What happens to Qaf's monthly AI bill when a model at about 5% of today's per-question cost answers non-paying users in the countries where Qaf earns almost nothing, and which country × platform segments should get it.
Today Qaf spends about $59.9k a month on AI. Moving non-paying Android users in the 14 recommended countries (scenario C) to tier 2 brings that to $48.2k–$50.1k: a saving of $9.7k–$11.6k a month (16–19%). In year 2, when Gemini Flash prices double, the saving is about $19k–$21k a month.
Revenue at risk is about $0 directly: every payer, trial, Play grace-period, gift, comp and lifetime user keeps the premium model. The only cost is possible lost future conversion, at most about $1.3k to $2.2k of MRR after a year even if Android billing were fixed, against a saving of $9.7k+ every month.
Update: corrected for a rerank double count, the bill is $54.5k today and $102.3k at 2027 prices. Scenario C alone leaves most of it; see why it is still high and how to reach $12.6k–$21.8k.
Corrected, Qaf's AI bill is $54.5k a month today and $102.3k at 2027 prices. The earlier $59.9k / $108.9k counted every rerank call twice. Scenario C only reaches 21% of it. The cost per question has to fall for everyone, and tier 2 then has to reach more of the free traffic.
Answer generation on Gemini 3.7 Flash is 88% of cost ($47.9k). Google's intro price ends on 31 Dec 2026, which takes this line to about $95.8k. Rerank is only ~$4.6k.
About 47k prompt tokens and 2.5k output tokens over ~3 Gemini calls. 66% of the prompt is ~42 retrieved passages (20 per search, sent with diacritics, footnotes and long IDs); 23% is chat history, a third of it old citation tags. Only ~21% is cached. That is $0.042 per question today and $0.079 in 2027.
~90k people asked questions in the window; the median free user asked 5 a month, the free tier is already 20 welcome questions plus 3 a day, and nobody averages 50 a day. Most free cost sits in markets that pay: Saudi Arabia alone is ~$12.7k a month and 41% of MRR. Two leaks: apps below 1.3.0 skip the free cap (~$5.3k), and guests cost ~$5.2k (up to ~$8.7k counting signed-out questions of people who sign in later).
A payer asks ~162 questions a month: $6.73 of AI today against $9.17–11.13 after store fees. Break-even needs 11–18% of users paying; the actual rate is 0.71%. At 2027 prices a payer costs $12.84, more than they pay after fees, so no conversion rate breaks even until the cost per question falls.
Monthly savings. Tier-2 ranges run from a realistic tier-2 model (tuned DeepSeek from the bench) to the 5%-of-generation target. Options overlap, so they cannot be added up; the packages below stack them properly.
| Option | Lever | Saving today | Saving 2027 prices | Risk | Effort | Confidence |
|---|---|---|---|---|---|---|
| C0 Fix the rerank double count in the cost model Langfuse logs every rerank call twice; counted once, rerank is ~$4.6k, not $9.1k. | measurement | — (baseline $59.9k → $54.5k) | — ($108.9k → $102.3k) | none: it is an overstatement, not a saving | none | high |
| E-A Efficiency package A (core) Short citation refs instead of long IDs, strip old citations from history, de-duplicate passages, cap history at 16k tokens, minimal thinking on the policy check. | efficiency | $8.4–11.7k | $16.9–23.3k | low: citation rendering, expand-chunk mapping, stream resume | ~1–1.5 weeks | medium-high |
| E-A+ Efficiency A, gated items Policy check on Flash-Lite; step cap 6. Each needs a targeted check (missed prohibited queries; truncated 3+-step answers). | efficiency | +$0.9–1.4k (A total $9.3–13.1k) | +$2.0–2.9k (A total $18.9–26.2k) | medium: safety false negatives, answer truncation | days + checks | medium |
| E-B Efficiency package B Top-12 instead of 20 passages, 1–2 searches instead of 2–4, drop diacritics/footnotes from passages, router in the policy call, shorter answers, cache work, response thinking low. | efficiency / retrieval | A+B $23.5–32.6k; top-12 alone $7.9–9.4k (lower if passages are a smaller share of the prompt: $4.8–7.2k) | A+B $46.3–63.6k; top-12 $15.7–18.9k | medium: completeness, coverage of differing madhab views; thinking-low was reverted once | 2–4 weeks incl. evals | medium-low |
