Localized list prices are making the list price unobservable
Geography now enters the price on four independent axes: the buyer's country, the serving region, the datacenter the workload runs in, and — newest — the country being called. Vercel Sandbox charges about 38% more in Paris and San Francisco than in Cleveland for identical compute; Apollo's dialer credits span 18x across regions with the buyer's own location playing no part.
What's happening — and why
What's happening: 'the list price' has stopped being a single observable number for a growing set of vendors, and the reasons have multiplied. The original mechanic was purchasing-power-adjusted local currency resolved client-side — Wispr Flow and Synthesia both moved India to rupee list prices on the same day, one to roughly a third of its US rate and the other onto a costlier local table.
Three more axes have since appeared. Serving region became a routing premium (Fireworks charges a flat 10% for US-only serverless). The workload's datacenter became a price input: Vercel Sandbox went multi-region in August 2026 and published a roughly 38% gap between Paris/San Francisco and Cleveland/Washington on both Active CPU and Provisioned Memory — set by a deployment setting rather than an IP lookup. And the callee's country became one too: Apollo prices dialer credits by where the call lands, from 2 credits per minute in the US to 36 in Africa, with the buyer's own location irrelevant.
The enabler for the first axis is also its problem. Client-side currency resolution means the localized price exists only in the rendered page, so no page states the country rule and a screenshot records whichever price the capture vantage happened to see.
How it works
Evidence over time
13 supporting · 6 counter — hover or tap a point for detail, click to jump to the row.
Evidence
| Company | Date | What happened |
|---|---|---|
| Clipdrop | Jun 2026 | The earliest capture in this cluster where localization erased the reference price outright. The consumer pricing page geo-localizes currency, and this capture from an Indian IP rendered Pro at ₹1,350/month while the USD/EUR headline was not displayed at all — recorded in the corpus as unknown rather than converted. The only USD figure available for comparison is an archived 2022 US render at $5/month. The corpus page still carries "₹1,350 /mo" with an "India price shown" badge, because that is genuinely all the page showed. |
| Snowflake Cortex | Jul 2026 | Geography priced as an explicit, published line rather than a rendering: AI Credits cost $2.00 per credit for global cross-region routing and $2.20 for regional (data-residency) routing — a 10% geography surcharge inside a consumption unit, edition-independent, applying across AI Functions, Cortex Search, Cortex Agents, Snowflake Intelligence, the Cortex REST API and AI Parse Doc. Warehouse compute and storage stay on edition-priced Platform Credits ($2/$3/$4). This is the discoverable end of the pattern: a buyer can read the rule. |
| Jul 2026 | A new consumer tier existed in a PPP market for three weeks before its USD price was knowable. Google AI Plus (400 GB storage, 2x usage limits, Pro-model access) was first captured on 2026-07-06 rendering only ₹399/month, with the USD figure resolved client-side and logged as unknown rather than converted. The 2026-07-28 US-region capture — the first non-India render in several capture cycles — confirmed $4.99/month and simultaneously corrected Google AI Ultra from rounded $100/$200 to exact $99.99/$199.99. Google had already introduced region-based token pricing with the Gemini 3 family on 2026-04-01, so geography sits on both its consumer seats and its developer rate card. | |
| Wispr Flow | Jul 2026 | The purchasing-power case, and the one where the rule is unobservable. The 2026-05-24 capture from this vantage rendered a single global $12/user/mo annual list; the 2026-07-22 capture from the SAME vantage renders ₹320/user/mo annual, ₹400 monthly and ₹160 for students — all at roughly 32% of the US figure at spot FX, i.e. one purchasing-power coefficient rather than a currency conversion (the savings calculator localized with it: $50/hr to ₹1,500/hr, $1,088/mo to ₹32,680/mo). The US list did not move ($12 annual / $15 monthly). No VAT or GST is added — the displayed figure is the total charged. Critically, the swap happens client-side: the served HTML still says $12, and nothing on the page states the country rule. Flow Basic stays free and Enterprise stays "Contact us" in every currency, so the geographic axis touches only the one self-serve seat. A 10-seat Indian team pays ~$38/mo where a US team pays $120/mo. |
