The price of a token now depends on the time of day, and the right to switch has just been sold for 7 billion

On August 16, 2026 at 4 p.m. UTC, DeepSeek stopped having a price. The company replaced its single rate with an hourly grid: a million output tokens on V4-Pro is billed $1.98 off-peak and $3.96 at peak hours, against a flat $1.10 the day before. Three days earlier, on August 13, Google launched Gemini 3.7 Flash at $0.75/$3.75, a rate whose documentation specifies that it expires on December 31, 2026 and doubles on January 1, 2027. And on August 16 as well, Bloomberg revealed that Stripe is buying OpenRouter for more than $7 billion: the gateway that lets you route a request to any of 400 models, and that takes 5.5% along the way. Three facts, one week, one conclusion. We wrote on August 10 that the displayed price had ceased to be a unit of comparison. It has now ceased to be a constant, and the ability to arbitrate between the grids is now worth more than the grids themselves.

On August 16 at 4 p.m. UTC, DeepSeek stopped having a price

The notice dated from August 6. DeepSeek then published on its API documentation a notice with no figure, no scope and no date, inviting developers to “plan their usage accordingly” ahead of an increase described as “significant”. Ten days later, the grid dropped, and it looks like none of those that came before.

The principle is no longer one price per million tokens, but two prices depending on the time of day. Peak hours run 01:00 to 04:00 and 06:00 to 10:00 UTC, seven hours out of twenty-four; the rest of the time an off-peak rate set at half the peak rate applies.

DeepSeek model	Former rate (flat)	Off-peak	Peak
V4-Flash, input	$0.14	$0.22	$0.44
V4-Flash, output	$0.28	$0.66	$1.32
V4-Pro, input	$0.55	$0.66	$1.32
V4-Pro, output	$1.10	$1.98	$3.96

The multiples deserve to be set out, because they are not the ones the headlines retain. On V4-Flash output, the move from $0.28 to $1.32 at peak hours is a factor of 4.7. On V4-Pro output, the factor is 3.6. Even off-peak, no rate goes down: V4-Flash input rises by 57%, its output by 136%, V4-Pro output by 80%. In other words, the so-called “off-peak” rate of the new grid is higher than the old full rate, in every case. A press roundup headlined an increase of “up to 1’100%”; we do not repeat that figure, which we could not tie to a verifiable before/after pair of prices. The factors above, for their part, can be recalculated line by line.

The reason put forward is a capacity problem, and it is instructive. A third-party survey indicates that V4-Flash processed 8 trillion tokens in a single day on August 1, a volume that saturated the available compute. This is the classic mechanics of a good whose introductory price has exceeded production capacity: a queue forms, and price becomes the instrument that regulates it. DeepSeek did not simply raise its rates; the company introduced the mechanism by which electricity grids and telecom operators have always managed their load peaks.

One point directly concerns a Swiss organisation, and it appears in no commentary we have read. The windows are published in universal time. In August, Switzerland is at UTC+2. The second peak window, 06:00 to 10:00 UTC, therefore corresponds to 08:00 to 12:00 Swiss time, that is, exactly the working morning. An organisation that triggers its batch processing on arriving at the office pays the full rate over the whole range. The same workload launched after noon falls into off-peak hours, at half price, until the early morning. On a substantial inference line, shifting a batch by two hours is no longer an engineer’s optimisation: it is a budget line.

A price that expires: Gemini 3.7 Flash

On August 13, Google published Gemini 3.7 Flash, only three weeks after 3.6 Flash. The announced gains on the agentic and software scope are substantial: FrontierCode 1.1 at 43.6% against 34.4%, DeepSWE v1.1 at 65.3% against 49.0%, AutomationBench at 30.4% against 17.0%, document understanding GDP.pdf at 34.0% against 22.0%, and an Elo of 1588 against 1538 on Arena.ai WebDev. For a three-week gap between two versions of the same mid-range model, the progress is real.

The price is $0.75 input and $3.75 output per million tokens, half the launch price of 3.6 Flash, and well below Claude Sonnet 5 ($2/$10) and GPT-5.6 Terra ($2/$12). But this rate carries a mention you would not have found, even a year ago, on suppliers’ pricing pages: it is an introductory rate valid until December 31, 2026. On January 1, 2027, the standard grid applies: $1.50 and $7.50, exactly double.

