Pricing methodology & sources

1. How prices get on this site

• Manual verification. A human opens the provider's official pricing page, copies the figures, and stamps the date you see above. There is no live price feed — we deliberately don't claim realtime sync.
• Weekly re-checks. An automated job re-opens the official pages every week and flags any drift; the stamp above moves only after a human confirms the new numbers.
• One data layer. Every calculator on this site reads from a single pricing file — a price can't be $2.00 on one page and $2.20 on another.
• No invented numbers. If a provider hasn't published a rate (or we haven't verified it), it doesn't appear on this site. Period.

2. GPT-6 prices we use (short context)

Source: OpenAI official pricing. Per 1M tokens, USD. These are the short-context Standard rates — the ones our calculators use.

3. Long-context rates exist — and we don't hide them

OpenAI's pricing page shows a second column: long-context rates that apply to very large prompts. They are higher:

Which one does the calculator use? Short-context rates — they cover the overwhelming majority of API workloads. If your prompts routinely exceed ~200K tokens, treat our estimate as a lower bound and check the long-context column on the official page.

4. Service tiers: what the 0.5× means

The official pricing page lets you switch between service tiers. Our calculator models three of them:

• Standard (1.0×) — full price, full speed. The rates in the tables above.
• Flex (0.5×) — half price, lower priority; your requests queue behind Standard traffic.
• Batch (0.5×) — half price, up to 24-hour turnaround for non-urgent bulk work. The 50% Batch discount is stated on the official pricing page.

Deliberately excluded: Fast mode (priority throughput) has no public price multiplier, so we don't guess one. FedRAMP and regional-processing endpoints carry a 10% surcharge — also excluded; our numbers assume standard global endpoints.

5. Exactly what the calculator computes

For one request, with c = cached-input share (0–100%):

Key point: cached input replaces the standard input rate for the cached portion — the two are never stacked. The tier multiplier (1.0× / 0.5×) applies once, to the total.

6. Other providers: Claude & Gemini

Our comparison pages also show verified Anthropic and Google prices — with one honest caveat: caching works differently per provider, so cross-provider comparisons are close estimates, not exact invoices.

• Anthropic: the "cached input" figure is the cache-hit rate. Cache writes cost 1.25–2× the base input price and are not modeled in our calculators — heavy cache churn will cost more than shown.
• Google: figures are short-context rates (prompt ≤ 200K tokens). Gemini 2.5 Pro charges more for longer prompts; Flash / Flash-Lite text rates shown (audio input costs more).
• Cross-provider "cached %" comparisons assume the same reuse pattern — in practice each provider's caching rules differ. Treat these as directional.

7. What we won't do

• No "live" or "realtime" price claims — the stamp at the top is the truth.
• No invented rates for tiers, regions, or models we haven't verified.
• No affiliation claims — this is an independent tool, not affiliated with or endorsed by OpenAI, Anthropic, or Google.
• No quality judgments dressed as math — comparisons here are strictly cost arithmetic.

8. Changelog

• September 26, 2026 — Initial publication. GPT-6 short/long-context rates, tier multipliers, Claude (Opus 5.5 / Sonnet 5 / Haiku 4.5) and Gemini (2.5 Pro / Flash / Flash-Lite) rates verified against official pricing pages.