| Synthesia | Jul 2026 | The same country, the same day, the OPPOSITE direction — this is the contradiction at the centre of the cluster. The 2026-05-31 capture served an Indian visitor Synthesia's discount-market USD table (Starter $19/month, $14/mo billed yearly, banner "plans now starting from $14/month - Save 34%"). The 2026-07-22 capture from the same location renders a rupee table instead: Starter ₹1,999/month (₹1,499/mo yearly), Creator ₹6,199 (₹4,649 yearly), banner discount cut to "Save 25%" — an effective increase on both cadences at prevailing rates. The USD list did not move (Starter $29/$18, Creator $89/$64). Two further page-level changes landed with it: cards now default to the yearly view with the monthly price struck through, and the pricing script gained a generic A/B experiment hook that can override the resolved price table per currency AND country tier — so a published Synthesia price is now market- and cohort-specific, not global. |
| Photoroom | Jul 2026 | Geography moved into the ENTITLEMENT rather than the price. In the same capture that replaced relative multipliers with absolute allowances (8,000/25,000/75,000 AI credits and 1,000/3,000/10,000 exports per month) and dropped the Free plan card, the credit FAQ added a warning that "the monthly credit amount included with your plan may vary depending on your country or region." The plan dollar fields render blank and prices are revealed only in the app checkout — so on this page neither half of the price/allowance pair is observable from the outside, and one of them is country-dependent by the vendor's own admission. |
| Fireworks AI | Jul 2026 | Geography as a priced serving path on an otherwise global rate card. Alongside Kimi K3 at $3.00/$0.30/$15.00 per 1M input/cached/output tokens (Standard) and a Fast variant at $4.50/$0.45/$22.50, Fireworks documented a US-only Serverless variant at a flat 10% premium over base serverless pricing for any model routed to US-only infrastructure — Kimi K3 US at $3.30/$0.33/$16.50. Same model, same rate card, a fourth pricing axis that is purely regional, and — unlike Wispr's — published and selectable by model ID. |
| Cursor | Aug 2026 | The eighth adopter, and the first to state a geo-differentiated ENTITLEMENT outright rather than warn that one might exist. Cursor's Models & Pricing docs now specify that its India-only Start plan (₹649) runs both Cursor Models in NON-FAST mode and fixes Grok 4.5 at a medium effort level — so the cheap regional SKU is cheaper because it is capability-limited, not because a currency converted. Photoroom's July case only warned that included credits "may vary depending on your country or region"; this names the limitation. Every other plan price held in the same capture (Hobby, Pro $20, Pro Plus $60, Ultra $200, Teams Standard $40 / Premium $120), and the two GPT-5.6 rate cuts shipped alongside it apply globally. |
| Fireworks AI | Aug 2026 | The geography premium escalated 10% to 1.5x, and moved from tokens to silicon. Fireworks was logged here on 2026-07-29 for a flat 10% premium on any model routed to US-only serverless infrastructure. Its on-demand dedicated GPU page now documents region-restricted deployments — GPUs pinned to US-only or Europe-only infrastructure — priced at a flat 1.5x the standard on-demand rate and gated behind a Contact Sales request rather than self-serve checkout. It landed in the same capture that added GB300 288 GB at $18.00/hr, the highest published on-demand rate on the card. One vendor now prices data residency two different ways at two different magnitudes: 1.1x on serverless tokens, 1.5x on dedicated GPUs. |
| Vercel | Aug 2026 | A THIRD geographic axis, distinct from the buyer-country pricing this trend was built on: the price now varies by where the WORKLOAD runs, not where the buyer sits. Vercel Sandbox expanded from one region (iad1) to four and disclosed its first regional price gap — iad1 (Washington D.C.) and cle1 (Cleveland) hold at $0.128/hr Active CPU and $0.0212/GB-hr Provisioned Memory, while cdg1 (Paris) and sfo1 (San Francisco) cost $0.177/hr and $0.0292-$0.0294/GB-hr, roughly 38% more. A US buyer and an EU buyer can pay different rates for identical work, decided by a deployment setting rather than by an IP lookup. |
