OpenAI Codex pricing: every plan, credit, and API cost explained

OpenAI Codex costs between $0 and $500 a month through ChatGPT plans: Free at $0, Go at $8, Plus at $20, and Pro tiers at $100, $200, and $500 since the September 29, 2026 DevDay reshuffle, with Business seats at $25 per seat monthly or $20 on annual billing, Business Premium seats at $125 monthly or $100 annual, and Enterprise pools above them 12345. Since April 2, 2026, plan usage bills through credits rather than messages 4, and the API bills separately at per-token rates, $1.75 per million input and $14.00 per million output on the Codex model, while GPT-6 Sol and its September 29 successor GPT-6.1 Sol undercut the GPT-5.6 Sol predecessor, listed at $4.00 in and $20.00 out per million, at $2.00 in and $10.00 out 6. The verdict this page defends: heavy users get tokens dramatically cheaper inside a seat than at the meter, and the crossover is computable from the published numbers.

The plan ladder: what each price actually adds

There is no standalone Codex subscription; Codex rides whatever ChatGPT plan you hold, on web, CLI, IDE extension, and iOS from Plus upward, with an API key as a fourth door that is automation-only, carries no cloud features, and bills at API rates 43. The CLI is free software drawing from the same pool as a web session in the same five-hour window, not a side door around plan pricing, and OpenAI's own pricing docs print that rule: local messages and cloud chats share one usage allowance 43. Read as deltas rather than feature lists:

• Free, $0: GPT-6 Luna at Standard speed in the desktop app, subject to rollout, for quick coding tasks; no card required 34
• Go, $8: light use on a weekly limit rhythm for lightweight coding tasks, no cloud task delegation, the background-agent feature that hands work off; Luna desktop access as on Free 453
• Plus, $20: cloud tasks, with OpenAI publishing per-model local-message ranges per five-hour window: 15 to 160 on GPT-6.1 Sol, 5 to 45 on GPT-6 Astra, 15 to 150 on GPT-6 Sol, 350 to 3,000 on GPT-6 Luna, with cloud chats using more allowance than local messages 1; since this page's October 6 read the estimates table has narrowed to those four GPT-6 rows, dropping the 10-to-100 GPT-5.6 Sol range it carried alongside GPT-5.5's 15 to 80, GPT-5.5 having left every plan on October 14, though the rate card keeps GPT-5.6 Sol, Terra, and Luna credit rates, which OpenAI notes remain unchanged 3; independent trackers had observed roughly 10 to 60 cloud tasks per window, with one Reddit user reporting 30 to 150 requests 4
• Pro 100, $100: five times Plus usage headroom, a multiplier the pricing page itself no longer prints since the September reshuffle and that trackers relay from OpenAI's own head of product 73
• Pro 200, $200: reopened to new subscribers on September 29 at a lower allowance, ten times Plus rather than the old twenty, per OpenAI's head of product as relayed by trackers, with subscribers active between September 22 and 10 a.m. Pacific on September 29 grandfathered on the old allowance through October 29 3
• Pro 500, $500: the new top tier, twenty-five times Plus usage, OpenAI's highest included allowance, and the only Pro plan with Ultrafast, which no amount of credit buying unlocks on the cheaper tiers 3

A reader holding a pre-April guide sees a single $200 Pro plan; OpenAI split it into Pro 5x and Pro 20x on April 9, 2026 so heavy users stop subsidizing light ones 4, then reshuffled the ladder again on September 29, 2026: Pro 200 reopened at a lower allowance, and a $500 Pro 500 tier took the top spot with Ultrafast attached 3. Priced per unit of headroom, the ladder lost its one bargain in that reshuffle: at the tracker-relayed multipliers, Pro 100 sells five Plus-multiples for $100, Pro 200 ten for $200, and Pro 500 twenty-five for $500, so every Pro rung now charges the same $20 per multiple of Plus usage that Plus itself charges, and the old Pro 20x's $10-per-multiple deal survives only for subscribers grandfathered through October 29 (ProvenBrief's arithmetic on the published prices and relayed multipliers). Pro plans also currently carry no five-hour limit, which turns the Pro allowance into a weekly budget spent at your own pace 3. Two seat families sit outside the personal ladder: Business Standard seats at $25 per seat monthly or $20 per seat on annual billing for 2 to 200 seats 35, and Business Premium seats at $125 monthly or $100 annual, which OpenAI prices at five times Standard included usage with no five-hour limit, the relationship OpenAI announced for the tier and that October reads of the business pricing page and third-party trackers still print, firming up Quesma's report of the August 25 launch at $125 to match Anthropic's identically priced tier 385. Enterprise, in CloudZero's framing, replaces fixed limits with a custom seat price plus a shared organization-wide credit pool billed at API list rates, and OpenAI's own docs confirm Enterprise agreements can bill credit-based or usage-based in dollars 453. OpenAI's own estimate for active use, relayed via CloudZero, runs $100 to $200 per developer per month 4.

