None
None
Long Story Video Skill
Turn your idea into 10s–10min video—AI writes scripts, prompts, and generates footage automatically.
Ads Video Skill
Generate professional ads and sales videos—AI auto-generates scripts, prompts, and footage.
3D Science Explainer Video Skill
Convert scientific concepts into stunning 3D explain animations
Feedback
freeTrialImage.bannerPity
freeTrialImage.upgradeUnlock
- ✓freeTrialImage.benefitHd
- ✓freeTrialImage.benefitWatermark
- ✓freeTrialImage.benefitUnlimited
claude fable
Claude Fable 5.2 is the escalation tier for punishing reasoning and marathon agent runs: a million-token context, thinking depth you dial by effort.
All Tools
Discover our comprehensive AI-powered animation toolkit
When Claude Fable 5.2 Is the Right Escalation
An escalation tier aimed at punishing reasoning and marathon agent runs — Claude Fable 5.2 mirrors Claude Mythos 5.2 spec for spec.
- Engineered for Punishing WorkloadsIf cranking effort on Claude Opus 5.5 still misses the bar, this is the next rung: Claude Fable 5.2 handles stubborn reasoning and agent tasks that run for hours.
- Million-Token Context, 128k RepliesA million tokens of context are available at both default and maximum settings, and replies can run to 128k tokens — the same on Claude Fable 5.2 and Claude Mythos 5.2.
- Thinking Depth Set by EffortThinking cannot be switched off. The effort parameter instead chooses how deep it goes — low, medium, high, xhigh, or max — and the API defaults to high.
Bringing Claude Fable 5.2 Into Production
Settle the build path, earn the upgrade with evals, then work through the breaking changes before real traffic arrives.
What Ships With Claude Fable 5.2
On-demand compaction, tool swaps mid-conversation, the computer use toolset, fast mode, strict tool use, and clean refusal handling all ship in Claude Fable 5.2.
On-Demand Compaction (Beta)
With the compact-2026-09-04 beta header you decide when older turns get summarized, which keeps history append-only and keeps the prefix check happy.
Swap Tools Mid-Conversation (Beta)
Rather than editing the tools array, add or drop tools through a system message; send the inline-tools-2026-09-15 header on the Claude API, Claude Platform on AWS, and Anthropic-managed Google Cloud.
Computer Use Through a Toolset
Access it via computer_toolset_20260801; declaring the older computer_20251124 tool returns a 400 invalid_request_error on every platform except Amazon Bedrock.
Where Fast Mode Runs
You get fast mode on the Claude API, on Anthropic-managed and Google-managed Google Cloud, and in Claude in Microsoft Foundry as a separate model, claude-fable-5-2-fast.
Schema-Valid JSON via Strict Tool Use
Keep tool_choice on auto and set strict: true, or move to JSON outputs; forcing tool use throws an error, and customer-managed encryption keys preclude structured outputs.
Refusals, Fallback, and What You Pay
Refusals arrive as HTTP 200 with stop_reason refusal and a stop_details object; eligible categories retry on Claude Opus 4.8 or Claude Opus 5, and fallback credit covers the prompt-cache cost.
Claude Fable 5.2 Questions, Answered
Straight answers on picking a starting model, surviving migration, and what Claude Fable 5.2 costs.
Which model should I start on — Opus 5.5 or Claude Fable 5.2?
Default to Claude Opus 5.5. Move up to Claude Fable 5.2 only after higher-effort runs on Opus still fail to clear your evaluation bar.
Why does touching an earlier turn trigger a 400?
The prefix check compares the system prompt, tools, and earlier messages ahead of a thinking block, and any edit fails the call. A stopgap: send thinking-binding-controls-2026-08-01 with prefix_mismatch_behavior set to drop_block.
Who can read thinking blocks written by Claude Fable 5.2?
Just Claude Fable 5.2 and Claude Mythos 5.2 read one another's blocks. Blocks the active model cannot parse are stripped before delivery, and you are not charged for them.
Is there a way to disable thinking or force a named tool?
Not directly. Thinking enabled with budget_tokens, thinking disabled, and tool_choice pinned to any or a named tool all return 400. Use effort levels and strict tool use instead.
How do Claude Fable 5.2 rates stack up against 5.1?
Both directions cost less than Claude Fable 5.1, which charged $10 per million input and $50 per million output tokens. Cache reads remain at 0.025x base input, and batch input runs $4 per million tokens.
What does a refused request actually look like?
You get HTTP 200 carrying stop_reason refusal and a stop_details object that names the policy area. Select categories can retry through server-side fallback, and charges vary by category.
Try Claude Fable 5.2 Before You Commit
Earn the upgrade: confirm Claude Opus 5.5 truly falls short, lock in a build path, then rehearse all four breaking changes inside your own integration before real traffic reaches Claude Fable 5.2.