| Apollo | Aug 2026 | A FOURTH axis: the price varies by the geography of the person being CALLED. Apollo cut per-minute Dialer credit costs in three regions — Middle East 35 to 28 credits/min (-20%), Africa 45 to 36 (-20%), Central & South America 15 to 14 (-6.7%) — while US/North America (2), Europe and APAC (10) held. Confirmed via Apollo's live pricing API. The buyer's own country is irrelevant; the callee's determines the burn rate, and the spread between the cheapest and most expensive region is 18x. |
| Sarvam AI | Aug 2026 | The INR-native counterexample cited in this hypothesis got more complicated rather than less. Sarvam rebuilt its marketing pricing page around a 2-card hero plus a 5-tier per-unit ladder and resolved a month-old ~7x LLM price gap with its own developer docs in the docs' favour (Sarvam-105B at INR 29.28 / 10.98 / 73.2 per 1M). It still publishes no USD table — so the "list price" for a non-Indian buyer remains unobservable — and the rebuild opened new marketing-vs-docs gaps on Translate and Dubbing. |
| Automattic | Aug 2026 | A currency-resolution failure that hid real prices, not a localization strategy. Forcing USD on cloud.jetpack.com revealed list prices for Security ($9.95/mo, $19.95 full), Growth ($9.95/$19.95) and Complete ($24.95/$49.95) plus most individual products at $4.95 first year — all of which had previously lazy-rendered to a US$0.01 placeholder for every product except CRM Entrepreneur. Client-side currency resolution did not produce a localized price here; it produced no price at all, which is the sharper form of this trend's measurement problem. |
Counterexamples
- Sarvam AI · Aug 2026 — Still not a geo case — but it stopped being an observability counterexample, which bounds this trend usefully. Sarvam is cited above as regionality-as-identity: INR-native, one price sheet, fully observable. On 2026-08-06 and again on 2026-08-14 its two first-party surfaces disagreed on the same day. docs.sarvam.ai/api-reference-docs/pricing prices Sarvam-105B at ₹29.28 input / ₹10.98 cached / ₹73.2 output per 1M tokens, split across two model IDs (`sarvam-105b` and `sarvam-105b-conversations`) and roughly 7x the ₹4 / ₹2.5 / ₹16 that sarvam.ai/api-pricing still advertises, captured the same day and byte-for-byte unchanged. The docs page also carries beta-priced Gemma-4 31B (₹36.6/₹13.73/₹91.5) and GLM 5.2 (₹128.1/₹23.79/₹402.6), plus a full Dubbing rate card (₹40/₹38/₹36 per minute by tier, doubling to ₹80/₹75/₹72 with `editor_flow: true`) live since 2026-07-30 that never reached the marketing page. Surface disagreement is a third route to an unobservable list price, alongside geography and live A/B — same country, same currency, same vendor, two prices.
- OpenAI · Jul 2026 — The market's largest anchor has not localized. The GPT-5.6 flagship line is published as a single global USD rate card — Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per 1M input/output tokens — and the ChatGPT consumer and Business tiers likewise carry one USD price each. No regional variant, no per-country table.
- Anthropic · Jul 2026 — Also unlocalized, and deliberately so at a moment it was repackaging: Claude Opus 5 shipped at the same $5/$25 per 1M rate as Opus 4.8 on one global card, and the packaging move that week was a seat-minimum cut (Claude.ai Team from 5 seats to 2), not a regional split. Where Wispr used geography to reach a smaller buyer, Anthropic used minimum quantity.
- Sarvam AI · Jul 2026 — The inverse of geo-differentiation: regionality as identity, not segmentation. Sarvam is INR-native with no USD card at all — Sarvam-30B at ₹2.5 in / ₹1.5 cached / ₹10 out per 1M, Sarvam-105B at ₹4/₹2.5/₹16, Saaras STT ₹30/hr (₹45 with diarization), prepaid ladders at ₹10,000 and ₹50,000. One price sheet, one market, fully observable — the opposite failure mode to Wispr's, and a reminder that a non-USD price is not automatically a localized one.
- Zhipu · Jul 2026 — Chose billing TERM instead of geography as its discount axis, in USD, globally: the GLM Coding Plan was repriced to $18/$72/$160 per month with a term ladder (monthly -10%, quarterly -20%, yearly -30%), up from roughly $10/$30/$80. A non-US vendor with every incentive to price by region discounting by commitment length instead.
- VEED · Jul 2026 — A global reprice with an experiment on top rather than a geographic split — and it bounds the pattern in a useful way, because it shows unobservable list prices can arrive without geography at all. VEED lifted yearly-billed USD seats for everyone (Creator $10 to $12, Pro $21 to $22, Studio $35 to $39) while running a concurrent live A/B variant at $24 for Pro. Same day as Synthesia's rupee move; different mechanism, same effect on discoverability.
