Fable 5 — compressed context

anthropic/claude-fable-5-compressed-context is an experimental mode of Claude Fable 5. It's the same model at the same per-token prices, but long context is handled more compactly: large conversations bill noticeably fewer input tokens.

Experimental — use at your own risk

This mode is an experiment. On most tasks the answers are indistinguishable from regular Fable 5, but tasks requiring ultra-precise verbatim recall of huge documents may diverge. Use at your own risk; for critical scenarios stick with regular anthropic/claude-fable-5.

How much it saves

Per-token prices match regular Fable 5 exactly ($1.548 input / $7.74 output per 1M). The savings come from long context being counted as fewer tokens. A live measurement through NordRouter — identical ~24 KB system prompt:

	prompt_tokens	Input cost
claude-fable-5	5,053	~$0.0078
claude-fable-5-compressed-context	1,287	~$0.0020

−75% input tokens in this example. The longer the context, the bigger the win: in agentic coding and long chats, where history is resent on every step, the budget savings reach 70% and more. On short requests there's almost no difference.

Anthropic format only

The mode works only via the Anthropic-compatible endpoint /v1/messages:

• ✅ Claude Code
• ✅ Anthropic SDK (Python / TypeScript)
• ✅ Any client using ANTHROPIC_BASE_URL
• ❌ OpenAI-format /v1/chat/completions — not supported for this mode

How to try it

Claude Code

Set the model in ~/.claude/settings.json (full config here):

"ANTHROPIC_MODEL": "anthropic/claude-fable-5-compressed-context"

Anthropic SDK / curl

curl

curl https://nordrouter.com/v1/messages \
  -H "x-api-key: sk-nr-YOUR-KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "anthropic/claude-fable-5-compressed-context",
    "max_tokens": 1024,
    "messages": [{ "role": "user", "content": "Hello!" }]
  }'

Python (Anthropic SDK)

from anthropic import Anthropic

client = Anthropic(base_url="https://nordrouter.com", api_key="sk-nr-YOUR-KEY")

resp = client.messages.create(
    model="anthropic/claude-fable-5-compressed-context",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello!"}],
)
print(resp.content[0].text)

Caching

Prompt caching works in this mode too (same rates: write ×1.25, read ×0.1). One nuance: cache matches can be partial — some of the repeated context occasionally bills at the regular input rate. Since the input itself is counted severalfold smaller, the total still comes out cheaper than regular Fable 5 with a perfect cache.

Which to pick when

Scenario	Model
Agentic coding, long chats, large documents in context	claude-fable-5-compressed-context — maximum savings
Verbatim quoting/auditing of huge texts, critical production pipelines	claude-fable-5 — the regular mode
Short requests without long context	No difference — use the regular one