Introduction
Anthropic's upgraded long-horizon reasoning and multi-repository engineering model at reduced token cost.
| Field | Value |
|---|---|
| Input | Text |
| Output | Text |
| Publisher | Anthropic |
| Model ID | anthropic/claude-fable-5.2/chat |
| API format | OpenAI-compatible /v1/chat/completions (SSE streaming supported) |
| Context window | 1,000,000 tokens (1M) |
| Max output | 64,000 tokens (default |
| Reasoning effort | low, medium (default), high, max |
| Input & output modalities | Text and |
| Tool calling | Up to 128 functions with |
| Four-tier pricing (per 1M) | $7.20 input · $0.72 cache read · $9.00 cache write · $36.00 output |
| Content policy & billing protection | Standard Anthropic moderation; refused or failed requests are not charged |
Key features & highlights
Four reasoning tiers on a single endpoint
Scale deliberation smoothly across low, medium, high, and max through one parameter instead of maintaining separate client integrations. Routine formatting stays crisp on low while intricate system audits tap max without altering your request format.
Prompt caching engineered for repeat reads
Pin large repositories, legal contracts, or manuscript bibles at the front of messages and settle repeat turns at $0.72/1M cached input tokens—90% below standard input and 10% below Anthropic list pricing across all four token tiers.
One million tokens across text and vision
Combine long codebases, conversation transcripts, and image_url captures inside a 1,000,000-token budget, with up to 64,000 tokens of structured analysis or prose returned in a single pass.
Drop-in SSE streaming and 128-tool control
Point existing OpenAI SDK code to /v1/chat/completions with model ID anthropic/claude-fable-5.2/chat for live SSE streaming, named tool_choice enforcement, custom stop sequences, and sequential or parallel tool calls.
Request parameters
| Parameter | Required | Description |
|---|---|---|
messages | Yes | Ordered conversation history for Claude Fable 5.2 supporting text and |
max_tokens | No | Maximum output tokens (1 to 128,000, default 4,096), including adaptive thinking tokens. |
reasoning_effort | No | Adaptive thinking level: low, medium (default), high, xhigh, or max. |
stop | No | Up to 4 custom stop sequences that terminate output immediately. at least 1 characters · at least 1 items · at most 4 items |
tools | No | Up to 128 function definitions callable by Claude Fable 5.2. at most 128 items |
parallel_tool_calls | No | Enable or disable parallel tool calls within a single assistant turn (default true). Default: |
tool_choice | No | Tool selection mode: auto, none, required, or a named function. |
Pricing
Input $7.20 / Output $36 per 1 million tokens
Use cases
Cross-repository refactoring and migrations
Load database schemas, API contracts, and integration test suites together inside the 1,000,000-token context window to draft synchronized migration steps and rollback checks.
Atmospheric long-form fiction and dialogue
Keep an entire novel draft and character style bible pinned in context to write slow-burn romantic or noir continuations with consistent voice, sensory lighting, and subtext.
Visual UI and diagram inspection
Pair screenshots, wireframes, or system topology images with source files via image_url blocks so Claude Fable 5.2 pinpoints layout regressions and state mismatches in one turn.
Stateful multi-tool engineering agents
Register up to 128 function definitions and toggle parallel_tool_calls to false whenever sequential database checks or deployment steps must wait for prior tool outputs.
Deep root-cause debugging
Switch reasoning_effort to high or max when untangling distributed race conditions, duplicate billing retries, or intermittent production deadlocks across multi-service logs.
Prompt guide & pro tips
Pin static context at the prefix
Place system rules, tool schemas, and reference codebases or manuscripts at the top of messages before dynamic user turns. Keeping that leading prefix byte-identical lets repeat requests settle at $0.72/1M cached reads instead of $7.20/1M.
Match
reasoning_effortto each turnLeave
reasoning_effortat medium for balanced drafting and routine code changes, drop to low for quick extractions, and step up to high or max when diagnosing gnarly defects. Skip prompts that tell the model to think step by step so you do not inflate internal token burn.Reserve
max_tokensfor thinkingBecause internal reasoning and visible text share
max_tokens(default 4,096, ceiling 64,000), raise the limit whenever you choose high or max effort. Under-sizing this cap can truncate the final answer mid-thought.Stream SSE and check tool arguments
Enable stream: true on long replies so clients stay responsive well within the 600-second request timeout. When using tools, validate every returned JSON argument in application code before running side-effecting commands.
Claude Fable 5.2 Chat Overview
Anthropic Claude Fable 5.2 (anthropic/claude-fable-5.2/chat) serves a 1,000,000-token context window, a 64,000-token output ceiling, text and vision inputs, SSE streaming, tool calling, and four adaptive reasoning_effort levels over OpenAI-compatible /v1/chat/completions. Teams use it for multi-repository engineering, visual UI audits, stateful agent loops, and long-form narrative drafting that demands continuity across hundreds of thousands of tokens.