Credits are the meter: what a task really costs

Credits are token billing under another name. Three token types feed the meter: input, cached input at a tenth of fresh input on every text-model row of both current rate cards except GPT-6.1 Sol, whose cached rate is half that, a twentieth of fresh input on both surfaces at 2.5 credits against 50 on the credit card and $0.10 against $2.00 per million on the API, and output at five to eight times input depending on family, 5x across the GPT-6 line and GPT-5.6 Sol, 6x on GPT-5.6 Terra and GPT-5.6 Luna, 8x on GPT-5.3-Codex 364. A typical GPT-5.6 Sol task still runs 2 to 15 credits by OpenAI's published average, a band this page carried at 5 to 30 last month 3; GPT-5.5 rode a 15-to-80 local-message range until the model left every plan on October 14, retired from Codex with ChatGPT sign-in at noon PDT while the API and API-key-authenticated Codex sessions keep access 39.

The dollar value of a credit is pinned by models priced on both surfaces, and the pin now holds across the whole GPT-6 and GPT-5.6 lines. OpenAI's credit rate card bills GPT-6 Astra at 250 credits per million input tokens and 1,250 per million output, GPT-6.1 Sol at 50 and 250, GPT-6 Sol at 50 and 250, GPT-6 Luna at 2.5 and 12.5, GPT-5.6 Sol at 100 and 500, GPT-5.6 Terra at 50 and 300, and GPT-5.6 Luna at 5 and 30 3; the API lists those seven models at $10.00 and $50.00, $2.00 and $10.00, $2.00 and $10.00, $0.10 and $0.50, $4.00 and $20.00, $2.00 and $12.00, and $0.20 and $1.20 per million 6. Every ratio divides to exactly $0.04 per credit, extending the cross-check CloudZero first ran on GPT-5.3-Codex's 43.75 and 350 credits against $1.75 and $14.00 46: a 2-to-15-credit task is $0.08 to $0.60 at list value, at or below the low end of the $0.50 to $2.00 typical API session CloudZero reports 4. Codex credit billing adds no separate cache-write charge 3. Data-residency endpoints, where eligible, add a 10 percent uplift on models released on or after March 5, 2026, and FedRAMP endpoints now carry the same 10 percent 6.

Limits sit in five-hour windows on Plus and Standard Business, while Pro plans currently have no five-hour limit at all, weekly limits may also apply, and Codex lets you finish the turn in progress when you hit the wall, subject to fair use, after which Plus and Pro users can buy extra credits without upgrading plans, Business, Edu, and Enterprise workspaces on flexible pricing can buy workspace credits, and any user can fall back to an API key at standard API rates 334. The five-hour rhythm also fixes the month's arithmetic, which none of the vendor surfaces runs: 30 days hold 144 windows, so the vendor-published GPT-6 Sol range of 15 to 150 local messages per window on Plus implies a theoretical ceiling of 2,160 to 21,600 included messages a month, 2,160 to 23,040 on GPT-6.1 Sol's 15 to 160. Spread $20 across that ceiling and an included message costs 0.09 to 0.93 cents, against the $0.08 to $0.60 of list value a typical 2-to-15-credit task carries, an embedded discount running from roughly 9-fold at the conservative pairing to roughly 670-fold at the generous one, a narrower spread than this page's September figures because the typical task got cheaper (ProvenBrief's arithmetic; a real month never maxes every window, and cloud chats cost more allowance than local messages). Independent trackers' 10-to-60 cloud tasks per window sits inside the same order 4. Image generation burns included limits three to five times faster than a normal turn 34, and desktop voice draws the same budget at $0.05 a minute, or 1.25 credits a minute under credit billing 3.