Trivia
-
Wispr Flow and Synthesia repriced the same country on the same day (2026-07-22) in opposite directions, and only one of them can be right. Wispr set India at ₹320/user/mo annual against an unchanged $12 US list — about 32% of the US rate at spot FX, one purchasing-power coefficient applied to every paid figure on the page. Synthesia moved India OFF the discounted USD table it had served in May ($19/mo, $14/mo yearly, "Save 34%") onto ₹1,999/₹1,499 with the banner discount cut to "Save 25%" — an effective increase on both cadences. PPP pricing and conversion pricing are incompatible answers to the same question; the corpus currently holds one instance of each, dated the same day.
-
Google AI Plus existed as a purchasable consumer tier for three weeks before its US price was observable: first seen 2026-07-06 rendering only ₹399/month with the USD figure resolved client-side and recorded as unknown rather than converted, and not confirmed at $4.99/month until the 2026-07-28 capture — which Google's own corpus entry notes was "the first non-India render in several cycles." The same run also corrected Google AI Ultra's previously-rounded $100/$200 to exact $99.99/$199.99 list prices.
-
Snowflake Cortex (2026-07-06) publishes the corpus's only geography surcharge as an explicit number rather than a rendering: AI Credits cost $2.00 for global cross-region routing and $2.20 for regional (data-residency) routing — a disclosed 10% premium, edition-independent, on a consumption unit. Fireworks AI (2026-07-29) does the same thing on the serving path at exactly the same magnitude: routing a model to US-only infrastructure costs a flat 10% premium (Kimi K3 US $3.30/1M input versus $3.00 Standard). Both are discoverable; Wispr's is not.
For buyers
Capture the price from the geography you will actually buy and run in, and record which one that was. If you deploy multi-region, price each region separately rather than assuming the headline rate — Vercel's Paris and San Francisco sandboxes cost about 38% more than Cleveland's for identical compute, and that is a deployment setting you control. If your workload calls out to other countries, get the per-region table rather than the headline rate. And treat any price that renders client-side with suspicion: Automattic's Jetpack prices were not localized but simply absent, lazy-rendering to a US$0.01 placeholder until USD was forced.
For vendors
If you localize, state the rule. The single biggest complaint this mechanic generates is not the price difference but the invisibility — a buyer who cannot find out why they see a different number than a colleague assumes the worst. Publishing a region table costs little and removes the ambiguity. If you price by workload region, make it a first-class column on the rate card rather than a footnote, because it is a number engineers can act on directly.
Outlook — what to watch
Expect more axes rather than convergence, because each one is enabled by a different piece of infrastructure — payment localization, edge routing, multi-region compute, telephony interconnect — and none of them is being retired. Watch the frontier labs: OpenAI and Anthropic still hold single global USD rate cards, and their adoption of any regional axis would move this from a specialist mechanic to a norm. Our own price verification remains blind to axis one, which is a measurement risk as much as a market observation.
Bottom line
Geography now sets the price four different ways, and only one of them depends on where the buyer sits. For the vendors that do this, 'the list price' has stopped being a single observable number — and our own captures record whichever price the capture vantage happened to see.
FAQ
Why does an AI vendor show me a different price than a colleague abroad?
Usually client-side currency resolution: the localized price exists only in the rendered page, so no page states the country rule. Wispr Flow and Synthesia both moved India to rupee list prices on 2026-07-22 — Wispr to roughly a third of its US rate, Synthesia onto a costlier local table — so the direction is not consistent either.
Is regional pricing the same as purchasing-power pricing?
No, and conflating them is the mistake. Only one of the four axes is about the buyer. Vercel prices by the datacenter the workload runs in, Fireworks by the serving route, and Apollo by the country being called — in that last case the buyer's own location plays no part at all.
How big can the gap be?
Apollo's dialer credits span 18x, from 2 credits per minute for US and North American calls to 36 for Africa. Vercel's regional compute gap is about 38% for identical work. Both are published, which makes them the measurable end of the mechanic; the buyer-country axis is the one that is hardest to observe.
How should I record a price for a vendor that does this?
Record the price together with the vantage it was captured from — country, currency and, for infrastructure, the region. A screenshot with no vantage attached is not a price observation for these vendors, it is a sample from an unknown distribution.