ChatGPT Pro 5x vs 20x for Codex: compare rate limits, token burn, and when Plus + API or $100/$200 Pro fits agentic coding, monorepos, and parallel agents.
If you have been searching for a “Codex 100” or “Codex 200” plan, stop. Those SKUs do not exist. Codex is OpenAI’s coding agent—write, review, ship—bundled into ChatGPT plans. The $100 and $200 prices are ChatGPT Pro tiers that multiply Plus-level Codex and agentic usage by 5× and 20×. Same Pro feature set; the difference that matters for daily agentic coding is how much of the shared usage allowance you get.
One thing has to come first: as of 10 September 2026, Pro 20× ($200) is closed to new buyers. OpenAI temporarily paused new sign-ups and upgrades to the Pro $200 plan, including upgrades from Free, Go, Plus and Pro $100, citing demand for GPT-6 Astra after its 3 September 2026 launch. Existing $200 subscriptions keep renewing; Pro $100 is unaffected and remains purchasable. So the 5× versus 20× decision is only live for people who already hold $200. For everyone else the ladder now stops at 5×, and overflow means credits or an API key.
That reframes the question engineers are actually making right now: whether ChatGPT Pro 5x vs 20x Codex coding spend is justified, when Plus is enough, and when the API is the cleaner overflow valve.
How Pro and Codex actually connect
Codex is not a separate product. Sign in with a ChatGPT account and the same plan pool powers:
- Codex CLI
- IDE extension (VS Code and forks)
- Codex cloud, ChatGPT Work on the web, and the ChatGPT desktop app for local chats
- Mobile surfaces where the plan card allows it
Usage from Codex and other agentic features (ChatGPT Work, ChatGPT for Excel, Workspace Agents where available) draws from one shared agentic usage and credit pool. Local messages and cloud chats share your plan’s usage allowance; weekly limits may also apply. Hitting the wall mid-session is a plan-pool problem, not a “CLI-only” problem.
The five-hour window deserves a precise statement, because the docs and the product disagree. OpenAI’s pricing tables publish every range as “local messages per five-hour period.” But OpenAI removed the rolling five-hour cap in July 2026 and reinstated it on 25–26 August 2026 for Plus accounts only, leaving it switched off on Pro $100 and Pro $200 for the coming months. On Pro right now, the weekly limit is the ceiling that bites; on Plus, both apply.
Auth has two paths. ChatGPT sign-in gives the subscription allowance plus cloud integrations on eligible plans (GitHub review, Slack, and similar). An API key path runs CLI/SDK/IDE work on pure token billing—no subscription usage table, and no those cloud integrations.
Pro’s official framing is simple: choose 5× or 20× higher rate limits than Plus, starting from $100/month. Both Pro tiers include the same core Pro capabilities, with one published exception—unlimited desktop ChatGPT Voice is $200-only (Plus gets roughly 30 minutes–1 hour, Pro 5× about 3–6 hours), and Voice-started tasks still draw on your Codex budget. Pro also still lists GPT-5.3-Codex-Spark (research preview) with a separate usage limit, but do not weight it: OpenAI announced on 11 September 2026 that Spark would be retired the following week, and the docs simply have not caught up. You can move from $200 to $100 in Settings → My Plan; you cannot move the other way while the pause holds.
What 5× and 20× mean in practice
Plus is positioned for “a few focused coding sessions each week.” Pro multiplies those allowances. OpenAI publishes planning ranges—not hard message caps—for local messages per five-hour period:
| Model | Plus | Pro 5× ($100) | Pro 20× ($200) |
|---|---|---|---|
| GPT-6 Astra | 5–45 | 25–225 | 100–900 |
| GPT-5.6 Sol | 10–100 | 50–500 | 200–2,000 |
| GPT-5.6 Terra | 25–200 | 125–1,000 | 500–4,000 |
| GPT-5.6 Luna | 250–2,000 | 1,250–10,000 | 5,000–40,000 |
5× is five times Plus on the same model row. 20× is twenty times Plus—about four times the 5× tier on those tables, not “twice 5×.” Temporary promo multipliers in spring 2026 confused community math, and the “10× on $100” framing expired on 31 May 2026; current official docs state clean 5× / 20× versus Plus. Note the shape of the Astra row: OpenAI’s most capable model is also the one that empties a window fastest, and it is not available for Codex cloud tasks—Sol is what cloud chats run on. For the wider plan picture, see ChatGPT Pro $100 vs $200 usage explained.
Metering is token- and credit-based; OpenAI replaced per-message metering with API-aligned credits on 2 April 2026, and all but a small subset of Enterprise customers have been migrated off the legacy per-message rate card. Credits scale by model for input, cached input, and output. Rates per 1M tokens:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6 Astra | 250 | 25 | 1,250 |
| GPT-5.6 Sol | 100 | 10 | 500 |
| GPT-5.6 Terra | 50 | 5 | 300 |
| GPT-5.6 Luna | 5 | 0.5 | 30 |
Two footnotes matter for budgeting. GPT-5.6 Sol’s rate is promotional at least through 21 November 2026, so do not treat 100/10/500 as permanent. And Fast mode costs more—on GPT-6 Astra it is charged at 2.5× the Standard rate.
