As checked 2026-10-25: Two OpenAI model-specific primary documentation pages were checked on October 25, 2026. This is a dated economics guide, not a permanent price guarantee or a personal hands-on benchmark.
Scope: global English reference; USD list prices where shown; local eligibility, currency, taxes, negotiated terms and model release dates require separate confirmation. Dynamic source-page prices have a check date, not a verified initial effective date.
The scope of this price snapshot
Two OpenAI model-specific primary documentation pages were checked on October 25, 2026. The GPT-5.4 model page displayed standard text rates of $2.50 per million input tokens, $0.25 per million cached input tokens and $15 per million output tokens. The GPT-5.5 model page displayed $5, $0.50 and $30 respectively. These are observations of current first-party model pages, not a claim about when those tariffs began. The pages also call out nonstandard rules for sufficiently long prompts, so their headline rates are not a complete invoice prediction.
Example of an actual dimensional comparison
Consider a proposed month of 10 million fresh input tokens and 3 million output tokens with no caching and no tools. The documented headline schedule gives GPT-5.4 a calculated text-token component of 10 × $2.50 + 3 × $15 = $70. For GPT-5.5 the same hypothetical token counts imply 10 × $5 + 3 × $30 = $140. This is only an arithmetic comparison under standard short-context rates, not a measured quality comparison or a promise that these exact models can complete the same tasks with equal token counts.
Watch the whole-request long-context rule
The GPT-5.4 and GPT-5.5 pages specify a threshold above 272,000 input tokens with higher rates for the full session. A cost estimator that calculates most tokens at short-context rates and just the excess at a premium would misread the documented rule. Test whole-request threshold selection, and account for the difference between the supplied input and the output tokens actually generated. Regional processing can also have an uplift where the offer supports data residency.
Processing tiers are separate products
Standard calls are not automatically batch calls. OpenAI’s Batch API documentation presents an asynchronous route with discounted processing and a completion-time policy. For a nightly classification workload, compare exact batch model rates and queue constraints; for an interactive application with a human waiting, standard or faster service may be needed. Tool charges such as web search, containers and file search can be additional and must be priced from the same dated OpenAI tariff rather than inferred from the base text tokens.
Selection and purchasing caution
The cheaper model on this narrow input/output scenario is not necessarily a cheaper accepted answer. Validate quality, latency, token consumption, retry rate and compliance for the task. A consumer ChatGPT plan is a different purchasing channel from this API billing ledger. Model pages are live documents; archive and re-verify them before publishing fresh price claims, and keep unknown contract and regional values null.
A production-side calculator query
For an editorial backend, isolate the first-pass drafting calls from research retrieval and validation. Suppose the application has one million cached input tokens in addition to the example’s uncached traffic. On the model pages’ cited headline schedules, the cached-input component is billed at its own rate and must not be included again in fresh input. Before calculating a final bill, check whether cache eligibility requirements were actually met and whether the request crossed the long-context threshold. A source-review call that uses a paid web tool must be accounted for separately.
Operational risk outside token price
Budget output validation, rejected JSON, refusal or safety conditions, reruns after model upgrades, latency limits and changing model aliases. Record the concrete version used for every job rather than preserving only a consumer-facing marketing title. An API-model tariff is a useful procurement reference but does not guarantee product suitability, uptime in a given country or the same features available through a ChatGPT consumer seat. If a model is retired or a regional multiplier changes, append a new snapshot and leave prior invoices linked to the earlier rate ID.
Decision checklist
- Identify the specific plan or API endpoint, model version, region and currency before pricing a job.
- Measure input, output, caching, generated media and failed-job units separately.
- Check the official pricing page again before procurement, especially if the SKU has no effective-date notice.
- Keep source evidence and a timestamped price ID; do not treat unknowns as zero.
Pricing evidence boundary
This cost discussion concerns OpenAI. Reproduce a quote only with its exact model or plan identifier, chargeable unit, billing region, processing mode and timestamp; a changing vendor rate card is not a permanent price guarantee. If any one of these dimensions is missing, keep the associated rate and worked scenario under editorial review instead of filling the gap with an estimate.
Sources & Further Reading
- gpt-5.4 text-token standard list rates and threshold above 272k input tokens — official first-party source, accessed 2026-10-25; original effective date not specified in source.
- gpt-5.5 text-token standard list rates and threshold above 272k input tokens — official first-party source, accessed 2026-10-25; original effective date not specified in source.
- API token rates depend on exact tier, modality, model, tool and region — official first-party source, accessed 2026-10-25; original effective date not specified in source.
- Batch API asynchronously handles eligible requests with stated cost reduction and turnaround — official first-party source, accessed 2026-10-25; original effective date not specified in source.