Facts dated November 21, 2026 — the long-context rule and the provider list are now stated

Fable 5.2 is verified against Anthropic's pricing docs and model card. GPT-6 Astra's column is the site registry row, re-read on OpenAI's model page and pricing page on 2026-11-21: output, cache, batch and context figures that earlier passes could not confirm are now printed there and match the table, the 272K long-context rule is spelled out, and the provider list was corrected to what the model page actually says (Amazon Bedrock plus a Google Cloud private preview). OpenAI also documented GPT-6.2 Astra and GPT-6.2 Sol on November 17 — those are different models with their own rate card, covered on a separate page rather than folded into this comparison.

GPT-6 Astra vs Claude Fable 5.1 (September 2026)

Re-verified November 21, 2026

Claude Fable 5.2 vs GPT-6 Astra

Anthropic's Fable-tier flagship against OpenAI's reasoning model for complex coding and professional work. The prices are published; the benchmark numbers are not — here is what each vendor has actually confirmed.

The short answer

On published list prices Fable 5.2 is the cheaper of the two: $8 input / $40 output per million tokens against GPT-6 Astra's $10/$50. Fable 5.2's specs are confirmed against Anthropic's docs (1M context, 128K output, 30-day mandatory retention). Astra's row was re-read on OpenAI's model and pricing pages on November 21, 2026 and every figure in it is published there: 1,050,000-token context, 128,000 max output, $1 cached input, $12.50 cache write, $5/$25 Batch, and 2x input / 1.5x output above 272K input tokens. Neither vendor publishes a benchmark table that compares these two models, so neither does this page. Choose on price, context, channels and data handling, then settle it with your own evals.

Specs and pricing side by side

API list prices per million tokens. Fable 5.2 from the model registry (Anthropic's pricing page, verified 2026-11-11); GPT-6 Astra from the registry's row, re-read on OpenAI's model and pricing pages on 2026-11-21.

SpecClaude Fable 5.2GPT-6 Astra
API model IDclaude-fable-5-2gpt-6-astra
Released2026-11-092026-09-03
Input / MTok$8.00$10.00
Output / MTok$40.00$50.00
Cache read / MTok$0.20$1.00
Cache write / MTok — 5 min on Anthropic, 30 min on OpenAI$10.00$12.50
Batch API (input / output)$4.00 / $20.00$5.00 / $25.00
Context window1M1.05M
Max output128K128K
Long-context billingFlat — no long-context surcharge>272K input bills 2x input/cache and 1.5x output for the whole request
Reasoning effortFive levels; default highlow / medium / high / xhigh / max
Where it is soldClaude API, AWS, Google Cloud, Microsoft Foundry, BedrockOpenAI API, Amazon Bedrock, Google Cloud (private preview)
Data retention30 days, mandatory — a Covered ModelZero Data Retention available to approved organizations

Fable 5.2 re-read from Anthropic's pricing documentation and model card on 2026-11-11; GPT-6 Astra from the registry row last verified 2026-11-21. OpenAI's model page and pricing page, re-read on 2026-11-21, publish every figure in Astra's column: $10 input, $50 output, $1.00 cached input, $12.50 cache write for 30 minutes ($25 for 12 hours), $5/$25 Batch, 1,050,000-token context, 128,000 max output, and 2x input / 1.5x output for the whole request above 272K input tokens. GPT-6 Astra remains a current model — OpenAI's deprecations page retires nothing in the GPT-6 line.

OpenAI documented two successors on November 17, 2026: GPT-6.2 Astra (gpt-6.2-astra, $8/$40) and GPT-6.2 Sol (gpt-6.2-sol, $2/$10), both with a 1.05M context. GPT-6.2 Astra is not GPT-6 Astra — different model ID, different rate card, different knowledge cutoff — so a figure from one does not describe the other. GPT-6.2 Astra and Sol (documented Nov 17)

Benchmarks: what is actually published

Neither vendor has a number you can set against the other on a shared harness, so this page prints none.

  • OpenAI publishes claims, not a table

    OpenAI's announcement describes GPT-6 Astra as leading on cost-efficiency across Terminal-Bench 4.0 and the Artificial Analysis Intelligence Index, but the figures sit in charts the announcement does not quote. No SWE-bench Verified or Terminal-Bench score for Astra appears on OpenAI's model page, re-read 2026-11-21.

  • What this page will not do

    It will not put a score in a cell. A table where one column came from a vendor chart and the other from a third-party leaderboard is not a comparison — it is two different measurements of two models, presented as one.