| T2-C Tier 2: scenario C Non-paying Android users in the 14 weak-payment countries. | audience | $7.7–9.6k | $17.8–19.7k | very low (~0.5% of converters) | 1–2 dev days + eval | medium-high |
| T2-S1 Tier 2: C + all guests + legacy apps Adds signed-out users and signed-in users on unmetered app versions below 1.3.0. | audience | $11.9–14.7k | $27.3–30.3k | low; about a fifth of payers asked as a guest first | + hours | medium-high |
| T2-S2 Tier 2: S1 + free users outside the 18 payer markets The 18 countries that produce 90% of MRR keep premium for free users. | audience | $18.0–22.3k | $41.3–45.7k | 9.6% of MRR and 22% of trials start outside the list (~$0.4–3.1k first-year MRR); VPN, fairness | low–medium | medium |
| T2-S4 Tier 2: S2 + payer markets after 50 lifetime questions Free users in payer markets get premium for their first 50 lifetime questions. | audience | $22.7–28.1k | $52.1–57.7k | 38–50% of converters pass 50 questions before paying; $1.8–7k first-year MRR | medium (lifetime counter) | medium |
| T2-ALL Tier 2: every non-paying user Only paid, trial, grace, gift, comp and lifetime users keep premium. | audience | $32.8–40.6k | $75.3–83.4k | high: every converter's first impression; $4.7–14k first-year MRR, more on lifetime value | 1–2 dev days + eval | high on cost, low on conversion |
| T2-M Which tier-2 model Tuned DeepSeek ≈30% of a question today / 16% in 2027 (searches 2.4× more). Gemini 3.1 Flash-Lite ≈35% / 17.5% of Flash generation. Claude Haiku 5.5 ≈25–30% / 13–15% once its 5× price on prompts over 100k tokens is counted. Only GLM no-think reaches ~5%, and it lost 88% of comparisons. | vendor | decides where each tier-2 range lands: low end ≈ DeepSeek, high end needs ~5% | same | no candidate passes quality yet (DeepSeek: serious errors in 66% of answers vs 33%) | 2–3 days bench | medium |
| L1 Force update to app 1.3.0+ Meters the legacy apps that skip the free cap today. | limits | $1.8–2.1k (up to ~$3.4k with drop-off); overlaps S1 | $3.4–4.0k | locks out users who cannot update (Iran sideloads); needs an APK/web fallback | hours | high |
| L3 Fewer starter credits (A/B) Starter 20 → 10; or 2/day + starter 10. | limits | $3.1–4.0k; $5.7–7.4k (upper bounds) | $5.8–7.6k; $11.0–14.3k | walls precede 31.6% of first paid starts; causal effect unknown | <1 day + 1–2 week test | medium on cost, low on revenue |
| L4 Sign-in after 3 guest questions; 50/day ceiling; meter reference routes | limits | $1.7–3.3k; $0.5–0.7k; $0.1–0.3k | $3.2–6.3k; $0.8–1.4k; $0.2–0.5k | top-of-funnel loss (sign-in wall) | <1–2 days each | low–medium |
| R1 Cheaper reranker rerank-v4.0-fast instead of pro; or a token-priced reranker. | retrieval | $0.9k (token-priced $2.4–3.8k) | same | small relevance loss; new vendor for token-priced | trivial / medium | high on price, unknown quality |
| V1 Vertex Flex PayGo for free traffic 50% off the same model, with Standard fallback. | vendor | $12.8–20.3k | $25.6–40.6k | latency and throttling on streaming chat; test on 1–5% first | low–medium | medium on price, low on latency |
| V2 Savings plan, price negotiation, startup credits Savings plan 10–20%; ask Google to extend the intro price; startup credits up to $350k one-off over 2 years (≤$250k in year 1, ≈3.6 months of 2027 generation; VC-funded only). | vendor | plan $4.8–9.6k | plan $9.6–19.2k; negotiation $0–48k | lock-in while prices fall; eligibility uncertain | low | low |
| V5 Titles and non-chat agents off gpt-5.6-terra | vendor | $1.2–1.3k | same | low | low | medium |
| X Ads / rewarded ads One rewarded video earns far less than a question costs outside the Gulf. | other | ~$0 net | ~$0 net | brand risk in a religious app | 1–2 weeks | low |
The higher number uses a realistic tier-2 model and each option's low saving; the lower number assumes tier 2 at ~5% of generation and each option's high saving.
