Moonshot / Sep 15, 2026

Moonshot v1 Auto Route Pins for 128K API Buyers

Use live Moonshot and Kimi route rows, explicit model pins, and 128K acceptance tests before approving long-context API traffic.

Keyword report: 2026-09-14Tier 1/2 developer focusSources checked Sep 15, 2026

This guide uses source checks from Sep 15, 2026. Provider and gateway prices can change; preserve the checked date with every forecast.

Why This Topic Matters Now

The Sep 14 keyword report did not find many new generic searches, but one small signal matters: a Tier 1/2 query looking for Moonshot documentation around `moonshot-v1-128k`. That is not a broad awareness query. It is an evaluator searching for an exact route family, probably to understand long context, OpenAI-compatible setup, and whether a 128K-class model belongs in a production stack. The previous AIWave Moonshot article covered context receipts; this one narrows the problem to route pinning.

Route pinning is the difference between testing a named long-context model and letting an auto route hide the decision. Auto selection can be useful, but it is hard to forecast unless the receipt records the resolved model, checked date, key group, token classes, output cap, and reviewer acceptance. A Tier 1 buyer should decide when to use `moonshot-v1-128k`, when to test a vision preview, and when an auto route is allowed to resolve on the team's behalf.

Source Facts Checked Today

AIWave /api/pricing checked on Sep 15, 2026 returned success=true, 64 live rows, top-level pricing_version a42d372ccf0b5dd13ecf71203521f9d2, auto_groups=['default'], and group_ratio default=1 and vip=0.9. The static /api/v1/pricing endpoint checked during the same run reported checked=2026-09-10, currency=USD, unit=per_1m_text_tokens, pricing_version 8c7a0c0b30661ccbc13d142cb54d1e4ae445fe774b2c6fa501080db97c7a3e56, and notes that dated base rates are adjusted by the effective account group. Selected static rows were moonshot-v1-128k at $1.80 input and $4.50 output per 1M tokens, moonshot-v1-auto at $75 input and $75 output, kimi-k3 at $4.50 input, $0.90 cache-hit input, and $22.50 output, GLM-5.1 at $2.10 input, $0.680001 cache-hit input, and $6.5999997 output, and DeepSeek Flash at $0.70 input, $0.0233 cache-hit input, and $2.10 output. VIP-key estimates multiply the same base rows by 0.9.

The Kimi API quickstart opened for this run says the platform is compatible with OpenAI and Anthropic API formats, asks users to prepare an API key, choose a model, and configure a base_url, and presents Kimi K3 as a model for long-horizon programming and knowledge-work scenarios with 1M token context and vision understanding.

Kimi API pricing opened for this run states that API billing is based on token consumption, that input and output tokens are billed separately, that Web Search is billed independently at $0.004 per invocation, and that context caching can discount repeated system prompts and reference documents. Those provider facts are useful context, while AIWave route rows and request receipts remain the gateway evidence for AIWave forecasts.

Planning Matrix

A source-dated planning matrix keeps the page useful for engineers and procurement reviewers. It turns a search query into an auditable route decision instead of a loose model preference.

Route choiceWhen to use itEvidence required
moonshot-v1-128kStable long-context text evaluationsroute string, source date, input and output cap
moonshot-v1-128k-vision-previewImage-aware long-context trialsmedia policy plus route receipt
moonshot-v1-autoManaged model selection after acceptanceresolved model and variance review
kimi-k3Coding-agent and 1M-context trialscache plan and search policy
default keyBase-rate trialgroup_ratio 1.0 receipt
VIP keyDiscounted key-group estimategroup_ratio 0.9 receipt

Implementation Pattern

The implementation pattern keeps credentials as placeholders, pins the AIWave base URL, records the model, and leaves room for route-specific controls. Production applications should move credentials into environment or secret storage.

from openai import OpenAI

client = OpenAI(api_key="YOUR_API_KEY_HERE", base_url="https://aiwave.live/v1")

ROUTE_PINS = {
    "long_context_review": {"model": "moonshot-v1-128k", "max_tokens": 900},
    "coding_agent_trial": {"model": "kimi-k3", "max_tokens": 1100},
}

def run_pinned_route(task: str, prompt: str):
    route = ROUTE_PINS[task]
    response = client.chat.completions.create(
        model=route["model"],
        messages=[{"role": "user", "content": prompt}],
        max_tokens=route["max_tokens"],
        temperature=0.15,
    )
    return {
        "model": route["model"],
        "api_key": "YOUR_API_KEY_HERE",
        "pricing_checked_at": "2026-09-15",
        "pricing_version": "a42d372ccf0b5dd13ecf71203521f9d2",
        "usage": response.usage,
    }

Pin Before You Forecast

The first control is a model pin. A forecast that says Moonshot or Kimi is too vague for production approval. Use the exact route string in configuration, store the checked pricing source, and keep the applied key group with the receipt. If the team later allows auto selection, the ledger should record which model resolved for each request and why the route was allowed to vary.

Treat Auto as a Policy

An auto route is not a shortcut around review. It is a policy that needs boundaries: eligible task classes, maximum output, allowed data classes, fallback behavior, and variance review. A long-context summarizer may tolerate auto selection if the final receipt records the resolved route. A finance-sensitive workflow should stay pinned until enough receipts prove stable behavior and acceptable spend movement.

Use 128K Acceptance Tests

A 128K route should be tested with realistic but redacted context packs. Use one legal-policy pack, one codebase map, one support-history pack, and one retrieval-heavy knowledge task. Measure answer quality, input size, output size, rejection cases, and whether the route obeys instructions at length. Passing a short prompt does not prove a route is ready for long-context production.

Keep Vision Preview Separate

Vision preview routes should have their own acceptance path. Screenshots and images introduce data classification, resolution, redaction, and storage questions that plain text routes do not. Store whether the image came from a user, a test fixture, or a sanitized screenshot. Do not let a successful text route silently authorize image-aware traffic.

Control Search and Cache

Kimi documentation makes search and context caching visible cost dimensions. Coding-agent workflows should decide when Web Search is allowed, how many calls can occur, and which prompt prefix is stable enough for cache measurement. Search calls, cache-hit input, cache-miss input, and output should be separate fields in the receipt rather than one blended line.

Procurement Review

Procurement should ask for the route string, source URL, checked date, public base row, key group, output cap, resolved model for any auto route, and a sample receipt. If the route is auto, the review should also ask what task classes are eligible and what variance would trigger a rollback to pinned models. Without that evidence, the budget is a trial note, not an approved route.

Final Checklist

A Moonshot or Kimi long-context route is ready when the model is pinned, 128K acceptance tests pass, vision preview traffic is separated, search and cache are governed, auto selection has a policy, and every receipt preserves route, date, key group, and usage. Recheck live AIWave pricing before traffic promotion and keep provider documentation separate from gateway evidence.

Source Links

Related AIWave Links