Same brief. Two separate experiments.
The starting prompts match byte for byte. Each model was asked for 100 standalone HTML websites, inline CSS and JavaScript, no frameworks or external image assets, and a distinct design for every page. Google Fonts were allowed.
Each model chose its own subjects and execution strategy. Astra used ten builders and completed its turn in 22 minutes, 34 seconds. Fable used 50 builder sessions across the original attempt and retries, hit usage limits, and was resumed twice. Its published gallery contains 72 pages.
The screenshots above show two different subjects. They are examples to open and explore, not a matched design test. This single experiment does not establish which model is generally faster, cheaper or better at design.
Read the exact shared prompt · Read Astra’s manifest
Shared prompt SHA-256: b82f0a1d12eaa503f208993930a8d8a78a4220dc73b89b4989346f5b59688caa
A correction to the earlier Fable receipt.
The previously published $306.77 / 58.12M tokens estimate summed every assistant content record in the builder logs. Multiple records repeated the same request’s usage. That calculation also omitted the coordinator and priced all cache writes as five-minute writes.
The corrected estimate is $227.40 / 34.35M tokens. It deduplicates requests, includes the coordinator and all 50 builders, and uses the actual five-minute and one-hour write buckets. The earlier social image is superseded by this audit.
Claude Fable 5.1 · Standard API prices in USD per million tokens| Token category | Tokens | Rate / 1M | Estimate |
|---|
| Uncached input | 172,602 | $10.00 | $1.73 |
| Cached input | 26,221,467 | $0.25 | $6.56 |
| 5-minute cache writes | 4,147,686 | $12.50 | $51.85 |
| 1-hour cache writes | 779,237 | $20.00 | $15.58 |
| Output, including reasoning | 3,033,775 | $50.00 | $151.69 |
| Total | 34,354,767 | | $227.40 |
Original coordinator session and all 50 builder logs, including retries. Count each request ID and message ID once; retain greatest reported cumulative output for that request. Use actual 5-minute/1-hour cache-write buckets. Ignore zero-usage synthetic errors. Exclude the later publication fork. Run hit usage limits and was resumed twice.
There are 394 distinct billable requests in the recovered original build logs. Output includes reasoning; cached input and cache writes are separate input categories.
Source: Anthropic pricing, checked September 6, 2026. Prices assume standard global API rates.
What the dollar figures mean.
Standard API-rate estimates from original build logs; not invoices or subscription charges. Deployment, marketing assets, and subsequent publishing work excluded. They are estimates of the recorded usage at public standard API prices. They are not provider invoices, measurements of subscription credit deductions, or evidence of a general efficiency advantage.
Per-site amounts divide each run’s estimate by its published page count; they include the cost of unsuccessful work and retries in that run. Line items are rounded separately, while totals use unrounded values.
Download the token counts, rates and per-builder Astra totals.
Try the black-hole lens
Try the Gramercy typewriter