On this page
OpenAI Codex Pricing Explained: What You'll Really Pay 2026
OpenAI Codex's April 2026 token billing overhaul created massive cost variance for engineering teams, with the $20 Plus tier functioning as a short-term trial rather than a sustainable plan. Most regular users need the $100 Pro 5x tier to avoid excessive overage fees, with realistic monthly spend ranging from $100 to $200 per developer.
OpenAI Codex crossed 5 million weekly active users on June 1, 2026 — just six weeks after hitting 3 million. That growth didn’t happen because pricing got simpler. It happened despite a deliberate April 2026 billing overhaul that replaced the old per-message metering with token-based credit billing, creating a system where two developers on the same $20 Plus plan can see effective monthly costs ranging from pocket change to over $200. If you’re trying to forecast what Codex will actually cost your team, the sticker price is the least important number on the page.
The Dual-Track Billing Architecture
Codex doesn’t have one pricing model. It has two entirely separate billing paths, and which one you’re on determines everything: your rate limits, your model access, your feature availability, and how predictable your monthly bill looks.
Track 1 — ChatGPT subscription (most developers). Codex is bundled into every ChatGPT plan from Free through Enterprise. You pay a monthly fee and get a usage allocation that resets on a rolling 5-hour window. This track includes the cloud features that make Codex function as an agent: GitHub integration, automatic PR code review, Slack integration, and long-horizon task execution in isolated sandboxes. If you’re using Codex as a daily engineering tool, this is almost certainly your path.
Track 2 — API key (pay-per-token). Developers building Codex-powered tools, integrating it into CI/CD pipelines, or automating workflows can use an API key and pay per token. The tradeoff is stark: no cloud features, no Codex Cloud, no GitHub review, and newer models arrive later through the API than through ChatGPT subscriptions. The gpt-5.3-codex model bills at per-token rates — but you lose the entire agent infrastructure that makes Codex more than a chatbot with a code interpreter.
The subscription track is where the real pricing complexity lives, and where most teams will spend their time trying to understand what they’re actually paying for.
The April 2026 Token Billing Overhaul
On April 2, 2026, OpenAI switched Codex billing for Plus, Pro, and Business plans from per-message to token-based credit billing. The old model was structurally similar to GitHub Copilot’s premium request system: one message roughly equaled one fixed cost. The new model connects usage directly to token consumption — input tokens, cached input tokens, and output tokens each carry separate credit rates.
The practical impact is a 9x cost variance between task types. A one-line code fix burns approximately 5 credits. A multi-file refactor that reads 30 files and generates 10 modified files costs around 45 credits. Both count as “one Codex task” under the old mental model. Under the new system, they’re radically different line items.
GPT-5.5 usage burns 125 credits per 1 million input tokens and 750 credits per 1 million output tokens, with credits tracking API token prices at roughly 4 cents each. GPT-5.4 is cheaper at 62.50 credits per 1M input and 375 per 1M output. GPT-5.4-Mini drops further to 18.75 and 113 respectively. The model you pick — and whether you’re reading large files or generating long outputs — matters far more than which plan tier you’re on.
OpenAI itself estimates average monthly Codex spend at $100-$200 per developer for teams using it as a primary engineering tool. That’s not a per-session number. That’s a monthly average, and the variance around it is substantial.
Plan Tiers and What They Actually Include
As of June 2026, Codex is bundled into ChatGPT plans with no standalone subscription option. The full ladder:
| Plan | Price | Codex Messages / 5h Window (GPT-5.5) | Key Notes |
|---|---|---|---|
| Free | $0/mo | Very limited | No cloud integrations |
| Go | $8/mo | Low | Lightweight tasks only |
| Plus | $20/mo | 15–80 | Credit top-ups available |
| Pro 5x | $100/mo | 75–400 | GPT-5.3-Codex-Spark preview access |
| Pro 20x | $200/mo | 300–1,600 | Highest consumer limits |
| Business | $20/user/mo (annual) or $25/user/mo (monthly) | Same as Plus base | Cloud features, admin controls, SSO |
| Enterprise / Edu | Custom | No fixed limits | Shared credit pool, SCIM, EKM |
The message counts per 5-hour window deserve scrutiny. Those are ranges, not fixed allocations. Where you land inside the range depends on task complexity, model choice, and how much context each task consumes. The 5-hour rolling window also means you can burn your full allocation in a single intense morning session and wait — not until tomorrow, but until the specific requests you made start aging out of the window.
For teams evaluating the jump from Plus to Pro 5x, the math is straightforward at the subscription level but murky in practice. A 50-developer team on ChatGPT Business at annual pricing incurs $12,000 in annual subscription costs (50 × $20 × 12). Moving everyone to Pro 5x tiers pushes that to $60,000 annually (50 × $100 × 12) — a 5x multiplier that matches the usage multiplier on paper but may not match actual utilization for every seat.
