Claude Fable 5.2 Review 2026: Pricing, Caching, Fast Mode and 1M Context Limits

Claude Fable 5.2 reached general availability on November 9, 2026, with a 1M-token context window and pricing that undercuts its predecessor. The model targets technical and agentic workloads such as long-horizon coding, research and multi-agent coordination. Users access it through the Claude API, paid plans in Claude apps, AWS Bedrock, Google Cloud and Microsoft Foundry.

Claude Fable 5.2 pricing and caching costs

Standard API rates stand at $8 per million input tokens and $40 per million output tokens. These figures represent a 20 percent reduction from Fable 5.1 rates of $10 and $50. Cache reads cost $0.20 per million tokens. Five-minute cache writes run $10 per million, while one-hour writes cost $16 per million. The batch API halves these prices to $4 and $20.

Fast mode, which increases output speed up to 2.5 times, carries rates of $16 input and $80 output per million tokens. All caching and batch discounts apply on top of the base or fast-mode rates. Anthropic requires a minimum 30-day data retention period for most accounts, though zero-retention exceptions exist for verified users.

Real costs depend heavily on effort level and reuse patterns. High-effort runs increase token consumption, while repeated prompts benefit from cache hits that drop effective input cost to $0.20. Independent tests show average task costs of roughly $9 at high effort and $26.54 at maximum effort on the Artificial Analysis Index.

Performance benchmarks and context limits

Fable 5.2 leads or matches top models on several technical benchmarks while using adaptive thinking at low, medium, high, xhigh or max effort. On FrontierSWE v2 it scores 69.9 percent compared with Opus 5.5 at 62.3 percent. FrontierCode 2.0 reaches 49.1 percent at xhigh effort versus 45.5 percent for Opus. OSWorld 2.1 hits 89.3 percent and Terminal-Bench-Science records 74.6 percent.

The 1M-token context supports large codebases and extended agent sessions, with a maximum output of 128K tokens. Outputs run up to 30 percent faster than Fable 5.1, and safety blocks drop by about 55 percent. One documented run coordinated roughly 1,000 agents to formalize a proof of Fermat’s Last Theorem in Lean within 30 hours.

Reviewers note strong results in game development, design and agent reliability, yet Opus 5.5 remains preferable for routine tasks because it finishes faster and cheaper. High-effort settings raise per-task token counts, so teams must monitor usage when scaling to production workloads. The model carries a June 2026 knowledge cutoff and accepts text-plus-image input while producing text output.

Teams running repeated prompts or long agent chains should enable caching early and test fast mode on time-sensitive jobs. Effort level defaults to high via API and medium in apps and Code environments. Production deployments benefit from monitoring token usage per task, because maximum-effort calls can quickly exceed baseline estimates.