Dev.to AI 🤖 Ai 👁 0 📖 5 min read

GPT-6.2 Astra vs Claude Fable 5.2 at $8/$40: check the cache rules, not just the list price

For repository changes, evaluate Fable 5.2 first; for legal-document agents with human review, evaluate Astra first. Artificial Analysis's published max-effort results favor different models on those workloads. A routing

For repository changes, evaluate Fable 5.2 first; for legal-document agents with human review, evaluate Astra first. Artificial Analysis's published max-effort results favor different models on those workloads. A routing decision needs a task and success criteria before it needs a flagship ranking. Both models are available through one AIHubMix key.

Start with the task and evaluation configuration

FrontierCode-AA 2.0 gives Astra at max 45.38%, versus 48.76% for Fable, labelled max, default fallback. AA runs 132 private tasks in real repositories three times each, using Stirrup with a 500-turn cap and equal task weights. Fable's higher score makes it a useful first coding candidate, but max labels do not prove equal compute or predict every repository workflow.

On Harvey LAB-AA v1.1, Astra at max records 15.0% at $6.35 per task; Fable, labelled max, default fallback, records 7.2% at $39.84 per task. These are evaluated-task costs, not costs per successful task. AA uses 120 private legal tasks across 24 practice areas and a three-judge LLM panel. Its strict Hallucination-Gated All-Pass Rate averages the share of judges passing every rubric criterion and zeros a task for any material hallucination. Do not substitute the much higher criterion-pass rate.

The directly published evaluator results underlie OpenAI's Harvey summary; they are not a second independent reproduction of the same run. They justify an Astra-first legal evaluation, with human review for either model because the absolute pass rates remain low.

For interactive 3D, AA-3D's 55 offline three.js tasks use blind human pairwise votes. Astra at max has Elo 1318 (95% interval 1286–1349). Fable, labelled max, default fallback, has 1271 (1246–1297). Overlapping intervals and the narrow task scope make this a reason to evaluate Astra on your visual briefs, not decisive general coding superiority. The source is grouped below.

Official prices that match, and official prices that do not

Both official Standard short-context rate cards list $8 per million input tokens and $40 per million output tokens. The figures below come from OpenAI pricing and the Fable 5.2 overview. They are official prices, not AIHubMix billing.

Item GPT-6.2 Astra Claude Fable 5.2
Input $8 / MTok $8 / MTok
Output $40 / MTok $40 / MTok
Cache read $0.40 / MTok $0.20 / MTok
Short cache write $10 / MTok (30 min) $10 / MTok (5 min)
Long cache write $20 / MTok (12 h) $16 / MTok (1 h)

Fable's cache reads are half Astra's price, but the short write lasts 5 minutes versus 30 minutes. Longer writes last one hour versus 12 hours. Waiting for a user or tool can change which cache window fits a run.

Astra's short-context rates apply at up to 272,000 input tokens. Above that, input and cache rates double and output is 1.5× for the full request, not just the excess. Context windows are 1,050,000 tokens for Astra and 1,000,000 for Fable; both cap output at 128,000. The base-price tie does not cover every context length or service tier.

A synthetic no-cache example isolates the base rate: ten requests totaling 818,000 ordinary input tokens and 30,000 billed output tokens including reasoning, each request below 272,000 input tokens, with no cache writes or tool fees. Assuming identical billed counts, both cost $7.744: 0.818 × $8 + 0.030 × $40. Different tokenizers need not produce identical counts on the same text.

Set effort explicitly before comparing runs

Fable's API defaults to high with adaptive thinking always on. Its announcement specifies medium in Claude apps and Claude Code. OpenAI's GPT-6.2 guide recommends explicitly starting new integrations at medium; it does not establish the default when the field is omitted. Compare quality, latency, and completed-task cost across budgets; identical effort labels do not match compute.

Keep the Agents' Last Exam caveat with its result

OpenAI reports Astra 5.5 percentage points ahead of Fable at 5% of the cost on Agents' Last Exam. Fable's leaderboard result excludes nine tasks that used a fallback model. The announcement does not establish identical task sets or matched compute. The vendor-reported ratio is not a general agent-cost forecast, nor can it be derived from the tied base rates.

That exclusion belongs to Agents' Last Exam. Harvey's Fable result has its own max/default-fallback label; do not apply the nine-task rule to it.

Compare the available models through AIHubMix

GPT-6.2 Astra, Sol, and Claude Fable 5.2 are available through one AIHubMix API key. Public sources checked on 18 November 2026 at approximately 15:04–15:06 UTC (23:04–23:06 Beijing time) show providers, pricing, and Try this model on the AIHubMix Astra page and Sol page. Both exact GPT-6.2 IDs appear in the parameter index, and their llms documents link the resolved schemas.

AIHubMix documents Chat Completions, Responses, and Anthropic Messages for these models. GPT-6.2's Messages route is gateway translation, not an OpenAI-native endpoint; OpenAI's GPT-6.2 tool-calling API requires Responses. No hands-on gateway test was performed for this article. A shared key or protocol does not guarantee the same backend or tool billing.

Fix prompts and success criteria, record effort and fallback behavior, and compare completed-task cost on the route you use. Start with representative repository tasks for Fable, reviewed legal tasks for Astra, and your own briefs for visual work. Direct vendor access is a separate channel.

Keep reading: the GPT-6.2 series

Sources

  1. Artificial Analysis: FrontierCode-AA 2.0, Harvey LAB-AA v1.1, and AA-3D — evaluator results and methods; snapshots accessed 18 November 2026, approximately 02:30–02:40 UTC. Max variants; Fable labelled default fallback.
  2. OpenAI: Introducing GPT-6.2 Astra and Sol — vendor-reported Agents' Last Exam comparison and exclusion; Harvey summary refers to the evaluator results above.
  3. OpenAI: Pricing, Astra model, and Using GPT-6.2 — rates, limits, and integration advice.
  4. Anthropic: Fable 5.2 overview and Fable and Mythos 5.2 announcement — specifications, prices, and API versus app/Claude Code defaults.
  5. AIHubMix: Astra, Sol, Fable, Astra llms document, Sol llms document, and schema index — GPT-6.2 model pages, llms documents, catalog entries, and resolved GPT-6.2 schemas checked 18 November 2026 at approximately 15:04–15:06 UTC (23:04–23:06 Beijing time); Fable model page and llms document checked at approximately 15:10 UTC (23:10 Beijing time) the same day. Availability and three gateway protocols are documented. No inference request was sent.
📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.