Fable 5.2 vs GPT-6 Astra vs Gemini 4 Carbon: Independent Scores

Who leads once outside evaluators run the tests?

By 11 November, Artificial Analysis, Vals AI and ARC Prize had each published Fable 5.2 results. Every row below is independent and comes from the evaluator's live page. Effort settings differ because each evaluator tests its own mix, so the setting is printed beside the score. Rupee figures use the RBI reference rate of Rs 93.8579 per US dollar on 9 November 2026, rounded.

How much of that lead is real?

Less than the rank numbers suggest. Fable 5.2's Vals Index lead over Carbon is 0.86 points, and the two published margins overlap, so treat that pair as tied. Carbon reaches the tie for 41% less per test ($10.51 against $17.94). Vals's Carbon page lists token costs of $4 and $20, but it does not say which rate produced that figure. The Artificial Analysis lead is wider and it is expensive: Fable's top score comes only at max effort, the costliest setting of all 73 entries on that measure. At high effort the gap to Carbon is 2 points for $1.32 more per task. Artificial Analysis's Carbon page lists $4 and $20 per million tokens, the rate that applies from 1 December, which suggests its $7.76 already reflects the higher price. The page does not say so outright. GPT-6 Astra is the weakest of the three on most of these rows. Its case rests on low cost per task, the provider-adapter ARC result, where it comes first, and science work on Vals AI, where it ranks first on the Vals Science Index (65.09% against Fable 5.2's 63.74%) and on Terminal-Bench-Science.

What do the vendors' own tables say?

Anthropic's launch table is the fullest vendor comparison, so it needs careful labelling. Some rows were run by third parties and only published by Anthropic; others are Anthropic's own runs. Its footnote says scores show each model's highest-available effort unless noted.

• •Google prints no scores for these in its 10 November post. It says Carbon "ranks #1 on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench at more than 2X lower cost per task compared to the next best external model", without naming that model.
• •The fairest vendor claim for Fable 5.2 is on cost: Cognition's FrontierCode 2.0 run has it at 45.3% for $6.72 per task at high effort, against Astra's best of 42.8% for $6.68.

Where does the same benchmark give a different answer?

Harnesses, effort settings and task subsets move scores by more than the gaps vendors advertise. When a vendor table and an independent board disagree, plan on the independent board.

• •SWE-bench Multimodal: Anthropic's system card reports Fable 5.2 at 74.1% and Opus 5.5 at 61.4%. Vals AI measures Fable 5.2 at 62.29% and Opus 5.5 at 70.21%. The order flips.
• •Terminal-Bench-Science: 74.6% in Anthropic's table (version 0.1) and 55.71% on Vals AI at high effort. Vals scores the same 70 tasks from v0.1.0 with one general-purpose harness and says its figures tend to sit below the official board's. The order against Astra flips too: Fable leads 74.6% to 68.1% in Anthropic's table, and Astra leads 62.86% to 55.71% on Vals.
• •FrontierCode 2.0 at max: Fable 5.2 scored 45.7%, but 22 of 660 attempts timed out and were graded anyway. Counted as zero, the score would be 44.4%, below Opus 5.5's 45.5%.
• •ARC-AGI-3 at max: 65.66% on the standard harness against 81.2% at xhigh. ARC Prize notes that 28 of 55 games at max ran into the 12-hour limit.

Where does Fable 5.2 lose?

Its losses cluster in document-heavy professional tasks. If your workload looks like contracts, returns or financial analysis, this section matters more than the headline index. Nothing on the Vals model pages suggests these tasks cover Indian GST or Indian case law, so test on your own documents before relying on them.

• •Vals Legal Index: Fable 5.2 69.17%, rank 20 of 37. Carbon 75.93%, rank 3. Astra 68.39%, rank 25.
• •Tax Agent Bench: Fable 5.2 32.09%, rank 27 of 84. Carbon 51.78%, rank 1. Astra 20.72%, rank 54.
• •Legal Research Bench: Fable 5.2 46.15%, rank 19 of 93. Carbon 55.77%, rank 1.
• •Vals Finance Index: Fable 5.2 63.58%, rank 6 of 39. Carbon 67.05%, rank 1. Fable does win the narrower Finance Agent v2.1 at 69.47%.
• •GDPval-AA v2.1: Opus 5.5 leads at 1866 Elo, Kimi K3.1 Pro follows at 1861, and Fable 5.2 is sixth at 1831.
• •AutomationBench: Carbon's 53.9% beats Fable's 51.4%, and both vendors agree on that order.
• •Speed and spend at max: 44.4 minutes per Artificial Analysis task against a 12.7-minute median, and 275M output tokens against a 72.8M median.

