Model availability: October 28, 2026
GPT-4.1 mini is deprecated, with shutdown scheduled January 14, 2027; OpenAI announced this October 14, so this corrects an earlier omission. Gemini 2.5 access is limited to prior active users; Google says these models are not deprecated and has not announced their shutdown date. That restriction’s effective date was not established. Named-model rates below were rechecked and are unchanged; availability is not guaranteed for a new project.
One million tokens do not have one universal price. Separate the model, input and output quantities, processing mode, and billable extras. For the three standard text API examples below, one million input tokens cost $0.30 to $1.00; one million output tokens cost $1.60 to $5.00. Those are selected examples, not the market’s minimum, maximum, or a quality ranking.
Official rates rechecked October 19, 2026, in USD. These are direct API text rates with uncached input, not a ChatGPT, Claude, or Gemini consumer subscription. Batch, priority, cached input, audio, tools, taxes, and regional add-ons are outside these calculations.
Three models, one consistent calculation
| Model and official rate source | One million input | One million output | 800,000 input + 200,000 output |
|---|---|---|---|
| OpenAI gpt-4.1-mini (deprecated) | $0.40 | $1.60 | $0.64 |
| Anthropic claude-haiku-4-5 | $1.00 | $5.00 | $1.80 |
| Google gemini-2.5-flash (restricted access) | $0.30 | $2.50 | $0.74 |
The mixed column is one million total tokens across a workload, not a recommendation to send a single 800,000-token prompt. Stay within the model’s request limits and check whether your actual request size enters a differently priced context tier. The Google output rate includes thinking tokens. These rows do not establish that the models perform equally well on your task.
The formula is: input tokens / 1,000,000 x input rate + output tokens / 1,000,000 x output rate.
For gpt-4.1-mini, the mixed example is 0.8 x $0.40 + 0.2 x $1.60 = $0.64. Applying only the input rate to every token would produce $0.40 and understate this particular workload by $0.24.
What would 1,000 requests cost?
Assume every request uses 2,000 billable input tokens and 500 billable output tokens. That produces 2 million input and 0.5 million output tokens. This is a deliberately fixed scenario, not a measured average user or a word-count conversion.
| Model | Input subtotal | Output subtotal | Total for 1,000 requests |
|---|---|---|---|
| gpt-4.1-mini | $0.80 | $0.80 | $1.60 |
| claude-haiku-4-5 | $2.00 | $2.50 | $4.50 |
| gemini-2.5-flash | $0.60 | $1.25 | $1.85 |
The Haiku example is 2 x $1 + 0.5 x $5 = $4.50. If all 1,000 requests are repeated once with the same billable quantities, this token-only scenario doubles to $9.00. Actual retries may have different context and output lengths, so use observed usage for your own forecast.
Why the lowest input rate may not give the lowest bill
In this sample, Gemini’s input rate is lower than gpt-4.1-mini’s, but its mixed-workload total is higher because output costs more. At the same fixed rates, their token-only costs match when output tokens are one ninth of input tokens: $0.10 x input millions = $0.90 x output millions. That arithmetic explains the reversal; it does not select a better model.
A useful comparison has a second denominator: cost per acceptable result. If a low-priced model needs repeated corrections, longer instructions, or more attempts, the initial per-token saving may disappear. Test a small representative workload and record both billable consumption and whether the result meets your requirements.
What this estimate leaves out
- Repeated conversation history and attachments can add input beyond the latest typed message.
- A token is not a word. Different tokenizers can count the same text differently; equal token counts do not mean equal content.
- Use the provider’s billable usage totals, including any billable thinking output, rather than only the text you can see.
- Search, execution, storage, and other tools may have separate charges. Consult the OpenAI pricing schedule, Claude pricing, and Google pricing for the features you actually enable.
- Cached input and cache writes need their own applicable rates. Do not count the same input once at the full rate and then add a full cached charge again.
For a working budget, first collect one completed task’s usage. Multiply it by expected task count, then add a clearly labeled contingency for rework and any fixed fees. Replace the contingency with actual evidence as the workload matures.
Subscriptions have different limits: see Claude usage resets and Cursor extra charges. Video uses a different billing unit; see AI video cost per minute.
Review record
October 19, 2026: rechecked the cited official sources and recalculated the displayed examples. Earlier observations below retain their dates. No account purchase, API benchmark or video generation was performed. AI sources enter a weekly review queue.
October 6, 2026: Initial rate observations and independently recalculated examples. No provider API calls were made and no output-quality benchmark was performed. Rates are observations, not a guarantee for future requests. Browse AI Costs & Limits.
October 28, 2026: re-read the linked official sources. Historical observations and calculations retain their original dates. No purchase, API execution or quality benchmark was performed. Next source-review target: November 4, 2026.