Anthropic's Claude Fable 5.2 Undercuts OpenAI on Price, But Token Usage Tells a Different Story
Anthropic's new Claude Fable 5.2 launched with lower API rates than OpenAI's GPT-6 Astra, but real-world testing reveals the model consumes significantly more tokens, erasing the pricing advantage on actual development tasks.

Anthropic unveiled Claude Fable 5.2 on November 9 as a model tailored for intensive coding and technical applications. The company's benchmarks showed Fable 5.2 outperforming OpenAI's GPT-6 Astra on two coding measures, while also advertising lower standard API rates. However, a closer examination of token consumption during real-world tasks reveals that the lower per-token pricing does not necessarily translate to lower total costs.
On FrontierCode 2.0, Fable 5.2 achieved 49.1% compared to Astra's 42.8%. On FrontierSWE v2, the gap was 69.9% to 65.5%. Fable 5.2's 49.1% score came at the model's highest effort setting; at its default High setting on the Claude API, performance dropped to 45.3%. Benchmark results do not always reflect performance on production work.
The pricing structure appears favorable to Anthropic at first glance. Fable 5.2 charges $8 per million uncached input tokens and $40 per million output tokens under standard API rates. OpenAI set GPT-6 Astra's standard rates at $10 per million uncached input tokens and $50 per million output tokens when it launched in September. Both services bill for reasoning tokens as output and maintain separate pricing for cached input. OpenAI President Greg Brockman stated at Astra's launch that "The price per task is what matters." In practice, Anthropic's models have historically consumed more tokens than OpenAI's equivalents, a pattern that emerged in testing.
Testing methodology
Both models were evaluated on three identical tasks, each run once through the respective vendor's API using default configuration. A scoring script graded responses on a 50-point scale per test.
- Bug fix: Ten bugs were inserted into a 2,600-line billing codebase. A support ticket described ten customer complaints with incorrect guesses about root causes. The codebase contained two faulty tests, and one bug only manifested after another was resolved.
- Code review: Forty-five defects were planted across VAT ID validation for 27 countries, IBAN rules, European Central Bank payment calendars, and date formats for 50 locales in an EU invoicing package. Models had to demonstrate each defect with a failing input and provide the correct result. False positives reduced scores.
- Code writing: Models built helper functions from scratch, including VAT ID validation for 27 countries, IBAN checks for 34 countries, the payment calendar, and extended date formatting.
Results by task
Bug fix
Both Fable 5.2 and Astra fixed all 10 bugs and cleared all 50 hidden checks in 12 turns each. The convergence on turn count was unexpected. Token consumption diverged sharply: Fable 5.2 deployed more than 17,000 reasoning tokens versus Astra's 819. Astra completed the task in approximately two minutes at an estimated cost of 98 cents. Fable 5.2 required more than five minutes and cost an estimated $2.51.
- Fable 5.2: 12 turns, 696,863 input, 36,135 output (17,387 reasoning), $2.51, 326 seconds
- Astra: 12 turns, 231,854 input, 6,823 output (819 reasoning), $0.98, 122 seconds
Code review
This task produced the only meaningful score difference. Fable 5.2 scored 49 while Astra scored 45. Fable 5.2 identified 44 of 45 defects and generated no false positives across 71 total findings. Astra identified 41 defects and incorrectly validated one Spanish ID. Astra missed check-digit rule violations in Spain, Ireland, and Luxembourg. Both models failed to catch the same Spanish defect: two transposed letters within a 23-character lookup table.
Fable 5.2's superior detection came at substantial cost. The model spent more than six times what Astra did and consumed more than four times as much duration. For systems processing invoices or approvals where a single missed validation rule causes real invoices to be rejected, Fable 5.2's four additional points may justify the expense. For routine review work, Astra identified 91% of defects for 16% of the cost. Since each model ran once, the four-point gap could shift in subsequent runs.
- Fable 5.2: 26 turns, 2,207,564 input, 103,240 output (77,641 reasoning), $6.05, 993 seconds
- Astra: 12 turns, 192,018 input, 8,664 output (4,307 reasoning), $0.95, 227 seconds
Code writing
Both models achieved perfect scores of 50 out of 50. Each successfully wrote working validation logic for 27 countries' VAT IDs and 34 countries' IBANs without relying on external libraries. Neither encountered difficulties. Both cleared all checks on their first attempt.
Astra required 23 turns, nearly matching Fable 5.2's 24 turns. Despite similar turn counts, Fable 5.2 cost 2.3 times more and took 1.7 times longer for an identical score. The additional tokens consumed did not improve the outcome.
- Fable 5.2: 24 turns, 1,495,712 input, 87,542 output (47,458 reasoning), $4.83, 838 seconds
- Astra: 23 turns, 460,271 input, 24,007 output (9,328 reasoning), $2.06, 483 seconds
Overall assessment
Fable 5.2 prevailed on code review and tied Astra on the remaining two tasks. Across the two tied tasks, Fable 5.2's estimated API cost reached $7.34 versus Astra's $3.04—a 2.4-fold difference. Execution time showed a similar pattern, with Fable 5.2 requiring 1,164 seconds compared to Astra's 605 seconds.
For development work of this nature, starting with Astra appears prudent. The model matched Fable 5.2's performance on bug fixing and code writing at roughly 41% of the estimated cost and delivered results faster across all three tasks. Fable 5.2 warrants consideration for thorough code reviews, where it uncovered three additional defects in this evaluation. Both models missed defects.
Although Fable 5.2's standard token rates are lower on paper, the model consumed approximately five times as many tokens as Astra across the three test runs. Brockman's assertion about price per task—rather than price per token—proved accurate in practice.