Claude Fable 5.2: more capable agents, lower prices and practical deployment

Claude Fable 5.2 targets work that needs more than a good first response: long software migrations, difficult incident investigations and contradictions across business documents. It adds capability while lowering base prices from Fable 5.1. The business question is whether that translates into completed work with fewer corrections.
Anthropic released the model on November 9, 2026. This article focuses on publicly available Fable 5.2; the Opus 5.5 and GPT-6 Astra tables put its performance and pricing in context. Figures were checked on November 10.
What Fable 5.2 brings
Anthropic reports improvements in coding, reasoning and vision, with output generation up to 30% faster than 5.1. Faster generation does not guarantee the same improvement for an entire agent run, which also waits for tools and may take additional steps.
The model has a 1M-token context window and up to 128K output tokens. Adaptive thinking is always enabled; effort controls reasoning depth. Fable can tackle difficult work directly or advise a cheaper agent. I would introduce it where the normal workflow encounters ambiguity and test several effort settings.
Fable and Mythos 5.2 share an underlying model, but Mythos retains full cyber and biology capabilities and is restricted to verified organizations. Its specialized results should not be attributed to publicly available Fable.
API pricing
| Model | Input / 1M | Output / 1M | Cache read / 1M |
|---|---|---|---|
| Claude Fable 5.2 | $8 | $40 | $0.20 |
| Claude Fable 5.1 | $10 | $50 | $0.25 |
| Claude Opus 5.5 | $4 | $20 | $0.20 |
| GPT-6 Astra | $10 | $50 | $1.00 |
These are standard USD API rates, not Claude or ChatGPT subscriptions, checked against vendor documentation and OpenRouter. Cache reads exclude initial writes: a five-minute write costs $10/M for Fable and $5/M for Opus. Astra requests above 272K input tokens incur twice the input and cache rates and 1.5 times the output rate for the full request. Its base row therefore does not describe every long-context workload.
Fable fast mode costs $16/$80 per million input/output tokens. Pay for faster responses when waiting has a measurable cost.
Benchmarks, with the methodology attached
| Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| FrontierCode 2.0 | 49.1 % | 32.4 % | 45.5 % | 42.8 % |
| FrontierSWE v2 | 69.9 % | 56.3 % | 62.3 % | 65.5 % |
| AutomationBench | 51.4 % | 31.4 % | 42.5 % | 41.4 % |
| Terminal-Bench-Science 0.1 | 74.6 % | 52.6 % | 58.7 % | 68.1 % |
| HLE with tools | 74.2 % | 65.6 % | 67.7 % | 57.2 % |
| Chartography (no tools) | 73.4 % | 44.8 % | 64.4 % | 71.0 % |
Compared with Fable 5.1, base input, output and cache-read rates are 20% lower. AutomationBench rises from 31.4% to 51.4%, a gain of 20 percentage points. This is a vendor evaluation result, not a guaranteed improvement in a business workflow.
All rows come from Anthropic's launch evaluation. Higher is better. Scores use the highest available effort except the FrontierCode results for Fable 5.2 and 5.1, which use xhigh. This is neither an equal-budget comparison nor this blog's independent test. HLE uses tools; do not combine it with no-tools results.
Fable leads this selection. FrontierSWE is relevant to engineering and AutomationBench to business automation, but neither predicts your ticket completion rate. Restricted Mythos cybersecurity scores are intentionally excluded.
Independent evaluator Artificial Analysis reports roughly 65 Intelligence Index points for Fable at max, alongside 46.4 minutes and $26.54 per task in that evaluation. These are benchmark costs, not typical invoice-processing costs. They show why max effort should be a deliberate choice.
Cost per token versus cost per completed job
A hypothetical run with 100K uncached input and 10K billed output tokens costs $1.20 on Fable, $0.60 on Opus and $1.50 on Astra. Output includes billed reasoning tokens; tools and retries add cost. Equal token counts do not imply equal consumption on the same task.
Anthropic reports Fable at high scoring 45.3% on FrontierCode against Opus at max at 45.5%, for roughly half the cost per task. This is a vendor result on one evaluation, but a useful reason to compare model-and-effort combinations rather than price lists alone.
Where I would deploy each model
Use Opus 5.5 as the baseline for demanding everyday engineering and knowledge work. Escalate to Fable 5.2 for cross-repository migrations, unclear incidents or contradictions across contracts, purchase orders and invoices.
Keep Astra in the evaluation where an OpenAI integration already works. A better score has to pay for integration changes, operational work and regression risk before it justifies a migration.
In n8n or Make, a cheaper model can classify email and extract fields. A validator routes contradictions to a specialist. Payment approval or CRM writes remain explicit workflow steps. Log model, effort, usage, corrections and escalation reasons.
Migration and the production decision
Fable 5.2 enforces unchanged-history checks for all accounts. Only Fable/Mythos 5.2 read its thinking blocks; the older computer-use tool fails outside Bedrock, and a 5.2 executor needs a 5.2 advisor. Custom routers must respect these boundaries. Check endpoint-specific retention before sending sensitive data.
Select 30–50 completed real tasks and define acceptable outcomes before testing. Compare several effort levels, correct completion, human repair time, cost and p95 latency. Deploy Fable where that combined bill improves. That is the result worth paying for.
Testing Fable 5.2 on a website brief
Higgsfield added Fable 5.2 to its Supercomputer model picker on November 9. On November 15, Zaujaloma published a web design test using that integration. This gives a product team a way to try Fable on a familiar deliverable before changing its development workflow. Higgsfield changelog · Zaujaloma’s test
I’d give Fable and the team’s current model the same approved brief, two product photos and the exact offer copy. On a phone, follow each page from the offer through to submitting a test enquiry. Track the time someone spends correcting a price, fixing unreadable copy or adding a missing confirmation. Count those minutes when comparing the cost of a page the team can actually ship.
Sources
Anthropic: Fable 5.2 · Claude pricing · GPT-6 Astra documentation · API migration · Artificial Analysis
OpenRouter: Fable 5.1, Fable 5.2, Opus 5.5, GPT-6 Astra.