| Package | Today | Saving | 2027 prices | Saving | Revenue risk |
|---|---|---|---|---|---|
| P0 Baseline (corrected) No change. | $54.5k | – | $102.3k | – | – |
| P1 Scenario C only Tier 2 for non-paying Android users in the 14 weak-payment countries. | $44.9k–$46.7k | 14–18% | $82.7k–$84.6k | 17–19% | very low (~0.5% of converters) |
| P2 Quick wins Tier 2 for C + all guests + unmetered legacy (<1.3.0) apps; efficiency package A; rerank-fast; titles/agents off terra. | $28.4k–$33.7k | 38–48% | $51.8k–$59.7k | 42–49% | low |
| P3 P2 + full efficiency P2 + efficiency B after evals (top-12 passages, fewer searches, payload slimming, router, shorter answers, cache work). | $15.1k–$24.2k | 55–72% | $26.6k–$41.3k | 60–74% | low on conversion; medium on answer quality |
| P4 Payer-market focus P3 + tier 2 for every free user outside the 18 countries that make 90% of MRR. | $12.6k–$21.8k | 60–77% | $21.3k–$34.7k | 66–79% | low–medium: 9.6% of MRR and 22% of trials start outside the list |
| P5 Aggressive P4 + payer-market free users premium only for their first 50 lifetime questions; Vertex Flex for remaining free traffic; starter credits 20 → 10. | $8.2k–$17.4k | 68–85% | $12.2k–$24.8k | 76–88% | medium: first-50 rule touches 38–50% of converters, mostly in Saudi Arabia |
| P6 All non-paying users on tier 2 Everyone without a paid, trial, grace, gift or comp plan on tier 2, plus full efficiency. | $6.6k–$15.8k | 71–88% | $8.4k–$18.8k | 82–92% | high: every converter starts on tier 2 ($4.7–14k first-year MRR at risk) |
| R Reference: no tier 2 Efficiency A+B, rerank-fast and the terra move only; everyone keeps the premium model. | $20.0k–$29.1k | 47–63% | $36.9k–$54.1k | 47–64% | low on conversion; medium on answer quality |
P6 is cheaper than P4 even with a realistic tier 2: about $6.0k a month less today and $16.0k less at 2027 prices. It is held back for revenue risk, not because the gain is small.
Target $12.6k–$21.8k a month today and $21.3k–$34.7k at 2027 prices, against $54.5k / $102.3k. Treat Flex and starter 10 as experiments.
These are exactly the Android segments an objective rule flags: at least 50 active users, conversion and revenue per user both under 25% of Qaf's average, and payment feasibility of "none" or "poor" (sanctions, no card rails, or weak Play billing). Together they ask 22% of all questions and cost 21% of the AI bill, but bring in almost no revenue: across all their Android users there are fewer than 5 payers.
CF-IPCountry. GeoIP alone misses 97.7% of Iranian Android users, who are on VPNs (they appear as Germany or the US); the x-qaf-client-timezone header the apps already send catches 96.6% of them. This rule saves $10.1k (B) for C, within ~4% of the modelled figure; GeoIP alone saves only $7.3k.x-qaf-client-platform=android). Android web needs new user-agent parsing; it is a phase-2 item worth about $0.3–0.4k/month.hasPaidAccess as it is: it excludes trialing and past_due, so trial and Play grace-period users would be downgraded.| Country (Android) | Decision | Payment rails | Active users | Payers | Trials + grace | Questions / mo | AI cost / mo | Saving / mo (B) | Saving / mo 2027 (B) |
|---|---|---|---|---|---|---|---|---|---|
| Yemen YE | C | Poor | 4,164 | <5 | <5 | 79,118 | $3,687 | −$2,962 | −$5,909 |
| Iran IR | C | None | 3,170 | 0 | 0 | 70,467 | $3,075 | −$2,440 | −$4,867 |
| Egypt EG | conditional | Good | 3,367 | <5 | 20 | 51,099 | $2,527 | −$2,051 | −$4,092 |
