Anthropic unveiled Claude Fable 5.2 on November 9, positioning it as a solution for complex coding and technical tasks. The model outperformed OpenAI's GPT-6 Astra on two coding benchmarks at launch, scoring 49.1% versus 42.8% on FrontierCode 2.0 and 69.9% versus 65.5% on FrontierSWE v2. Fable 5.2 also carries lower standard API pricing than Astra.

However, benchmark performance does not always translate to cost efficiency in production. Fable 5.2 charges $8 per million uncached input tokens and $40 per million output tokens at standard rates, compared to Astra's $10 and $50 respectively, launched by OpenAI in September. Yet token consumption tells a different story. OpenAI President Greg Brockman stated at Astra's launch that "The price per task is what matters," and real-world testing suggests Anthropic's model consumes substantially more tokens to complete the same work.

Testing methodology

Three tasks were run through each vendor's API at default settings to compare performance and cost. Each test was scored out of 50 points using automated grading.

  • Bug fix: Locate and repair 10 intentional bugs in a 2,600-line billing codebase using a support ticket with incorrect guesses about causes, two broken tests in the codebase itself, and one bug that only surfaces after another is fixed
  • Code review: Identify 45 planted defects across VAT ID validation for 27 countries, IBAN rules, European Central Bank payment calendars, and date formats for 50 locales in an EU invoicing system, with false positives penalizing scores
  • Code writing: Generate working implementations from scratch for VAT ID validation across 27 countries, IBAN checks for 34 countries, payment calendar logic, and localized date formatting

Results by task

Bug fix

Both models fixed all 10 bugs and passed all 50 hidden checks in 12 turns each. The cost difference was substantial. Fable 5.2 consumed more than 17,000 reasoning tokens, taking over five minutes at an estimated cost of $2.51. Astra used only 819 reasoning tokens, completing the task in about two minutes for an estimated 98 cents.

  • Fable 5.2: 12 turns, 696,863 input tokens, 36,135 output tokens (17,387 reasoning), $2.51, 326 seconds
  • Astra: 12 turns, 231,854 input tokens, 6,823 output tokens (819 reasoning), $0.98, 122 seconds

Code review

This was the only test where scores diverged. Fable 5.2 scored 49 out of 50, proving 44 of 45 defects with no false positives across 71 findings. Astra scored 45, proving 41 defects but incorrectly classifying one Spanish ID as valid and missing defects in check-digit rules for Spain, Ireland, and Luxembourg. Both models failed to catch the same Spanish defect involving two swapped letters in a 23-letter lookup table.

Fable 5.2's superior accuracy came at significant cost. It spent more than six times what Astra did and required more than four times longer. For applications where validation errors have real consequences—such as invoice processing—Fable 5.2's four additional correct findings may justify the expense. For routine reviews, Astra identified 91% of defects at 16% of the cost. Each model ran only once, so the four-point gap could vary on subsequent runs.

  • Fable 5.2: 26 turns, 2,207,564 input tokens, 103,240 output tokens (77,641 reasoning), $6.05, 993 seconds
  • Astra: 12 turns, 192,018 input tokens, 8,664 output tokens (4,307 reasoning), $0.95, 227 seconds

Code writing

Both models achieved perfect scores of 50 out of 50, successfully writing working validation logic for 27 countries' VAT IDs and 34 countries' IBANs without external libraries. Both completed the task in a single run. Fable 5.2 required 24 turns while Astra needed 23, yet Fable 5.2 still cost 2.3 times more and took 1.7 times longer despite matching performance.

  • Fable 5.2: 24 turns, 1,495,712 input tokens, 87,542 output tokens (47,458 reasoning), $4.83, 838 seconds
  • Astra: 23 turns, 460,271 input tokens, 24,007 output tokens (9,328 reasoning), $2.06, 483 seconds

Overall assessment

Fable 5.2 outperformed on code review alone while matching Astra on bug fixing and code writing. Across the two tied tasks, Fable 5.2's estimated API cost was 2.4 times higher ($7.34 versus $3.04), and execution time was 1.9 times longer (1,164 versus 605 seconds).

For these types of coding tasks, Astra represents the more economical choice. It matched Fable 5.2's performance on bug fixing and code writing at approximately 41% of the estimated cost while completing all three tasks faster. Fable 5.2 warrants consideration only for careful code review work where it identified three additional defects in testing. Neither model achieved perfect defect detection.

Despite Fable 5.2's lower standard token rates, it consumed roughly five times as many tokens as Astra across the three test runs. This validates Brockman's observation that total cost per task, rather than per-token pricing, determines real-world economics.