Answer in brief
October’s AI API costs reveal different choices for budget and latency. A fixed workload calculation and independent measurements make the trade-offs visible without pretending quality is equal.
October budgets meet a moving price calendar
AI API costs now reward reading both the unit price and its expiry date. OpenAI and Google’s official pricing pages, checked on 1 October, show different economics for everyday production requests. Google also specifies a January change for Gemini 3.8 Flash. A budget approved using the current rate can therefore become outdated without any change in traffic.
The practical question is whether an endpoint delivers an accepted result within the product’s waiting-time budget. Customer chat, document processing and overnight classification need different response patterns. A low token price can suit work that tolerates delay, while a faster stream may matter when a person watches a long answer arrive. Neither alone establishes suitable quality.
Quoted rates and calculated costs stay separate
The first table uses standard paid text rates and a transparent scenario: each call consumes 10,000 native input tokens and 2,000 total billable output tokens. Multiply each quantity by its quoted rate per million, add the amounts, then multiply by the request count. OpenAI’s short-context column applies to this Luna scenario. The resulting totals are editorial calculations.
The output allowance includes any reasoning tokens charged as output; it does not promise 2,000 visible answer tokens. There is no cache saving, tool charge, batch discount or tax in this baseline. Repeated attempts and longer reasoning change actual consumption. Holding quantities fixed makes arithmetic legible, but it does not claim that all models solve the same request equally well.
| Model / rate period | Quoted input USD / million | Quoted output USD / million | Calculated USD / call | Calculated USD / 10,000 calls |
|---|---|---|---|---|
| GPT-6 Luna | 0.10 | 0.50 | 0.002 | 20 |
| Gemini 3.5 Flash-Lite | 0.30 | 2.50 | 0.008 | 80 |
| Gemini 3.8 Flash — through 2026-12-31 | 0.75 | 3.75 | 0.015 | 150 |
| Gemini 3.8 Flash — scheduled from 2027-01-01 | 1.50 | 7.50 | 0.030 | 300 |
Stream speed does not erase the initial wait
Artificial Analysis supplies the second table’s measurements through the named first-party APIs. Its performance method version 2.2.0, dated 2 March 2026, uses a default 10,000-input-token workload. The October observation retains Luna’s max and Flash’s high variants. Flash-Lite’s reasoning effort is not specified in its displayed name, so that setting remains unknown.
Output speed describes the stream after it starts. The latency column keeps the evaluator’s FAQ label, time to first token; its charts also distinguish time to first answer token. These concepts cannot be silently exchanged for a full task duration. Exact sample dates are unknown. VJOURNAL has not run these tests, and the figures are not regional service guarantees.
| Model / measured variant | Output tokens / second | Reported TTFT seconds | Measured endpoint |
|---|---|---|---|
| GPT-6 Luna (max) | 125.2 | 118.03 | OpenAI |
| Gemini 3.8 Flash (high) | 221.0 | 17.45 | |
| Gemini 3.5 Flash-Lite | 326.1 | 9.00 |
Token units and future rates need their own notes
Artificial Analysis standardizes performance token counts with o200k_base. Provider invoices use native token accounting, so identical text can consume different billable quantities. Dividing a quoted native-token price by a standardized streaming speed would hide that mismatch. The scenario fixes counts for planning rather than treating them as identical documents.
Google schedules doubled Flash input and output rates from 1 January 2027. The separate future row holds volume unchanged to show the budget exposure. It is an announced schedule, not a price already charged in October. Teams can keep current and scheduled budgets beside one another, with a source check before renewal.
Buy accepted work within a response budget
A useful local selection process starts with representative prompts and a clear acceptance rule. Record native billed tokens, time to useful text, time to completion and retries. Interactive requests deserve their own waiting-time threshold; background queues can prioritize cost. This is a proposed workflow, not a claim of measurements collected by this publication.
October’s sources identify candidates and expose assumptions, without providing a universal winner. A production owner can decide which failures need escalation, how much reasoning is allowed and when a delayed request should stop. Cost per accepted result then becomes a better operating question than a cheap token headline, especially when a published promotional rate has a known end date.
Questions and answers
Are the costs measured production invoices?
No. They are arithmetic for a fixed token workload using official standard paid rates checked on 1 October 2026. Actual token use, retries, tools and taxes can change a bill.
Does the fastest stream deliver the first answer fastest?
Not necessarily. Output speed starts after streaming begins. Reasoning and input processing can delay useful answer text; the published TTFT label is not a completed-task guarantee.
Why is there a January price row?
Google’s official schedule lists higher Gemini 3.8 Flash rates from 1 January 2027. The future cost uses the same assumed workload and is clearly separated from October prices.