How to Use Claude Fable 5 API with EvoLink
model: "claude-fable-5", verify that it works, then decide where Fable 5 belongs in your routing policy.Prerequisites
| Requirement | Why it matters |
|---|---|
| EvoLink account | Needed to create an API key |
| Server-side API key storage | Prevents exposing credentials in the browser |
| Model ID | Use claude-fable-5 |
| Test prompt | Pick a hard task where Fable 5 should matter |
| Cost logging | Fable 5 is a premium route, so measure tokens and retries |
Step 1: Create an EvoLink API Key
Create an API key in the EvoLink dashboard. Store it server-side as an environment variable:
export EVOLINK_API_KEY="your_api_key_here"Do not put API keys in client-side code, public repositories, or frontend environment variables that are exposed to the browser.
Step 2: Make a Minimal Request
Start with the smallest Claude Messages API request that proves the route works.
curl https://direct.evolink.ai/v1/messages \
-H "Authorization: Bearer $EVOLINK_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "claude-fable-5",
"max_tokens": 1024,
"messages": [
{
"role": "user",
"content": "Analyze this architecture decision: should we split billing, usage logs, and model routing into separate services?"
}
]
}'If the request succeeds, log the model, latency, input tokens, output tokens, and the quality of the answer. The first call proves availability. It does not prove that Fable 5 should become your default.
Step 3: Use a Real Test Prompt
Fable 5 is designed for harder work. A generic "hello world" prompt is not a useful evaluation. Use a prompt that resembles the work you would actually pay a premium for.
| Test prompt type | Good Fable 5 evaluation |
|---|---|
| Repo architecture | "Given these modules and constraints, propose a migration plan and risk order." |
| Coding agent trace | "Review this failed agent run and identify the next safest action." |
| Long-context analysis | "Synthesize this spec, log excerpt, and incident timeline into a root-cause plan." |
| High-risk decision | "Compare three implementation paths and identify where each can fail." |
Step 4: Add Routing Rules
Do not route all Claude traffic to Fable 5. Use an escalation policy.
| Request type | Suggested route |
|---|---|
| Simple extraction, classification, or short rewrite | Haiku or lower-cost route |
| Everyday coding and production assistant work | Sonnet or Opus depending on quality need |
| Complex coding, long agent loops, high-value reasoning | Opus 4.8 as the strong default |
| Frontier-difficulty tasks where failure is expensive | Fable 5 |
Step 5: Treat Advanced Parameters Carefully
Anthropic documents Fable 5 with adaptive thinking behavior, thinking output behavior, and fallback-related behavior. EvoLink users should verify which advanced parameters are supported by the current route before relying on them in production.
| Parameter or feature | Safe rollout guidance |
|---|---|
model | Use claude-fable-5 |
max_tokens | Required for Claude Messages API requests; start with a bounded output |
messages | Start with a standard Claude Messages API user turn |
system | Use a top-level system instruction when needed |
thinking | Verify support before depending on it in production |
| Tool use | Verify current EvoLink docs and account support |
| Vision input | Verify current EvoLink docs and account support |
| Streaming | Verify current EvoLink docs and account support |
| Prompt caching | Use when supported and your repeated context is stable |
Step 6: Measure Cost Per Completed Task
Official pricing is useful, but your production cost comes from the whole workflow.
| Cost driver | What to measure |
|---|---|
| Input tokens | Are you sending too much context? |
| Output tokens | Are responses longer than necessary? |
| Retries | Does Fable reduce failed attempts compared with Opus? |
| Cache usage | Are stable instructions and repeated context cached? |
| Fallbacks | Are failed or blocked requests visible in logs? |
The right comparison is not "Fable is more expensive per token." The right comparison is "Fable costs X per accepted task and reduces Y amount of review, retry, or repair work."
Step 7: Build a Production Checklist
Before sending real user traffic:
| Check | Pass condition |
|---|---|
| Route availability | claude-fable-5 works for your EvoLink account |
| Pricing visibility | Usage logs and pricing match expectations |
| Prompt replay | Representative hard prompts produce better outcomes |
| Safeguard test | Sensitive workflows behave predictably |
| Fallback route | Opus 4.8 or Sonnet remains available |
| Budget guardrail | Your app limits context, output, retries, and route escalation |
| Observability | Logs include model, latency, tokens, and failure reason |
Where This Fits in the Claude Family
Sources
- EvoLink Claude Messages API documentation
- Anthropic models overview
- Anthropic pricing
- Anthropic Fable 5 and Mythos 5 launch docs
FAQ
What model ID should I use for Claude Fable 5?
claude-fable-5.What endpoint should I call?
https://direct.evolink.ai/v1/messages. Check the current EvoLink Claude Messages API docs for account-specific endpoint guidance.Should I use Claude Fable 5 for every request?
No. Use Fable 5 only for the hardest requests where stronger reasoning is likely to justify the premium.
How do I test whether Fable 5 is worth it?
Replay real prompts and compare accepted output rate, retries, latency, token usage, and manual cleanup against Opus 4.8.
Can I use prompt caching?
Use prompt caching when the current EvoLink route supports it and your repeated context is stable. It is especially important for long-context premium routes.
Does Claude Fable 5 support 1M context?
Yes. Anthropic documents a 1M token context window.
What should I do if a request is blocked or changed by safeguards?
Log the prompt category, response, model route, and user-visible result. Keep fallback handling and product messaging ready for sensitive workflows.