Request Parameters
| Parameter | Type & Range | Default | Notes |
|---|---|---|---|
messages | array (1–256 items) | Required | Ordered system, developer, user, assistant, and tool turns with text or image_url parts. |
max_tokens | integer (1–64,000) | 4096 | Shared ceiling covering both internal adaptive thinking tokens and visible output text. |
reasoning_effort | low, medium, high, max | medium | Controls adaptive deliberation depth per request without swapping model IDs. |
tools | array (up to 128 items) | Omitted | OpenAI-style function definitions; your application runs approved calls and returns tool_call_id. |
tool_choice | auto, none, required, or named object | auto | Forces, suppresses, or pins a specific function call for structured extraction. |
parallel_tool_calls | boolean | true | Set false when dependent tool steps must run sequentially across turns. |
stop | string or array (1–4 items, 1–32 chars) | Omitted | Halts generation immediately on custom delimiters; matched sequences are omitted. |
stream | boolean | false | Set true for Server-Sent Events (SSE) streaming during long responses. |
Adaptive thinking manages token sampling internally, so temperature and top_p are omitted and rejected if submitted.
Four-Tier Prompt Caching & Token Math
Live billing verified on 2026-11-19 settles measured usage across four token tiers—10% below Anthropic's published list rates ($8.00 / $0.80 / $10.00 / $40.00 per 1M tokens). Check /models/claude-fable-5-2 for current rates before high-volume runs.
| Token Tier | SpicyAPI Rate (per 1M) | Anthropic List Rate | Relative Multiplier |
|---|---|---|---|
| Standard input | $7.20 | $8.00 | 1.00x input |
| Cached input read | $0.72 | $0.80 | 0.10x input (90% off) |
| Cache write | $9.00 | $10.00 | 1.25x input |
| Output (includes thinking) | $36.00 | $40.00 | 5.00x input |
Break-Even Math on a 200,000-Token Prefix
- Uncached across 2 turns:
2 × (0.2M × $7.20) = $2.88in input spend. - Cached across 2 turns (1 write + 1 read):
(0.2M × $9.00) + (0.2M × $0.72) = $1.80 + $0.144 = $1.944—already saving$0.936by turn 2. - Cached across 10 turns (1 write + 9 reads):
$1.80 + (9 × $0.144) = $3.096versus$14.40uncached (a 78.5% reduction).
Keep static system instructions, tool schemas, and reference files at the very top of messages with zero dynamic timestamps or random IDs ahead of them. Any byte change early in the prompt invalidates the remainder of the cached prefix, and pausing past the idle TTL evicts the prefix so the following turn incurs a fresh $9.00/1M cache write.
Choosing a Reasoning Effort Level
low: Minimal internal deliberation for rapid classification, formatting, or concise verification turns.medium(default): Balanced depth for routine coding, editorial revisions, and atmospheric narrative continuations.high: Thorough verification for cross-service architecture reviews, migration audits, and multi-file refactors.max: Deepest deliberation for stubborn distributed race conditions, formal invariants, and high-stakes debugging.
Internal reasoning tokens bill at $36.00/1M output tokens and draw from max_tokens. Avoid adding "think step by step" instructions when high or max is already active, and raise max_tokens well above your expected visible answer so thinking never truncates the reply.
Comparing Claude Fable 5.2, Claude Fable 5, and Gemini 4 Neon
| Dimension | Claude Fable 5.2 | Claude Fable 5 (/models/claude-fable-5) | Gemini 4 Neon |
|---|---|---|---|
| Context window | 1,000,000 tokens | 1,000,000 tokens | 1,048,576 shared tokens |
| Max output | 64,000 tokens | 128,000 tokens | 65,536 tokens |
| Reasoning tiers | low, medium, high, max | low, medium, high | low, medium, high, xhigh |
| Structured & tools | Up to 128 tools | Up to 128 tools | Tools + native JSON Schema |
Choose Claude Fable 5.2 when you want max effort depth and lower four-tier caching rates on 1M contexts. Pick Claude Fable 5 when a single pass needs between 64,001 and 128,000 output tokens. Consider Gemini 4 Neon when strict native JSON Schema enforcement is a core pipeline requirement.
Limits and Content Policy
Requests time out after 600 seconds and accept up to 256 messages containing text and image_url parts; convert audio, video, or binary files before sending, and note that this endpoint returns text or tool calls rather than generated images. Claude Fable 5.2 is not in the unrestricted tier: standard Anthropic publisher moderation applies, and refused or failed requests incur no charge.
API quick start
curl --fail-with-body -sN https://api.spicyapi.ai/v1/chat/completions \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $(uuidgen)" \
-d '{
"model": "anthropic/claude-fable-5.2/chat",
"messages": [
{
"role": "user",
"content": "Summarise the following release notes into three bullet points."
}
],
"stream": true
}' \
| jq --unbuffered -Rrj 'select(startswith("data: ")) | .[6:] | select(. != "[DONE]") | fromjson | if .error then "\n[error] \(.error.message)\n" else (.choices[0].delta.content? // empty) end'
See the API reference for authentication, the complete schema, polling, and error handling.