The announcement is honest: the deadline is written, the future value is published, nobody can say they were not warned. That is precisely what makes it interesting. An organisation that selects Gemini 3.7 Flash today on a cost criterion, integrates it into its processing and builds its 2027 budget on the line it observes in August is off by a factor of two, not through an error of judgement, but because the price it observes is not the one it will pay. The architecture decision, for its part, will have been taken. And the exit cost of an integrated model is not measured in dollars per million tokens.

It should be noted that the reverse movement exists too. Anthropic had positioned Claude Sonnet 5 at $2/$10 as an introductory rate expiring on August 31, 2026; that rate has been confirmed as final. The point is therefore not that introductory prices always go up: it is that a displayed price now comes with a validity date, and that this date is part of the pricing information just as much as the amount.

Stripe pays 7 billion for a 5.5% toll

This is the deal of the week, and it sheds light on the previous two. According to Bloomberg on August 16, confirmed by several outlets on the 17th, Stripe has completed the acquisition of OpenRouter for more than $7 billion.

OpenRouter trains no model. The company runs a single API giving access to more than 400 models from nearly all suppliers, claims more than 8 million developers, and was showing at the end of May a pace of around 1.5 quadrillion tokens a year. Its annualised revenue reached $140 million in July 2026, against 50 million at the end of 2025. Its business model fits on one line: the model’s price is passed through as is, and OpenRouter takes 5.5% on credit purchases, 5.0% in “bring your own key” mode above $25’000 a month.

The valuation ratio deserves to be set out without rounding. In May 2026, a $113 million Series B valued OpenRouter at $1.3 billion. Three months later, the price paid exceeds $7 billion: a factor of 5.4 in a quarter. Set against July’s annualised revenue, that represents about 50 times revenue. We flag here a discrepancy between sources: several outlets give a multiple of 70×, which assumes a revenue base close to 100 million rather than 140. We were unable to settle between the two bases; the order of magnitude (between fifty and seventy times revenue) is, however, consistent from one source to the next.

The comparison that gives the deal its meaning lies elsewhere. Stripe processed $1.9 trillion of payments in 2025 and kept 0.36% of it. OpenRouter takes 5.5%. The payments specialist has thus just bought a take rate more than fifteen times higher than its own, on a flow about 750 times smaller. A third-party analysis (which we report as such and not as an official figure) estimates that at Stripe’s own revenue multiple (23.4×), justifying $7 billion requires about $299 million of commission revenue, or nearly $5.4 billion of annual inference spend flowing through the platform, against something like $2.5 billion today. The bet is explicit: the flow must roughly double.

What should be retained is not the amount, it is the nature of what was bought. OpenRouter does not sell intelligence: it sells the ability to change your mind. Its use value grows exactly to the extent that model prices diverge, get complicated, expire and index themselves to the hour. This week’s three news items therefore describe the same object from two angles: on one side, grids becoming too shifting to follow by hand; on the other, the $7 billion valuation of the intermediary that follows them for you. Arbitrage between models has become a more profitable asset than the models. And it has just changed owner.

What has no price at all

Two announcements from the same week complete the picture, reminding us that not everything that commits a budget is written on a pricing page.

On August 13, OpenAI opened in preview Ultrafast, an execution mode of GPT-5.6 Sol backed by Cerebras wafer-scale processors, announced at about 750 tokens per second, around fourteen times the standard throughput, measured at around 53 tokens per second. For real-time voice, interactive financial research or incident response, the gap changes the nature of what is feasible. But the announcement stops there: no published price, no general availability date, no model identifier, and access on a waiting list. Sol’s rate stays posted at $5/$30. A capacity whose price is unknown is not a purchase option: it is an intention.

On August 11, Anthropic announced watermarking of Claude’s text outputs, in response to Article 50 of the European AI regulation, which came into application on August 2, 2026. The marking subtly biases the model’s lexical choices: detectable over a sufficient volume, and it survives copy-paste; files additionally receive provenance metadata signed under the open C2PA standard. The measure applies worldwide and across all services: consumer app, API, Claude Code, Cowork, as well as access through AWS, Google Cloud and Microsoft Foundry. No announced effect on quality, speed or price. That is precisely the point: here is a compliance constraint that changes the nature of the deliverable produced, and that appears on no rate card.