| Afghanistan AF | C | None | 2,524 | 0 | <5 | 54,723 | $2,384 | −$1,891 | −$3,772 |
| India IN | exclude | Good | 1,584 | 0 | 5 | 23,504 | $1,087 | −$872 | −$1,740 |
| Morocco MA | exclude | Good | 1,577 | <5 | 6 | 22,542 | $1,022 | −$817 | −$1,630 |
| Indonesia ID | exclude | Good | 1,183 | 5 | 13 | 17,826 | $843 | −$679 | −$1,354 |
| Algeria DZ | C | Poor | 1,546 | 0 | <5 | 19,958 | $840 | −$662 | −$1,321 |
| Iraq IQ | exclude | Fair | 1,230 | <5 | 11 | 18,016 | $769 | −$607 | −$1,212 |
| Bangladesh BD | exclude | Fair | 1,562 | 0 | <5 | 18,035 | $759 | −$598 | −$1,193 |
| Ethiopia ET | C | Poor | 668 | 0 | <5 | 14,592 | $612 | −$482 | −$961 |
| Pakistan PK | exclude | Fair | 678 | <5 | <5 | 9,229 | $467 | −$381 | −$760 |
| Syria SY | C | None | 573 | 0 | 0 | 9,227 | $398 | −$315 | −$628 |
| Mauritania MR | C | Poor | 287 | 0 | <5 | 8,999 | $394 | −$313 | −$625 |
| Palestine PS | C | Poor | 421 | 0 | <5 | 6,817 | $283 | −$222 | −$443 |
| Libya LY | C | Poor | 425 | 0 | 0 | 6,295 | $269 | −$213 | −$424 |
| Somalia SO | exclude | Poor | 178 | <5 | 0 | 2,912 | $228 | −$197 | −$393 |
| Sudan SD | C | None | 150 | 0 | <5 | 1,790 | $80 | −$64 | −$127 |
| Niger NE | C | Poor | 70 | 0 | 0 | 1,542 | $70 | −$56 | −$112 |
| Tajikistan TJ | C | Poor | 89 | 0 | 0 | 1,381 | $56 | −$43 | −$87 |
| Chad TD | C | Poor | 69 | 0 | 0 | 1,211 | $54 | −$43 | −$85 |
| Mali ML | C | Poor | 73 | 0 | 0 | 1,409 | $55 | −$42 | −$85 |
Questions and cost include Android users seen only in Langfuse (mostly signed-out guests). Payer and trial counts under 5 are shown as "<5".
Every scenario moves only non-paying users (payers, trials, grace, gifts, comps and lifetime keep premium). Bars show the monthly saving today: dark = interpretation B, light = interpretation A, tick = low case (B).
| Scenario | Users / mo | Share of questions | Today, B: after | Today, A: after | Low case (B) | 2027 prices, B | Month 12, base growth (B) | Deployable rule (B) | GeoIP only (B) | "Everyone" variant: payers / MRR exposed | Conversion risk after 12 mo (MRR) |
|---|---|---|---|---|---|---|---|---|---|---|---|
| A Founder's list Android app + web, non-paying users in YE, AF, IR |
12,747 | 16.1% | $52.6k −$7.3k · −12.2% |
$51.2k −$8.7k · −14.5% |
−$6.5k | $94.4k −$14.5k (A: −$15.9k) |
$152.0k from $181.3k · −$29.3k |
−$7.3k | −$4.8k | <5 / <$100 | $180–$300 billing fixed: $697–$1.2k |
| B A + Egypt A plus Egypt Android |
16,728 | 20.2% | $50.5k −$9.3k · −15.6% |
$48.8k −$11.1k · −18.5% |
−$8.3k | $90.3k −$18.6k (A: −$20.4k) |
$145.7k from $183.2k · −$37.5k |
−$9.4k | −$6.9k | <5 / <$100 | $587–$978 billing fixed: $1.3k–$2.1k |
| C Recommended core Android app + web, non-paying users in the 14 countries flagged by the weak-monetization rule |
18,389 | 21.9% | $50.1k −$9.7k · −16.3% |
$48.2k −$11.6k · −19.4% |
−$8.7k | $89.5k −$19.4k (A: −$21.3k) |
$144.5k from $183.6k · −$39.1k |
−$10.1k | −$7.3k | <5 / <$100 | $360–$600 billing fixed: $1.3k–$2.2k |
| C+EG Conditional add-on C plus Egypt Android, after a Play billing fix or an A/B test |
22,370 | 25.9% | $48.1k −$11.8k · −19.7% |
$45.8k −$14.0k · −23.5% |
−$10.5k | $85.4k −$23.5k (A: −$25.8k) |
$138.2k from $185.5k · −$47.4k |
−$12.2k | −$9.3k | <5 / <$100 | $691–$1.2k billing fixed: $1.9k–$3.1k |
| E C + all platforms in no-payment countries C plus iOS and desktop web in IR, AF, SY, SD (timezone or IP in-country required) |
