Current rate categories
| Model | Uncached input | Cache read | Cache write | Billed output |
|---|---|---|---|---|
| Luna | $0.20 | $0.02 | $0.25 | $1.20 |
| Terra | $2.00 | $0.20 | $2.50 | $12.00 |
| Fable 5.1 | $10.00 | $0.25 | $12.50 / $20.00 | $50.00 |
| Astra | $10.00 | $1.00 | $12.50 | $50.00 |
USD per million tokens, checked 12 September 2026. Fable cache writes show five-minute / one-hour lifetimes. OpenAI rows are short-context standard service.
The reproducible formula
cost_USD = (U × input_rate + R × cache_read_rate
W × cache_write_rate + O × output_rate) / 1,000,000U, R, W, and O must be disjoint billed categories. Output includes billed reasoning or thinking, not only visible text. Select service, region, context band, and cache lifetime before applying rates.
Three workflow baskets
| Illustrative workload | Luna | Terra | Fable 5.1 | Astra |
|---|---|---|---|---|
| 12K input + 2K output | $0.0048 | $0.0480 | $0.2200 | $0.2200 |
| 80K input + 10K output | $0.0280 | $0.2800 | $1.3000 | $1.3000 |
| Five calls; 250K aggregate input + 20K output, uncached | $0.0740 | $0.7400 | $3.5000 | $3.5000 |
The five-call row means five separate 50K-input/4K-output calls, not one 250K prompt. Each total represents complete billable input including instructions, schemas, retrieved files, and retained history.
A repeated-prefix cache example
Suppose five calls each contain a stable 40K-token prefix, 10K changing input, and 4K billed output. The prefix is written once and read four times: U = 50K, W = 40K, R = 160K, O = 20K.
| One write + four valid hits | Luna | Terra | Fable 5.1 | Astra |
|---|---|---|---|---|
| Token-only cost | $0.0472 | $0.4720 | $2.0400 | $2.1600 |
Fable uses its five-minute write rate in this basket. Valid reuse depends on an unchanged prefix, a cacheable boundary, and timing. A real debugging conversation may not preserve any of those conditions.
Long context changes the relationship
| One request: 300K input + 10K output | Cost |
|---|---|
| Luna | $0.1380 |
| Terra | $1.3800 |
| Fable 5.1 | $3.5000 |
| Astra | $6.7500 |
OpenAI’s greater-than-272K rule applies higher rates to the whole request. Fable’s advertised 1M window uses its standard rates. This still does not measure whether either model uses long context effectively.
What the totals exclude
Tool calls, hosted execution, agent runtime, subscriptions, taxes, currency conversion, negotiated prices, unmeasured retries, and human work are outside these examples. There is no cost-per-success figure because no compatible accepted-outcome dataset exists for this slate.
Use the buyer guide to choose a test and the comparison to decide which premium to investigate.
Reader questions
Frequently asked questions
Why can an agent cost more than input plus the visible answer?
Instructions, schemas, retrieved files, retained history, hidden reasoning, tools, and retries can all add billable usage or fees. Use provider usage fields rather than estimating only visible text.
Does caching always make repeated coding cheaper?
No. The prefix must remain cacheable and be reused before expiry. Edits, compaction, changed tool schemas, and timing can change the hit rate and total bill.
Why is there no cost per successful task?
The research has no same-protocol dataset pairing every attempt’s complete cost with accepted repository outcomes and review effort. Combining an illustrative bill with a benchmark score would create a misleading metric.
Primary reading
Sources and update conditions
Rates, availability, model identity, and provider behavior can change. Recheck the linked sources before making a consequential purchase or sending repository data.