Output costs far more than input; fast mode burns faster. Small scripts sit at the high end of a range; monorepo loops, big context, tool use, and long outputs sit at the low end—context load at session start counts, so a nominal five-hour bucket can feel like roughly an hour of serious large-repo work. OpenAI puts a typical GPT-5.6 Sol Codex task at 5–30 credits, and notes average Codex spend on the order of ~$100–$200 per developer per month, with large variance by model, instances, automations, and fast mode. When you hit the limit, the in-flight turn may finish under fair use; then wait for reset, buy credits (eligible Plus/Pro), switch to a smaller model, or continue local work with an API key.
When Pro 5× still throttles
5× is not “unlimited agenting.” It multiplies Plus—and Plus was never designed for multi-agent monorepo days. If you are weighing a switch, compare Claude Code vs OpenAI Codex limits at $100 and $200.
Parallel agents. The pool is account-level, not per-agent. Concurrent CLI + IDE + cloud, or multi-agent orchestration, compete for the same weekly budget. Three Astra or Sol agents on fast mode can empty a 5× allowance that would feel comfortable for one disciplined session.
Big repositories. Large context, heavy retrieval, and fat project instructions raise credits per turn. You get fewer “messages” and more burn per message. Trimming AGENTS.md and disabling unused MCP servers is real money. Cached input is billed at a tenth of the input rate, which is why prompt caching for agent costs matters.
Extended / all-day sessions. Weekly pressure is the constraint on Pro today, and it does not reset at lunchtime. Output-heavy scaffolding, full-suite loops, and multi-file refactors hit the expensive side of the rate card. Luna buys volume at 5 credits per 1M input; Astra and Sol buy depth at a steeply higher rate.
With the 20× upgrade unavailable, the cheap levers stop being optional:
- Serialize agents instead of running many in parallel
- Downshift model for bulk work; reserve Astra and Sol for deep tasks
- Narrow scope and lean agent instructions; nest AGENTS.md rather than bloating one file
- Disable MCP servers you are not using
- Avoid full test suites every turn
- Leave Fast mode off unless the latency is worth 2.5×
- Top up credits for spikes
Plus + API: the control alternative
Staying on Plus ($20) with an API key for overflow is a common pro pattern—subscription for interactive Codex and cloud glue, tokens for CI, automation, and bursts. It is also, with $200 closed, the main way to buy more headroom than 5× gives you. The ChatGPT API vs Plus break-even math shows where token billing starts to win.
| Dimension | Plus / Pro (ChatGPT sign-in) | API key |
|---|---|---|
| Cap model | Shared plan allowance—weekly on Pro, weekly plus five-hour on Plus—then credits | Pay per token; no those subscription windows |
| Cloud features | Integrations on eligible plans | None of those cloud glue features |
| Cost shape | Fixed sub + optional credits | Variable, usage-linear |
| Best fit | Interactive agentic coding with integrations | CI, automation, shared runners, burst overflow |
Standard Business seats publish Plus-class Codex ranges per seat, while Business ($100) seats use the Pro 5× estimates, and Enterprise/Edu workspaces on flexible pricing have no fixed rate limits at all—usage scales with credits. Every plan's allowance is compared in Codex and GPT-6 usage limits by ChatGPT plan. Teams buy admin, SSO, and workspace controls, not a personal 20× multiplier.
Check live usage—ignore invented credit packs
Do not plan from forum screenshots or fixed “you get N credits” myths. Allowance is live and workload-dependent.
- Codex Settings → Usage — remaining allowance, credits, auto-reload where your role allows
- Product limit banners near the wall
- In an active CLI session,
/statusfor a plan/window snapshot - Historical spend on the OpenAI usage dashboard (lags; not a live meter)
Watch your weekly remaining first on Pro, the five-hour remaining too if you are on Plus, then the model, and whether fast mode is on. Pricing-page message ranges are planning estimates only, and they still describe five-hour buckets even where Pro is not currently metered that way.
Who should pick what
| Profile | Intensity signals | Choice |
|---|---|---|
| Occasional agentic | A few focused sessions/week; small PRs | Plus (+ credits/API for rare spikes) |
| Daily single-thread heavy | Full-time one main agent; Plus interrupts mid-week | Pro 5× ($100) |
| Parallel / multi-repo / all-day | Multiple agents, monorepos, frequent 5× exhaustion | Pro 5× plus credits or API overflow — 20× is closed to new buyers since 10 September 2026 |
| Already on Pro 20× | Continuous Codex, large repos, parallel agents | Keep it — letting it lapse means you cannot repurchase during the pause |
| CI / automation / cost control | Scripted jobs, predictable token bills | API key (± Plus for interactive) |
| Team / compliance | SSO, admin, no-training defaults | Business / Enterprise |
Upgrade ladder: Measure one to two weeks with Usage and /status. If Plus interrupts real work several times a week, move to 5×. If 5× still trips on normal parallel or big-repo days—not experimental thrashing—the move today is credits, model downshift, or hybrid API overflow, because the 20× rung is not purchasable. Optimize workflow before you spend: model downshift, narrower tasks, leaner agent instructions.
On published tables, 20× costs twice 5× but delivers about four times the 5× allowance. That arithmetic still holds, and it is why an existing 20× subscription is worth protecting—but it is not a purchase decision available to you right now. Holders deciding whether to keep it can weigh ChatGPT Pro vs Claude Max at $200.
Bottom line
Size the tier against how hard you actually push parallel agents, large repos, and long sessions, then check what you can still buy: since 10 September 2026 that is Plus or Pro 5×, plus credits and API keys. Use official planning ranges and token burn drivers—the current ones, with GPT-6 Astra at the top and Luna at 5 credits per 1M input—then verify with Usage and /status, not folklore.