19,182 | 23.0% | $49.6k −$10.3k · −17.1% |
$47.6k −$12.2k · −20.5% |
−$9.3k | $88.4k −$20.5k (A: −$22.5k) |
$142.9k from $184.1k · −$41.2k |
−$10.8k | −$7.7k | <5 / <$100 | $401–$668 billing fixed: $1.3k–$2.2k |
| D Broad: 28 countries Android in all 28 low/middle-income candidate countries (includes payable EG, MA, ID, IQ, PK, JO) |
35,296 | 37.7% | $42.7k −$17.2k · −28.7% |
$39.4k −$20.4k · −34.2% |
−$15.3k | $74.6k −$34.3k (A: −$37.5k) |
$121.6k from $190.6k · −$69.0k |
−$17.7k | −$14.7k | 22 / $240 | $1.9k–$3.2k billing fixed: $3.7k–$6.1k |
| S1 Sensitivity: C + India, Bangladesh C plus India and Bangladesh Android (no payers yet, but payable markets) |
22,080 | 25.2% | $48.6k −$11.2k · −18.7% |
$46.5k −$13.4k · −22.4% |
−$10.0k | $86.5k −$22.4k (A: −$24.6k) |
$140.0k from $185.0k · −$45.0k |
−$11.6k | −$8.7k | <5 / <$100 | $648–$1.1k billing fixed: $1.8k–$3.1k |
| S2 Sensitivity: all signed-out guests Every signed-out guest, every country and platform |
24,620 | 10.3% | $55.3k −$4.5k · −7.5% |
$54.5k −$5.4k · −9.0% |
−$4.4k | $99.9k −$9.0k (A: −$9.9k) |
$160.5k from $178.6k · −$18.1k |
– | – | 10 / $104 | $0–$0 billing fixed: $713–$1.2k |
Baseline $59.9k/month today, $108.9k at 2027 Gemini prices with today's volume. "Low case" assumes Android-app questions cost 0.883× the average (see Method). "Deployable rule" = timezone OR CF-IPCountry instead of the modelled inferred country. "Everyone" variant also moves paying users (not recommended). Conversion risk = MRR lost after 12 months if tier 2 cuts free-to-paid conversion by 30–50%, counting trial and grace pipeline; the second line is an upper bound if Android billing were fixed and these markets converted at Qaf's average.
Cost per question today: 4.59¢ (3.87¢ generation + 0.72¢ retrieval). At 2027 prices: 8.44¢. Gemini 3.7 Flash list price doubles on 1 Jan 2027, when introductory pricing ends.
| Platform | AI cost / mo | Share | MRR | MRR ÷ AI cost |
|---|---|---|---|---|
| Android app | $32,836 | 54.9% | $2,039 | 0.06 |
| iOS app | $17,861 | 29.8% | $4,122 | 0.23 |
| Desktop web | $5,302 | 8.9% | $1,373 | 0.26 |
| Non-question AI | $1,700 | 2.8% | $0 | 0.00 |
| Android web | $1,471 | 2.5% | $26 | 0.02 |
| iOS web | $685 | 1.1% | $8 | 0.01 |
iOS and web cost per question is not measured directly (see Method); they are assumed to cost the average.
| Segment (inferred country · main platform) | Active users | Questions / mo | AI cost / mo | Share of cost | Payers | MRR | MRR ÷ AI cost | Weak monetization |
|---|---|---|---|---|---|---|---|---|
| Saudi Arabia · iOS app | 10,731 | 169,361 | $7,774 | 13.4% | 117 | $1,840 | 0.24 | |
| Saudi Arabia · Android app | 6,458 | 99,192 | $4,514 | 7.8% | 55 | $793 | 0.18 | |
| Yemen · Android app | 4,066 | 74,441 | $3,499 | 6.0% | <5 | <$100 | <0.03 | Yes |
| Iran · Android app | 2,777 | 64,888 | $2,848 | 4.9% | 0 | $0 | 0 | Yes |
| Egypt · Android app | 3,064 | 47,292 | $2,367 | 4.1% | <5 | <$100 | <0.04 | |
| Afghanistan · Android app | 2,424 | 46,219 | $2,068 | 3.6% | 0 | $0 | 0 | Yes |
| United States · iOS app | 1,617 | 24,751 | $1,136 | 2.0% | 37 | $367 | 0.32 | |
| Saudi Arabia · Desktop web | 1,167 | 23,016 | $1,057 | 1.8% | 32 | $389 | 0.37 | |
| United Kingdom · iOS app | 1,242 | 22,272 | $1,022 | 1.8% | 22 | $333 | 0.33 | |
| India · Android app | 1,350 | 20,480 | $956 | 1.6% | 0 | $0 | 0 | |
| Morocco · Android app | 1,387 | 19,878 | $916 | 1.6% | <5 | <$100 | <0.11 | |