What has not been published yet?

Fable 5.2 is two days old, so these gaps are normal. They are also a reason to avoid large migrations on this week's numbers.

• •METR: no time-horizon figure for Fable 5.2. Its time-horizons page, last updated 8 September 2026, says it is no longer actively updated.
• •SWE-bench Pro: Scale's public board lists no Claude 5.x, GPT-6 or Gemini 4 model. Its leader is Muse Spark 1.1 at 61.50. Anthropic's 93.0% for Fable 5.2 is vendor-reported, and we would not compare it with that board.
• •Arena text ranking: on 11 November the main Arena page showed Fable 5.2 only as first in WebDev at xHigh, with no score, and not in the Agent top ten. We could not read a text ranking that includes it.
• •Any independent test on Indian-language, GST or local legal work for these three models. We found none.

Who actually leads WANDR?

Nobody can say yet, and you should discount both claims. Anthropic prints 82.1% for Fable 5.2 on a WANDR "relaxed score" with no Gemini figure. Google says Carbon ranks first. Anthropic's system card explains why the two cannot be lined up: it ran a modified version with a frozen search index and its own judge model, and states that its results "are not directly comparable to reported numbers for competitor models". Perplexity's own launch write-up scores soft and hard F1 on the live web and does not list either model. The launches were a day apart, so each vendor may also be quoting a board from before the other shipped.

What do the three models cost, and what changes on 1 December?

List prices below are each vendor's standard rates per million tokens. Carbon's introductory period ends on 30 November, and from 1 December it matches Claude Opus 5.5 exactly. Astra also doubles input and adds 50% to output once a prompt passes 272K tokens; Google's page lists no long-context tier for Carbon, and Anthropic charges none for Fable 5.2.

Why doesn't the price list predict your bill?

The arithmetic above assumes each model uses the same tokens, and they do not. On Artificial Analysis, Astra lists 25% above Fable 5.2 yet costs $6.24 per task at max against Fable's $26.54, because Fable used 275M output tokens across the index. Carbon at high used 205M, so even at half of Fable's listed rate it costs $7.76 per task. Budget from cost per finished task at the effort you will actually run. For the Claude line-up, see /blog/claude-fable-5-2-vs-opus-5-5-cost-per-task-rupees.

Where do the cheaper models fit?

The three flagships are not your only choice for high-volume work. GPT-6.1 Sol at max scores 55 on Artificial Analysis for $1.27 per task (about Rs 119), which makes it the OpenAI model to test before paying for Astra. Muse Ember, launched on 9 November at $0.10 and $0.50 per million tokens, scores 36 for $0.14 per task. Kimi K3.1 Pro lists at 46 for $3.68 on the live board, down from the 55 that Simon Willison reported on 23 October; either the index was re-based or the listed variant differs. Its weights are on Hugging Face under a custom licence.

Which model should you pick for which job?

This is our reading of the published numbers, and Tech Arion has run no benchmark of its own, so treat it as a shortlist. Anthropic itself tells most coding and knowledge users to start with Claude Opus 5.5, its "strongest daily driver", before moving up to Fable.

• •Long coding and agent runs where finishing matters: Fable 5.2 at high. It scores 62 for $9.08 on Artificial Analysis, ahead of Opus 5.5 at max (59 for $10.04). Move to xhigh only where you have measured a gain.
• •Legal, tax and finance documents: Carbon first, and re-cost it after 1 December.
• •An existing OpenAI stack: test GPT-6.1 Sol before Astra.
• •Fable 5.2 access has one catch for regulated buyers: it requires 30-day retention unless Anthropic expressly authorises otherwise. See /blog/fable-5-2-30-day-retention-inference-hooks-dpdp-claude-code-security.

How should you run your own comparison in a week?

A small bake-off on your real work will tell you more than any board above.

What mistakes should you avoid when reading these boards?

Each of these comes from reading a single number without its conditions.

Frequently asked questions about Fable 5.2, GPT-6 Astra and Gemini 4 Carbon

Short answers as of 11 November 2026.

How can Tech Arion help you choose?

Tech Arion is a Hyderabad software and AI consultancy that builds with Claude Code and advises teams on model selection. We can set up a bake-off on your own tasks, with cost per finished task in rupees and a routing plan you can defend to finance.