unbiased ai

How much do tokens actually cost?

A token is about three quarters of an English word. So "$5 per million input tokens" means roughly 750,000 words in for $5. But nobody buys a million tokens of nothing: you buy finished tasks. This page translates both ways: paste your own text, then see what real tasks billed.

Rates are provider list, July 2026. Task bills are metered runs, not estimates. Anything with a dotted underline shows its arithmetic on hover or focus.

or match your spend in credits

Translate your text into dollars

Type or paste anything. We estimate tokens with the three-quarters-of-a-word rule and price it as input on every July 2026 rate card. Nothing you type leaves this page.

0 words 0 characters 0 tokens, estimate
Model $/Mtok in This text, one send Sent 1,000 times

Pareto is a blend priced in per-token credits, so it has no flat rate card; its metered bills are in the next section. Estimates price input only.

This is the quick estimate. The deep tool meters your exact workload: output tokens, batch sizes, caching, and monthly volume across nine models.

Open the LLM cost calculator

One real task, six real bills

The same dashboard build prompt, run to completion on each model on the same day, billed by each provider's own meter.

Model Rate card (in / out) The task's real bill Bill relative to the priciest
Fable 5$10 / $50$0.1462
Claude Opus 5$5 / $25$0.1158
Claude Opus 4.8$5 / $25$0.0724
Claude Sonnet 5$2 / $10 $3 / $15 from Sep 1$0.0410
Pareto (blend)per-token credits$0.0196
Claude Haiku 4.5$1 / $5$0.0162

Replication check. A second metered task, a small game build, came out about ten times apart too:

Fable 5$0.2269
Pareto (blend)$0.0220

This table only lists models we hold real bills for. Kimi K3 is queued for the same metered run QueuedRunningPublished and GPT-5.6 Sol, Terra, and Luna are rate cards only for now. The model card publishes seven same-harness benchmarks with both bills; per-task cost ran 6 to 38 cents on the Opus 5 dollar. Want your own workload priced? The full calculator takes any prompt.

Your month, at your volume

Drag the slider. Each bar is the metered task bill times your tasks per day times 30. Straight multiplication, no volume discounts modeled.

25 a day is 750 tasks a month

Claude Sonnet 5 at this volume: $30.75 a month today, $46.13 after September 1. Same tokens, both rates up 50%.

The three multipliers rate cards hide

5 to 6xwhat output tokens bill versus input on every July 2026 card here
+60%more tokens Claude Opus 5 spent than Opus 4.8 on the identical task
2.6xcheaper the same real agent session got with prompt caching on

Output is where the money goes

The left number on a rate card is the small one. Output bills at five to six times input on every card here, so long answers dominate real bills.

Fable 5 $10 in / $50 out
5x
Claude Opus 5 $5 in / $25 out
5x
Claude Opus 4.8 $5 in / $25 out
5x
Claude Sonnet 5 $2 in / $10 out until Aug 31
5x
GPT-5.6 Sol $5 in / $30 out
6x
Terra $2.50 in / $15 out
6x
Luna $1 in / $6 out
6x
Kimi K3 $3 in / $15 out
5x

Bars are drawn to one shared dollar scale, so you can compare across models as well as within one card.

Verbosity is a price you cannot see on the card

Same task, same rate card. Claude Opus 5 chose to spend about 60% more tokens than Opus 4.8, and its metered bill landed about 60% higher: $0.1158 against $0.0724. Every dot below is one token.

Claude Opus 4.8: 2,877 tokens
Claude Opus 5: 4,610 tokens, the orange 1,733 are the extra

Caching pays you back for re-reading

A real Claude Code session: five API calls, $0.59 total, and 78 times more tokens read than written. Prompt caching made that session 2.6x cheaper than the same calls uncached.

Cache off derived$1.54
Cache on billed$0.59

Kimi K3 publishes the cleanest cache line on any card here: $3.00 input, and cache hits bill at a tenth of that.

Price Kimi K3 input at its cache-hit rate: $3.00 per Mtok

The toggle also reprices Kimi K3 in the translator above. Cache economics are workload specific: the 2.6x is one real session, the $0.30 is a list price for hits, and your mix of hits and misses decides the rest.

What $10 of credits buys

Same metered dashboard task. $10 divided by each model's real bill, rounded down to whole runs.

68runs on Fable 5
86runs on Claude Opus 5
138runs on Claude Opus 4.8
243runs on Claude Sonnet 5
510runs on Pareto (blend)
617runs on Claude Haiku 4.5

Task counts are arithmetic on the July 2026 metered bills above, not a promise about your workload. Longer prompts buy fewer runs; the translator up top shows how fast that scales.

Every rate card on one table

Sortable. The last column translates each input dollar into words, using the three-quarters rule.

Fable 5$10.00$50.005x≈ 75,000 words
Claude Opus 5$5.00$25.005x≈ 150,000 words
Claude Opus 4.8$5.00$25.005x≈ 150,000 words
Claude Sonnet 5$2.00$3.00 from Sep 1$10.00$15.00 from Sep 15x≈ 375,000 words
GPT-5.6 Sol$5.00$30.006x≈ 150,000 words
Terra$2.50$15.006x≈ 300,000 words
Luna$1.00$6.006x≈ 750,000 words
Kimi K3$3.00$0.30 cache-hit$15.005x≈ 250,000 words

Provider list prices, July 2026. Kimi K3 rates are flat across context lengths; its API went live July 16, 2026. Pareto is a blend priced in per-token credits, so it has no flat card; its metered bills are above.

Ten months of price moves

Four dated moves since November, and a fifth that was cancelled. Select a marker for the details.

Why metered pricing is suddenly the whole conversation: weekly usage limits arrived on Claude subscriptions in summer 2025, with 5-hour session windows. On July 7, 2026 the frontier model left subscriptions entirely, moving to API usage credits at $10 in and $50 out. When capacity is metered, translating tokens to dollars stops being trivia.

Quick answers

How many words is a token?

About three quarters of an English word, as a working estimate. A million tokens is roughly 750,000 words. Exact counts vary by tokenizer and language, so meter anything that matters.

Why does output cost more than input?

On July 2026 provider rate cards, output tokens bill at five to six times input. Chatty models multiply the premium: Claude Opus 5 spent about 60% more tokens than Claude Opus 4.8 on an identical task at identical rates, and its metered bill was about 60% higher.

What does a real AI task cost in dollars?

A fixed dashboard build task, metered in July 2026, billed $0.1462 on Fable 5, $0.1158 on Claude Opus 5, $0.0724 on Claude Opus 4.8, $0.0410 on Claude Sonnet 5, $0.0196 on Pareto, and $0.0162 on Claude Haiku 4.5 per completed run.

Do token prices change often?

Four dated moves in ten months, and a fifth that was called off: Claude Opus dropped 67% in November 2025, Fable 5 moved to API usage credits on July 7, 2026, Kimi K3 launched at $3 in and $15 out on July 16, 2026, Claude Opus 5 launched at $5 and $25 on July 24, 2026, and Claude Sonnet 5's intro pricing ends September 1, 2026.

Four price moves in ten months. The fifth was cancelled.

When any provider changes rates, we re-meter the task and email the new table. That is the whole list.

One request in, one answer out: the blend picks per request, so you skip this choice entirely.

Get your credits matched