| France · iOS app | 1,111 | 18,114 | $832 | 1.4% | 7 | $83 | 0.10 | |
| Indonesia · Android app | 958 | 16,357 | $782 | 1.3% | 5 | $54 | 0.07 | |
| France · Android app | 888 | 16,330 | $766 | 1.3% | 7 | $67 | 0.09 | |
| Kuwait · iOS app | 1,258 | 16,578 | $761 | 1.3% | 21 | $303 | 0.40 | |
| Algeria · Android app | 1,474 | 16,617 | $700 | 1.2% | 0 | $0 | 0 | Yes |
| Iraq · Android app | 1,188 | 16,077 | $688 | 1.2% | <5 | <$100 | <0.15 | |
| United Kingdom · Android app | 624 | 12,287 | $674 | 1.2% | 14 | $186 | 0.28 | |
| Bangladesh · Android app | 1,354 | 14,851 | $626 | 1.1% | 0 | $0 | 0 | |
| Germany · iOS app | 688 | 11,785 | $541 | 0.9% | 10 | $146 | 0.27 | |
| Ethiopia · Android app | 628 | 12,761 | $541 | 0.9% | 0 | $0 | 0 | Yes |
| Egypt · iOS app | 648 | 10,440 | $479 | 0.8% | 6 | $65 | 0.14 | |
| Netherlands · Android app | 188 | 4,407 | $456 | 0.8% | <5 | <$100 | <0.22 | |
| Egypt · Desktop web | 510 | 9,850 | $452 | 0.8% | 5 | $53 | 0.12 | |
| Pakistan · Android app | 557 | 8,418 | $438 | 0.8% | <5 | <$100 | <0.23 | |
| Saudi Arabia · Android app (guests seen only in Langfuse) | 1,517 | 10,349 | $437 | 0.8% | 5 | $61 | 0.14 | |
| Turkey · Android app | 611 | 9,431 | $427 | 0.7% | <5 | <$100 | <0.23 | |
| Uzbekistan · Android app | 579 | 10,135 | $409 | 0.7% | <5 | <$100 | <0.24 | |
| Canada · iOS app | 509 | 8,698 | $399 | 0.7% | 11 | $119 | 0.30 | |
| UAE · iOS app | 608 | 8,250 | $379 | 0.7% | 14 | $176 | 0.47 |
Highlighted rows are in scenario C. Saudi Arabia is Qaf's largest cost and revenue market and stays on premium. Yemen's Android app costs about $3.5k a month against roughly $20 of MRR. MRR is gross (mobile includes VAT); segments with 1–4 payers show MRR as "<$100". Total MRR in the window $7,768.
| Scenario | In-scope questions, Aug | In-scope, Sep | Growth in scope | Growth elsewhere |
|---|---|---|---|---|
| A | 128,813 | 189,547 | +47% | +9% |
| B | 165,999 | 238,582 | +44% | +8% |
| C | 179,628 | 253,543 | +41% | +8% |
| C+EG | 216,814 | 302,578 | +40% | +7% |
| E | 179,628 | 253,543 | +41% | +8% |
| D | 325,948 | 421,557 | +29% | +7% |
Mixpanel chat_message_sent, calendar months. In-scope usage grew ~40–47% month over month vs ~8% elsewhere, so the saving grows over time. One month of growth can't be compounded, so the model uses assumed growth cases: base = +6%/month in scope, +4% elsewhere; high = +12% / +8%.
chat_message_sent) per user, platform, country signals and monthly trend.Monthly change = Σ over moved segments of questions × cost per question × (1 − tier-2 ratio). Per-segment cost per question varies only within the Android app (where Langfuse measures it; target countries sit at 0.92–1.09× the Android mean). Every other platform gets one flat index.
CF-IPCountry reaches the origin; how Autumn reports Play trial status; whether to apply tier 2 to the reference routes (Quran, hadith, entity), which are unmetered today.qaf-t2) or branch inside the dynamic route on request metadata, so no model id lives in code.x-qaf-client-platform (native), x-qaf-client-timezone (all clients, also persisted on the user), CF-IPCountry (api.qaf.ai is Cloudflare-proxied; confirm with one log line). Android web needs User-Agent or client-hint parsing, which nothing does today.tier, country and platform to the Langfuse trace metadata, and fix Langfuse coverage of iOS and web, so the real saving can be tracked after launch.