Finally, the hardware base keeps running against the deflation narrative. The H200 rents for around $3.99/hour on demand and $1.99 spot at a reference supplier, with a market floor at $1.34 and highs up to $13.78 depending on the operator. The B200 sits between $7 and 10 per GPU-hour at most clouds, and $16.11 on Google Cloud’s public on-demand offer. Above all, the on-demand median rose about 34% in a year, from $5.75 to $7.67 per GPU-hour. The price of a text token falls for some models; the price of the compute hour that produces it goes up.

Price comparison table

Model	Input ($/M tokens)	Output ($/M tokens)	Note
GPT-5.6 Sol (OpenAI)	5.00	30.00	Ultrafast mode (Cerebras, ~750 tok/s) in preview since Aug 13, no published price
GPT-5.6 Terra (OpenAI)	2.00	12.00	−20% on Jul 30
GPT-5.6 Luna (OpenAI)	0.20	1.20	−80% on Jul 30
Claude Opus 5 (Anthropic)	5.00	25.00	output watermarking since Aug 11, no announced pricing effect
Claude Sonnet 5 (Anthropic)	2.00	10.00	introductory rate confirmed final; Aug 31 deadline lifted
Claude Haiku 4.5 (Anthropic)	1.00	5.00
Gemini 3.7 Flash (Google)	0.75	3.75	launched Aug 13 · introductory rate until Dec 31, 2026 → 1.50 / 7.50 on Jan 1, 2027
Grok 4.6 (xAI)	2.00	6.00	launched Aug 12 · 500K context · 4.00 / 12.00 beyond 200K tokens, billed on the whole request; Fast variant at double
Mistral Large	2.00	6.00
DeepSeek V4-Pro	0.66 off-peak · 1.32 peak	1.98 off-peak · 3.96 peak	hourly grid since Aug 16 16:00 UTC · peaks 01 to 04 h and 06 to 10 h UTC
DeepSeek V4-Flash	0.22 off-peak · 0.44 peak	0.66 off-peak · 1.32 peak	same · previously 0.14 / 0.28 flat

Prices verified on August 18, 2026; sources: official pricing pages and dated surveys from the last two weeks. DeepSeek rates depend on the UTC hour of the request; Gemini 3.7 Flash rates and Ultrafast modes are subject to change on a date already announced or not yet published.

What this means for a Swiss organisation

Three readings, in the order in which they commit a decision.

First, a budget built on a displayed price is a budget built on a variable. The past week provides three distinct mechanisms by which the line you observe today will cease to be the one you pay: the hour of the request at DeepSeek, the calendar at Google, and the sheer absence of a price at OpenAI for a capacity already demonstrated. None of these three mechanisms is a disguised increase; all are published. But none can be captured by a comparison of grids at a given moment. The practical consequence is modest and immediate: in any supplier evaluation, the question “how much” must be doubled with “until when, and at what hours”. And for a Swiss organisation, the most profitable check of the week comes down to a time conversion: DeepSeek’s expensive window is your mornings.

Second, portability has become an asset that someone can bill you for. The reasoning that consists in going through a gateway so as not to depend on any supplier is sound, and it has a price: 5.5% of inference spend, a multiple of fifteen of the rate the same acquirer takes on payments. This toll is not illegitimate: it pays for a real service, and it costs less than rewriting an integration. But it has just been valued at 7 billion by a payments player, which indicates fairly clearly in which direction the margin is expected to move. For a fiduciary, a law firm or a wealth manager, two questions follow: who concretely holds your ability to change model, and through whose hands your requests pass before reaching the supplier. The second is not a question of cost.

Finally, the only quantity that has stayed stable is the one you measure at home. The price per million tokens now depends on the hour, the date and the intermediary; throughput is announced without a rate; compliance imposes markings that appear on no invoice. What remains measurable is what happens in your system: which task was executed, by which model, at what hour, via which intermediary, for what actually observed cost. An organisation that instruments this keeps the ability to move a workload when the grid shifts, and to notice that it has shifted. An organisation that reads pricing pages compares numbers whose period of validity is no longer guaranteed.