."
That matters more than it sounds, because the two meters attribute differently and neither is visible in the other's report. To reconstruct the true cost of a review you need the AI credit figure from the AI usage page in billing settings and the Actions minutes from Actions metrics, filtered to the review workflow. Most teams look at one and conclude reviews are cheap.
Codex has no equivalent second meter in the published material. What it does have is a distinct set of things you should be watching. Since 7 August 2026 GitHub's Copilot usage metrics API reports third-party agent activity separately through a totals_by_3rd_party_agent array, one entry per recognised agent app, in the enterprise, organization, enterprise-user and organization-user 1-day and 28-day reports. If your teams run Codex agents inside GitHub workflows, that is where their activity shows up, and GitHub's own guidance is to group on agent_id rather than agent_name because display names change.
What each one is good at
Reviewing across repositories is Codex's clearest current advantage. The ChatGPT desktop app update of 30 July 2026 added the ability to "See all repositories in a multi-folder project and the lines changed in each one. Select Review to inspect diffs across those repositories without switching between separate review views." Cross-service changes are exactly the class of change where a per-pull-request reviewer is structurally blind, because the breaking half of the change is in the other repository.
Enforcing a standard across hundreds of repositories is Copilot's. The organization-level effort default is one setting that every repository inherits unless it opts out, and the review's effort level is labelled in the timeline and the pull request overview comment, so you can audit review depth across an estate without instrumenting anything. Codex's AGENTS.md model gives finer control per repository and correspondingly weaker central enforcement.
Predicting the bill is Codex's, decisively. You choose the model ID, the rate is published, and a batch tier exists. Copilot's undisclosed model selection makes per-review cost unpredictable by design.
Watching spend across a mixed estate is GitHub's. The Copilot impact dashboard added a return-on-investment section on 7 August 2026 showing cost per developer per month derived from actual credit consumption, that cost as a share of a chosen compensation band, and pull requests per developer per month, split between chat-and-completion users and agent-first developers. GitHub is careful to caveat it: "Cost figures are estimates based on AI credit consumption, and the salary selector is a modeling input rather than actual payroll data. Treat these metrics as directional."
A workable configuration
The honest answer for most teams running both is not to pick one. It is to give each the work it prices well.
Route the high-volume, low-stakes sweep to the cheapest published model on the cheapest tier. Dependency bumps, generated-code diffs, changelog and documentation pull requests, and lockfile churn do not need a frontier model. GPT-5.6 Luna at $0.20 and $1.20, halved again on the batch tier to $0.10 and $0.60, is a different order of magnitude from Sol at $5.00 and $30.00.
Set the Copilot organization default to Lite and reserve Balanced for the pull requests that earn it, meaning security-sensitive paths, cross-service changes, and anything touching authentication or payments. This is the cheapest structural saving on the GitHub side because it requires no change in how anyone works.
Keep review rules in AGENTS.md so they are reviewed like code. A review policy that lives in a settings screen drifts silently; one that lives in a versioned file gets a pull request when it changes.
Instrument both meters before you tune anything. AI credits from the billing AI usage export, Actions minutes filtered to the review workflow, and third-party agent activity from the usage metrics API. Tune one variable at a time afterwards.
Do the 31 August migration this week, not on 30 August. Workspace defaults, saved model settings, managed configurations, custom agents and scheduled tasks, in that order, because the first four fail loudly and the fifth fails at 3 a.m.
India-specific considerations
Three points land differently for Indian engineering organisations.
Data residency carries a published surcharge. OpenAI's pricing page states that "Regional processing (data residency) endpoints are charged a 10% uplift for models released on or after March 5, 2026, that are eligible for data residency." For teams whose source code or review context is subject to residency commitments under the Digital Personal Data Protection Act 2023 or a client contract, that 10% is a line item to model, not a footnote. It applies to the model rate, so it compounds with whatever tier you chose.
The batch tier is worth more here than the headline suggests. Offshore delivery teams often run review sweeps against a backlog rather than interactively per pull request, which is precisely the workload the batch tier exists for. A team already working asynchronously across time zones loses nothing by accepting batch latency and halves the token cost.
Currency exposure differs between the two products. GitHub Copilot's seat fee is fixed in dollars per user and easy to budget in rupees; its metered overage is not, and GitHub enables additional paid usage by default for organizations and enterprises. Codex, bought against an API key, is metered from the first token but every rate is published, so a spend model built from your own token counts is arithmetic rather than a guess. Neither is inherently cheaper. One is predictable by construction and the other is predictable only after you have measured it.
What we would not do
We would not standardise on a single AI reviewer because a comparison table said one won. The two price different things well, and the meaningful saving comes from routing, not from picking.
We would not treat effort or reasoning levels as free. GitHub says a higher reasoning level "consumes more tokens, and therefore more credits" and publishes no multiplier for it. OpenAI publishes Fast mode rates but no multiplier rule, and GPT-5.5 breaks the pattern the other models follow. In both cases the only reliable number is one you measured on your own diffs.
We would not defer the 31 August work on the grounds that a replacement was named. A named replacement handles the model ID. It does not handle the scheduled task that still references gpt-5.4-mini, and it does not handle the half of your team authenticating with ChatGPT rather than an API key.
The cost of AI code review is not the model rate. It is the reviews you run at frontier depth on diffs that never needed it.
FAQ
How eCorpIT can help
eCorpIT builds the routing layer that decides which diff gets which reviewer at which depth, then wires both meters into reporting your finance team already reads. We are a Gurugram technology consultancy founded in 2021, senior-led, CMMI Level 5 assessed, MSME certified and ISO 27001:2022 certified, and we work with Indian and global engineering organisations. If you are running Codex and Copilot side by side without knowing what either costs per pull request, talk to our engineering team. Related reading: Copilot AI credits drop on 1 September, choosing an AI coding agent harness, hybrid LLM routing for API spend and our model comparison of Gemini 3.5 Pro, GPT-5.6 and Claude Fable 5.
References
- OpenAI release notes - the 31 July 2026 Codex sunset notice for GPT-5.4 and GPT-5.4 mini, the 30 July 2026 desktop multi-repository review update, and the 30 July 2026 API changes.
- OpenAI API pricing - standard, batch, flex and fast-mode rates per model ID, cache-write costs, the data-residency uplift and the Fast mode rename.
- Codex changelog - Codex CLI and app releases, including review rules moving into
AGENTS.mdand auto-review behaviour.
- Models and pricing for GitHub Copilot - GitHub's AI credit rate card, the undisclosed model note for code review, and the absence of a batch tier.
- Usage-based billing for organizations and enterprises - included allowances, pooling, default paid usage and budget controls.
- GitHub Copilot is moving to usage-based billing - Mario Rodriguez, 27 April 2026, on Actions minutes for code review and the removal of fallback experiences.
- Copilot code review effort levels are generally available - GitHub Changelog, 7 August 2026, Lite and Balanced levels, organization defaults and effort labelling.
- Customize the reasoning level for Copilot cloud agent - GitHub Changelog, 3 August 2026, reasoning level and its credit consequence.
- Copilot usage metrics API adds agent app activity - GitHub Changelog, 7 August 2026, third-party agent reporting and the
agent_idguidance.
- Copilot impact dashboard adds a return on investment section - GitHub Changelog, 7 August 2026, cost per developer and its stated caveats.
- Copilot users can now see AI credits used per billing cycle - GitHub Changelog, 20 July 2026.
- Retiring the Copilot Billing Preview app - GitHub Changelog, 4 August 2026, where AI usage data now lives.
Last updated: 9 August 2026.
