`epithre-prme`

Premium tier for long-context workloads. Use when prompts exceed 32K tokens or when you need the strongest reasoning available.

Capabilities

Capability	Notes
Tier	Premium
Context window	180,000 tokens (~135K English words / ~140K Indonesian)
Max output	16,384 tokens
Modalities	Text only (no vision)
Tool use	Yes
Extended thinking	Yes; recommended for hard problems
Structured output	Yes; json_object + json_schema
Prompt caching	Yes; especially valuable here given long stable contexts

When to use

Whole-codebase analysis (e.g., "review all changes in this PR")
Long-document Q&A (multi-pasal legal analysis, 100-page contracts)
Multi-step reasoning where chain-of-thought adds real value
RAG with very large retrieved context (top-50 chunks not top-5)

When NOT to use

Image / video input - use epithre-omni or epithre-lyt (prme is text-only).
Short chat (under 8K tokens) - epithre-omni is faster and cheaper.
High-throughput - prme is single-stream serial; latency under load is higher than omni.

Pricing

Input: Rp10,000 / 1M tokens
Output: Rp35,000 / 1M tokens
Cache: same multipliers (1.25x write, 0.1x read)
Batch: 0.5x

Performance characteristics

Strongest in our lineup at:

Multi-document synthesis (10+ documents in context)
Long-chain reasoning (5+ logical steps)
Code review / generation in unfamiliar codebases
Indonesian legal pasal cross-reference analysis

Latency

Non-streaming, 1000-token reply on 100K-token input: ~30-90s
With thinking enabled: +20-60s
Streaming first-token: ~3-15s depending on input length

Plan for higher latency vs omni. Use streaming UX or background-job pattern.

Caveats

Single-stream serial backend. If your traffic burst exceeds the per-key concurrency cap (default 10), requests queue.
No vision. If you need to analyze images alongside text, use epithre-omni (49K context) for the vision step, then pass extracted text to prme for deeper analysis.
min_p parameter is stripped server-side (backend quirk; doesn't affect output quality).

See also