Speed modes are the multiplier to watch, and the fast lane widened since this page's October 6 read: the speed table now prints two Ultrafast rows, GPT-6 Astra Ultrafast and GPT-6.1 Sol Ultrafast, each drawing included subscription usage at 8x and billing purchased credits and Enterprise pay-as-you-go at 6x, while Fast mode draws included usage at 2.5 times the Standard rate and bills credits at 2x; OpenAI prints that table itself and labels the numbers billing multipliers, not speed increases 3. On the API, Fast still doubles rates across the board, the Codex model to $3.50 in and $28.00 out per million and GPT-6.1 Sol to $4.00 and $20.00, while Ultrafast lists GPT-6 Astra at $60.00 in and $300.00 out per million and GPT-6.1 Sol at $12.00 and $60.00 6. At the $0.04-per-credit list value, the plan-side 2.5x turns a typical 2-to-15-credit task into 5 to 37.5 credits, $0.20 to $1.50 of list value, to finish the same work sooner (ProvenBrief's arithmetic on the vendor-printed multiplier and the GPT-5.6 Sol task average). Ultrafast stays gated: among self-serve plans it reaches only Pro 500, buying credits on Pro 100 or Pro 200 does not unlock it, eligible Enterprise and Edu workspaces get it off by default until an owner enables it, and workspaces that require inference residency outside the United States cannot get it at all 3310. Quesma's refactor accidentally ran its first half in fast mode, $110 of a $283 day, and Business organizations cannot disable fast mode org-wide, where Anthropic requires organizational approval before anyone turns it on 5.

The arithmetic the pricing pages leave to you

The cited guides print the seat ladder and the token meter separately; the crossover between them is this page's arithmetic. Quesma's Jacek Migdal measured one refactor-class pull request at $283.44 of tokens at list rates, consuming $270 of allowed overflow inside 24 hours 5, which pins overflow at roughly list pricing, a pin that rests on that single measured PR. A SemiAnalysis benchmark Quesma cites found a $200-a-month plan delivered up to $14,000 a month in tokens 5. Divide one by the other and included usage prices a list-rate dollar at about 1.4 cents, while overflow prices it at about a dollar (ProvenBrief's arithmetic), with one October caveat: that benchmark measured the pre-reshuffle twenty-x allowance, and on the ten-x tier new Pro 200 subscribers get, the implied ceiling halves toward $7,000 and the included dollar rises toward 2.9 cents (ProvenBrief's adjustment; the token multiple rests on a benchmark month that may also have benefited from extra limit resets granted by OpenAI, as Quesma notes 5). The upgrade gap from Plus to Pro 200 is $180 a month, and one Quesma-class PR in overflow bills at $270. The rule that falls out: a month containing even one heavy PR past your Plus allowance makes the $200 seat cheaper than the $20 seat plus overflow.

Assembled, the system has two regimes no single cited source states: inside an allowance, a list-rate dollar costs about 1.4 cents, the $200 plan against its $14,000 token ceiling as benchmarked at the old allowance; outside one, every meter converges on a full dollar, overflow at roughly list, Enterprise pools billed at API list rates, and the API itself at list. It falls out of dividing numbers each vendor surface carries separately, and it compresses the purchase decision into one sentence: buy enough seat to cover ordinary weeks, because every path past the allowance charges the same list the seat exists to avoid. The GPT-6 and GPT-5.6 rate cards tighten the pin rather than moving it: every model named above that is priced on both surfaces converts at the same $0.04 per credit 36.

The API has its own crossover, published as directional by CloudZero: below roughly 10 to 40 sessions a month the API tends to beat Plus, below 50 to 200 it can beat the $100 Pro tier, and above that a flat subscription usually wins 4. The API route drops automatic GitHub code review, Slack integration, and, typically, day-one access to new models 4.

Codex versus Claude Code: same prices, different meter

The price grid still mirrors at the price points: free entry, $20 mid tier, $100 and $200 power tiers, $25 team seats that both vendors discount to $20 on annual billing, and $125 premium seats, OpenAI's added August 25, 2026 to match Anthropic's identically priced tier 4533. The multipliers no longer mirror: Anthropic's Max plan still tops out at Max 20x for $200 a month, which since September 29 out-headrooms OpenAI's ten-x Pro 200 for new subscribers, while OpenAI's twenty-five-x answer lives a rung higher at $500 7. The meters differ: CloudZero describes Codex as optimized for background agent work that reads many files before writing few, and Claude Code as spending tokens in interactive, turn-by-turn sessions 4. On the budget tier, Quesma finds ChatGPT Go's weekly limit rhythm more usable than Claude Pro's five-hour limit, which it reports often cannot cover a basic pull request in a big repository 5. Most engineers hedge rather than choose: 70 percent run two to four AI tools at once in the Pragmatic Engineer survey Quesma cites 5.