  • Fable 5.2's numbers live on their own page

    Anthropic does publish a twelve-row launch table for Fable 5.2 — FrontierCode 2.0, FrontierSWE v2, AutomationBench, Terminal-Bench-Science, WANDR, OSWorld, RoboDojo, AA-Briefcase, the Remote Labor Index, Humanity's Last Exam and Chartography. It compares Fable 5.2 with Opus 5.5, not with Astra, so this site reproduces it, with sources, on the Fable 5.2 benchmark page instead of re-quoting it here.

  • Run both on your own tasks

    Both models take a system prompt, tools and images, and both expose a reasoning-effort setting, so the same harness runs against both with a base-URL and parameter swap. Run 50–100 of your own tasks at each model's default effort and compare cost per passing task — that number is measurable in an afternoon, and it outranks any table.

Sources: OpenAI's GPT-6 Astra model page and its pricing page (re-read 2026-11-19) for the rates and the long-context rule; Anthropic's Fable 5.2 launch post (2026-11-09) for Fable's own table. No third-party or aggregator figures are used — Artificial Analysis, OpenRouter and LiteLLM numbers measure different harnesses and would sit beside vendor numbers as if the two were comparable.

Fable 5.2 benchmarks: the launch table

When to use which

Route on published price, context, channels and data handling — the four things both vendors document.

  • You are building long-running coding agents

    Fable 5.2 is the model Anthropic positions for multi-day, open-ended technical work, with a confirmed 1M context and 128K output. Astra is the model OpenAI calls “a reasoning model for complex coding and professional work”, and it exposes five effort levels including xhigh and max, a wider deliberation range than Fable's default high. Without a shared benchmark, settle it on your own agent traces.

  • Your prompts run past 100K tokens

    Fable 5.2 serves the full 1M window at one flat rate, with no long-context surcharge. OpenAI publishes short- and long-context tiers, and Astra's threshold is confirmed at 272K input tokens: above it the whole request bills 2x input and cache rates and 1.5x output. A 300K-token prompt on Astra therefore costs roughly double the short-context input rate — price that tier before assuming Astra is cheaper.

  • You need a specific cloud

    Fable 5.2 is sold on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry. GPT-6 Astra's model page (re-read 2026-11-21) lists two cloud providers besides the OpenAI API: Amazon Bedrock as openai.gpt-6-astra and Google Cloud in private preview. If your data cannot leave a particular cloud, check that provider's own availability page — the model page only names the two.

  • You need zero data retention

    Fable 5.2 is a Covered Model: 30-day retention is mandatory and zero data retention is not available unless Anthropic expressly authorises it, with Enterprise Frontier Safeguards as the documented exception path. OpenAI offers Zero Data Retention to approved organisations on eligible endpoints and features; its data-controls page treats ZDR as an organisation- or project-level setting, so confirm eligibility for your own account rather than reading a blank cell as a guarantee.

  • Price is the deciding factor

    At $8/$40 against $10/$50, Fable 5.2 is the cheaper item on the published rate card for most workloads — the arithmetic is below. Astra's $50 output rate and its 272K long-context tier can still reverse that on output-heavy or long-prompt traffic, so re-price with your own token mix.

The routing descriptions are the vendors' own positioning; the price and context figures are the registry's, dated in the table above. Nothing here is an independent performance measurement.

What the published prices mean per task

List-price arithmetic on a representative coding workload — not a measured cost per task, and no allowance for retries.

Worked example — 100 tasks, 10K input tokens each with a 50% cache hit, 8K output tokens each, at list rates: Fable 5.2 ≈ $0.36 per task (≈ $36.10 for 100), GPT-6 Astra ≈ $0.46 per task (≈ $45.50 for 100) — about 26% more. Both sides now use rates published on the vendors' own pages, so the gap is arithmetic, not an estimate.

Priced at API list rates from the registry, global routing, no data-residency multiplier. Both models bill cached input at a steep discount, and a real agentic trace spends far more input than output, so the cache rows matter as much as the headline rates. The cost calculator has the exact math for your volumes.

Open the cost calculator

Sources and verification dates

Fable 5.2: Anthropic's pricing documentation (https://platform.claude.com/docs/en/about-claude/pricing) and model overview (https://platform.claude.com/docs/en/models/fable-5-2/overview), verified 2026-11-11 and re-checked 2026-11-21. GPT-6 Astra: OpenAI's model page (https://developers.openai.com/api/docs/models/gpt-6-astra), pricing page (https://developers.openai.com/api/docs/pricing) and deprecations page, re-read 2026-11-21. Where the two disagree, the vendor's own page wins.

Price your own token mix

Astra's long-context tier against Fable's flat 1M window changes which model is cheaper, depending on your prompt sizes. Run your real numbers before you commit.

Frequently asked questions