API cost monitoring

Three jobs of cost monitoring

Most cost-monitoring work falls into three buckets: project future cost, react to price changes, audit historical spend. TensorFeed has a dedicated endpoint for each.

Job 1: Project future cost across providers

Pure math, but agents pay 1 credit for the canonical abstraction so they do not have to maintain pricing tables in their own code:

from tensorfeed import TensorFeed

tf = TensorFeed(token="tf_live_...")

projection = tf.cost_projection(
    models=["Claude Opus 4.7", "GPT-5.5", "DeepSeek V4 Pro"],
    input_tokens_per_day=5_000_000,    # busy production agent fleet
    output_tokens_per_day=1_500_000,
    horizon="monthly",
)

# Cheapest at the top
for r in projection["ranked_cheapest_monthly"]:
    print(f"{r['model']}: ${r['monthly_total']}/mo")

Use the same call to validate budget changes: re-run with new token volume estimates before signing off on a quarterly budget. Returns daily/weekly/monthly/yearly breakdowns so finance teams can plug into their existing reporting cadence.

Job 2: React to price changes

The AI pricing market moves fast. New models drop every few weeks; existing models get repriced; budget tiers compete on input cost. Three reactive patterns:

Realtime price watch:

tf.create_watch(
    spec={
        "type": "price",
        "model": "GPT-5.5",
        "field": "blended",
        "op": "lt",
        "threshold": 20,  # blended $/1M tokens
    },
    callback_url="https://your-finops.example.com/webhooks/price",
    secret="any-shared-secret",
)

Daily/weekly digest:

tf.create_digest_watch(
    cadence="weekly",
    callback_url="https://your-finops.example.com/webhooks/weekly",
)
# Fires every 7 days with a curated summary of pricing changes,
# new/removed models, regardless of whether anything dramatic happened.

Forecast:

forecast = tf.forecast(
    target="price",
    model="Claude Opus 4.7",
    field="blended",
    lookback=60,
    horizon=14,
)
# forecast["confidence"]["label"]  - low / medium / high
# forecast["forecast"]              - day-by-day predicted with 95% CI

The forecast is a conservative linear-regression projection with explicit "statistical inference, not a guarantee" disclaimers. Treat low-confidence forecasts as no signal rather than a directional call.

Job 3: Audit historical spend

For every TensorFeed bearer token your agents use, the per-token usage history is one free call away:

usage = tf.usage()
# usage["total_calls"]
# usage["total_credits_spent"]
# usage["by_endpoint"]   - per-endpoint count and credits
# usage["recent"]        - last 25 calls with timestamp

The same data renders in the human-facing /account dashboard if you prefer a UI. Token-scoped, no audit dance, no extra access requests.

Job 4 (optional): Compare cost-effectiveness

When picking between two models for a workload, compare them side by side:

tf.compare_models(ids=[
    "Claude Sonnet 4.6",
    "GPT-4o",
    "DeepSeek V4 Flash",
    "Gemini 2.0 Flash",
])
# Returns rankings.cheapest_blended, plus benchmarks side-by-side

For the daily TensorFeed slack channel

A common pattern: a digest watch fires every morning to a Slack webhook. The digest includes any pricing changes overnight and any new models launched. The Slack channel then has a daily message at 7am UTC that takes the team 10 seconds to read. No human has to remember to check the pricing page.

Recommended TensorFeed endpoints (in priority order)

• /api/premium/cost/projection — workload cost across 1-10 models, 4 horizons
• /api/premium/watches — realtime price-drop notifications
• /api/premium/forecast — linear-regression price forecasting with CI
• /api/payment/usage — per-token audit log (free with bearer token)
• /account — human-facing dashboard for the same data