Every dated change, on one line each

• August 2021: OpenAI ships the original Codex model, first to market with an LLM generating usable programming code 5
• April 2, 2026: per-message billing retires for Plus, Pro, and Business, replaced by credits 4; the same day OpenAI announced pay-as-you-go Codex-only seats for Business and Enterprise workspaces and cut annual Business seats from $25 to $20 11
• April 9, 2026: the single $200 Pro plan splits into Pro 5x at $100 and Pro 20x at $200 4
• April 23, 2026: legacy Enterprise, Edu, Health, and Gov plans move onto credits 4
• June 24, 2026: pay-as-you-go Codex-only seats stop being offered to new Business signups; existing seats keep working, and standard Business seats still include Codex within plan limits 114
• July 30, 2026: Priority processing is renamed Fast mode in the API 6
• August 2026: paid limit resets appear at $8 on Plus, $40 on Pro 5x, and $80 on Pro 20x, the two Pro tiers now named Pro 100 and Pro 200 5
• August 25, 2026: Business Premium seats launch at $125 per seat to match Anthropic's identically priced tier 5
• September 3, 2026: GPT-6 Astra is released at $10.00 in and $50.00 out per million on the API 126
• September 22, 2026: GPT-6 Sol and GPT-6 Luna roll out to Codex and ChatGPT Work at lower token prices than their GPT-5.6 predecessors, reaching Plus, Pro, Business, Enterprise, and Edu users, with Luna on Free and Go in the desktop app 13
• September 29, 2026: at DevDay, Pro 500 launches at $500 with OpenAI's highest included allowance and exclusive Astra Ultrafast, while Pro 200 reopens to new subscribers at a lower allowance, grandfathering subscribers active from September 22 through 10 a.m. Pacific that day 27
• September 29, 2026: GPT-6.1 Sol launches in ChatGPT Work and Codex, not Chat, and on the API at $2.00 in and $10.00 out per million, carrying a 15-to-160 local-message range on Plus 143
• October 14, 2026: GPT-5.5 retired from ChatGPT, ChatGPT Work, and Codex on all plans at noon PDT; the API and API-key-authenticated Codex sessions are unaffected; OpenAI points paid plans to GPT-6 Sol and Free and Go to GPT-6 Luna 39
• October 29, 2026 (scheduled): grandfathered Pro 200 allowances end, and every active Pro 200 subscription moves to the lower included usage at the same $200 price 3
• November 21, 2026: GPT-5.6 Sol promotional pricing runs at least through this date 6

Note what the sequence does to the April 2 seat program: announced as flexible pay-as-you-go access, closed to new Business signups by June 24, while Quesma's August signup shows seat-based Business access continuing at $25 per seat with usage included 45. Anyone budgeting off the announcement alone is reading a door that has since narrowed. And the model churn has a budget consequence: the task anchor this page long used, 5 to 30 credits on GPT-5.6, is superseded by OpenAI's current 2-to-15-credit average for a typical GPT-5.6 Sol task, and GPT-5.5 has now left every plan, on October 14, so credit budgets built on either should be re-run on the current average or GPT-6's published rates 33.

Who should buy what

The two regimes do the sorting. Free answers whether agentic coding sticks at all. Go at $8 suits light use that never needs cloud delegation. Plus at $20 is the practical default for developers 4. Pro 100, Pro 200, and Pro 500 are for genuinely heavy, parallel workloads, and since September 29 all three price a unit of headroom identically, so the choice is how much scale you need, with Pro 500 adding Ultrafast and grandfathered Pro 200 holders keeping the old half-price headroom through October 29 3. Sparse users can skip seats for the API under the session thresholds above, accepting the convenience losses 4. Teams pilot on Business Standard seats at $25 monthly or $20 annual and graduate to Premium seats, $125 monthly or $100 annual at five times Standard usage, or to an Enterprise pool when usage runs uneven 53. Quesma's bottom line after paying real money: stay on seat-based plans as long as you can, and watch the fast-mode default 5. Neither cited guide prints regional plan prices such as an India-specific ladder; those live on OpenAI's own bot-gated plan pages, which this page's refresh loop re-reads when they open.

Verified October 2026.