StreamGist's recommendation engine is deterministic, with LLM narration at the edges. The user request path makes zero AI calls. Games are ranked by deterministic code, never written by an LLM.
Three tiers use LLMs, each on a different cost and risk profile. Only the user-request tier serves games, and it never calls an LLM.
Logged figures come from production telemetry; estimated figures are marked. Logged spend sums the per-batch cost ledger and per-call cost logs on the ops droplet. Each Anthropic API call writes its token usage and computed cost. Estimates apply where per-call cost is not directly logged.
Known logged spend: $114.13. Estimated total including internal briefings and on-demand narration: ~$138–$142. Tier 1 (overnight, amortized) and Tier 4 (internal ops) absorb nearly all spend. The user request path is $0.
Separate from the personalized engine above. The anonymous Best Games page is served from a single precomputed row, not a per-user ranking. Figures below are point-in-time, as of the report date.
Rebuilt once per day by a deterministic SQL job, read in a single indexed lookup, with a live SQL fallback if it goes stale.
| Component | Tier | Failure mode | What users see |
|---|---|---|---|
| Recommendation serving | Tier 2 | Anthropic + OpenAI both down | No degradation to recommendation serving. Cached results and the deterministic SQL ranking continue without LLM calls. |
| Catalog enrichment | Tier 1 | LLM unreachable during overnight job | Newly-discovered games not enriched until next successful run. Existing catalog and recommendations continue serving. |
| Public-page game copy | Tier 1 | Copy generation fails for a game | Public page renders the game without a marketing blurb. The game itself still appears. |
| “Why this game for me?” rationale | Tier 3 | LLM call fails on click | The personalized sentence is absent. The ranked game itself is unaffected. |
| Weekly saved-game recap | Tier 3 | Weekly cron LLM call fails | Recap email not sent that week. Catalog and recommendations unaffected. |
| Daily ops briefing | Tier 4 | LLM unreachable | Operator does not receive that day's summary. Product unaffected. |
No. The recommendation API serves each user from a short-lived cache, and on refresh ranks the games with a deterministic SQL query and Python scoring. No LLM client imports or calls are present in the request path, verified by source-level review.
Recommendations keep serving to every user, from cached results and a deterministic SQL ranking that is also LLM-free. Only AI-generated narration (the per-click "Why this game?" text and the weekly recap email) is affected.
No. StreamGist makes inference-only API calls to third-party model providers. Personal user data is not used as training material for any model.
Yes, and it is disclosed. Blog articles are drafted by an LLM from StreamGist's own data, then reviewed by a person before publishing. Per-game public copy and rationales are LLM-generated overnight (Tier 1). The recommendations themselves are never written by an LLM.
About $138 to $142 over the most recent 30 days. $114.13 is directly logged from production telemetry; the remainder is estimated for internal ops briefings and on-demand narration. Nearly all spend is absorbed by overnight precompute and internal ops. The user request path costs $0.
The report is reviewed at least quarterly, and republished sooner when the underlying figures or architecture change materially. The figures in this version reflect the 30 days ending 2026-05-23.
Not at this time. The claim is verified by source-level review of the recommendation handler; the full source under review is private. The verification process is documented in the methodology section below. Independent audit may be added in future revisions.
batch_runs), 30 days ending 2026-05-23UTC. Rationale ($3.15) & public-copy ($0.41): per-call cost=$X markers from droplet cron logs over the same window.Related reading: How StreamGist builds recommendations · Execution systems (the framework this report applies) · The operating loop (Tier 4 as a live system) · Partners.
Prepared by Matthew Juszczyk, Founder · Reviewed 2026-05-23 · v1.0