The Pro 5x Tier Is the New Minimum for Regular Use
Here’s the uncomfortable truth about the $20 Plus tier: 15 to 80 GPT-5.5 messages per 5-hour window sounds reasonable until you actually try to use Codex as a primary tool. A single agentic workflow — the kind where Codex reads multiple files, chains tool calls, and produces a diff — can consume multiple messages within one task. Developers running regular agentic workflows report exhausting Plus limits within the first week, and the cost of frequent credit top-ups pushes effective monthly spend above $100 within the first month.
The Pro 5x tier at $100/month isn’t a luxury upgrade for power users. For anyone using Codex daily for real engineering work, it’s the minimum viable plan. The 75-400 message window provides enough headroom that you’re not constantly rationing tasks, and the effective per-task cost drops because you’re not paying premium rates for overage credits.
This is the pattern I’ve seen across the pricing data: the entry-level tier functions as a trial, not a sustainable plan. Teams that stick with Codex migrate to Pro within the first billing cycle or two.
The Pay-You-Go Reversal and Bedrock Expansion
OpenAI’s April 2026 announcement introduced pay-as-you-go Codex-only seats for Business and Enterprise — no fixed seat fee, token-based billing, no rate limits, plus a promotional $100 in credits per new team member (up to $500 per team). It was framed as a scalable path for small teams to pilot Codex without commitment.
On June 24, 2026, OpenAI discontinued new pay-as-you-go Codex-only seats for Business plans. Existing seats are unaffected, but the product is no longer available for new team adoption. The earlier framing of pay-as-you-go as a permanent, scalable option for team adoption directly contradicts this update. If you’re evaluating Codex for a team and were counting on pay-as-you-go seats as your entry point, that door has closed.
Simultaneously, Codex expanded to Amazon Bedrock on June 1, 2026, with GPT-5.5, GPT-5.4, and Codex itself reaching general availability at per-token pricing identical to OpenAI’s direct rates — no AWS platform markup. Usage counts toward existing AWS spend commitments. For engineering leaders already committed to AWS infrastructure, this eliminates the procurement friction of adding a separate OpenAI vendor relationship. Codex on Bedrock inherits AWS enterprise controls (IAM, encryption, VPC isolation, CloudTrail logging) and operates in the customer’s selected AWS region for data residency.
The Bedrock launch matters for pricing predictability. Pure pay-per-token usage with no seat licenses means teams that have been blocked by per-seat procurement cycles can now consume Codex as a line item on existing cloud commitments. Whether that actually reduces your total spend depends on your utilization pattern, but it removes a structural barrier to adoption.
The Rate Limit Reset Gap
On June 12, 2026, OpenAI introduced bankable rate limit resets for Go, Plus, Pro, and Business Codex users. Instead of waiting for the 5-hour window to refill, you can now save resets for later use. Each user gets one free reset, with additional resets available through a referral program (up to 3 friends through June 24, 2026).
The feature was announced as available for all paid tiers. There’s just one problem: as of June 22, 2026, the bankable rate limit reset feature is only available via Windows and macOS desktop Codex apps — Linux CLI users have no way to access it. Multiple verified reports in the OpenAI Developer Community confirm this gap, with Linux users describing it as effectively locked out of a feature marketed as universally available.
If you’re running Codex on Linux (which, let’s be honest, is a significant portion of the developer audience), this feature doesn’t exist for you yet. Plan your usage patterns accordingly.
What This Means for Your Decision
The Codex pricing structure in 2026 rewards commitment and punishes casual usage. The Plus tier is a trial that becomes expensive quickly. The Pro 5x tier is the actual starting point for regular use. The Business tier makes sense only if you need the admin controls and cloud integrations — not if you’re optimizing purely for cost. And the Bedrock option is worth serious evaluation if your organization already has AWS spend commitments to draw down.
If you’re choosing between Codex and a competitor like Claude Code, the pricing comparison is more nuanced than the headline $20-per-month numbers suggest. Our analysis of Claude Code vs OpenAI Codex metering architectures covers why identical sticker prices hide fundamentally different usage patterns — and why dual subscriptions may be the most cost-effective choice for professional teams.
For teams already committed to the OpenAI ecosystem, the question isn’t whether to use Codex but which tier to commit to. Start with Pro 5x for any developer using Codex more than a few times per week. Budget $100-$200 per developer per month as a realistic range. And watch your token consumption for the first two billing cycles before setting expectations — the variance between task types is large enough that your actual spend will depend far more on what you’re building than which plan you picked.