Separate authentication, usage and API costs
Work and Codex share usage and limits within ChatGPT access; ChatGPT for Word also uses the relevant allowance. Opening another working interface therefore does not create a separate free budget. API usage follows its own billing. GPT-6 Sol and Luna have different API prices, with further differences by context length and processing tier. Do not apply these token prices to a ChatGPT subscription. Record account, model, applicable price source and remaining allowance for the pilot.
Use the accepted change as the cost unit
Record consumption plus preparation, correction, testing and review time for several comparable tasks. Do not compare a difficult defect with a simple text change. The denominator should be an accepted task or another clearly defined deliverable. This reveals whether low model consumption merely creates more human rework.
Account for peaks and operating effort
Separate recurring cost from one-off setup. Add parallel tasks, retries and maintenance as explicit assumptions. Show a baseline and a higher-usage scenario. Verify actual limits and consumption controls in the account used. An example calculation is a planning aid, not a guaranteed monthly invoice.
Pro 200: September 2026 allowance change
As of 29 September, Pro 200 subscriptions are available again with less included usage at the unchanged $200 monthly US list price. Eligible existing subscribers retain the earlier allowance through 29 October. Current help does not quantify the reduction. Codex estimates labelled Pro 5x/20x, checked before DevDay, do not establish a comparable before-and-after weekly limit. Check your account and change notification before budgeting a plan switch.
Budget separately for Pro 500 and Ultrafast
On 29 September, OpenAI announced Pro 500 at US$500 per month. Initially available to subscribe to on the web, it provides 25 times the included usage allowance of Plus according to the DevDay recap, alongside Astra Ultrafast in Work and Codex. This multiplier is not a fixed token or message count. Among Pro subscriptions, Ultrafast is exclusive to Pro 500; buying credits does not unlock it on Pro 100 or Pro 200. Pro 200 returning with a smaller allowance is a separate change. Public sources still do not reliably quantify that reduction.
Ultrafast runs GPT-6 Astra at a faster service tier. Speed claims concern token generation and do not guarantee an equivalent reduction in every complete task. The Work/Codex documentation specifies that included usage is consumed at eight times the Standard rate; purchased credits and Enterprise pay-as-you-go usage use a sixfold rate. Eligible Enterprise and Edu plans can also qualify. Ultrafast starts disabled in Enterprise. Required inference residency outside the United States rules it out; the company’s location alone does not determine eligibility.
Decision matrix
| Decision point | Proceed when | Stop when |
|---|---|---|
| Cost assumptions | Documented: current plan, usage pattern, review time and avoided work. | Scope, data or accountability remains unresolved. |
| Measurement baseline | A cost calculation separates verified plan assumptions, observed usage and measured review time. The comparison uses the same delivered engineering task and makes uncertainty visible. | There is only an unevaluated demo without acceptance evidence. |
| Budget decision | Owner, approval, fallback and next review date are defined. | Avoid: stale price values without a review date. Undated prices or unmeasured usage assumptions must not be presented as reliable budget commitments or guaranteed savings. |
Keep it verifiable
Primary sources
- OpenAI: Codex and Work pricingSource checked:
- FinOps Foundation: Unit economicsSource checked:
- OpenAI: ChatGPT for WordSource checked:
- OpenAI: GPT-6 SolSource checked:
- OpenAI: GPT-6 LunaSource checked:
- OpenAI: About ChatGPT Pro tiers — transition termsSource checked:
- OpenAI: ChatGPT release notes — 29 September 2026Source checked:
- OpenAI: Speed — Astra Ultrafast usage and residencySource checked:
