<style>
.avi-wrap{font-family:Georgia,'Times New Roman',serif;max-width:760px;margin:0 auto;padding:0 18px;color:#1a1a1a;line-height:1.65;font-size:17px}
.avi-wrap img.avi-hero{display:block;width:100%;max-width:760px;margin:0 auto 22px;border-radius:8px}
.avi-eyebrow{font-family:Arial,Helvetica,sans-serif;font-size:12px;letter-spacing:.10em;color:#c8a24b;text-transform:uppercase;margin:0 0 6px}
.avi-dateline{font-family:Arial,Helvetica,sans-serif;font-size:13px;color:#888;margin:0 0 18px}
.avi-wrap h1{font-size:30px;color:#0f2740;margin:0 0 16px;line-height:1.25}
.avi-wrap h2{font-size:21px;color:#0f2740;border-bottom:2px solid #c8a24b;padding-bottom:6px;margin:32px 0 16px}
.avi-wrap h3{font-size:18px;color:#0f2740;margin:26px 0 10px}
.avi-box{background:#0f2740;color:#f4f6f8;border-left:4px solid #c8a24b;padding:18px 20px;border-radius:6px;margin:0 0 26px}
.avi-box strong.avi-k{color:#c8a24b}
.avi-tablewrap{overflow-x:auto;margin:0 0 28px}
.avi-wrap table{border-collapse:collapse;width:100%;font-family:Arial,Helvetica,sans-serif;font-size:13.5px}
.avi-wrap th{background:#0f2740;color:#f4f6f8;text-align:left;padding:9px 11px;border:1px solid #d9e2e6}
.avi-wrap td{padding:9px 11px;border:1px solid #d9e2e6;vertical-align:top}
.avi-wrap tr.avi-alt td{background:#f7f5f0}
.avi-wrap p{margin:0 0 16px}
.avi-lede{font-weight:bold;margin:0 0 6px!important}
.avi-cta{display:inline-block;background:#c8a24b;color:#0f2740;font-family:Arial,Helvetica,sans-serif;font-weight:bold;font-size:15px;text-decoration:none;padding:13px 24px;border-radius:6px}
.avi-faq h3{font-size:17px;margin:22px 0 8px}
.avi-rule{border:none;border-top:1px solid #d9e2e6;margin:34px 0 18px}
.avi-sources{font-size:13px;color:#666;line-height:1.55}
@media (max-width:600px){.avi-wrap{font-size:16px}.avi-wrap h1{font-size:25px}}
</style>
<div class="avi-wrap">
<img class="avi-hero" src="https://storage.googleapis.com/cio-asset-manager-eu/images/client-env-206818/01M0802D0BCZN84T4G5771E7GJ.png" alt="Avalerion Intelligence Edition 011: DeepSeek raised API prices by up to 1,100 percent">
<p class="avi-eyebrow">Avalerion Intelligence · Edition 011</p>
<p class="avi-dateline">Tuesday 18 August 2026 · AI, Data & Enterprise Transformation for the Nordics, UK & Ireland</p>
<h1>The cheap model raised its prices 1,100 percent</h1>
<div class="avi-box">
<strong class="avi-k">The One Thing.</strong> In seven days the price of AI capability moved in both directions, and every move came with a date attached. Google shipped its newest model at half price until 31 December. DeepSeek raised API rates by as much as 1,100 percent. Nvidia arranged more than 500 billion dollars of third party money against compute. And Stripe paid over 7 billion dollars for the switch that lets you change vendors.
</div>
<h2>This week at a glance</h2>
<div class="avi-tablewrap">
<table>
<thead><tr><th>Move</th><th>Who</th><th>Why it matters to you</th></tr></thead>
<tbody>
<tr><td>Gemini 3.7 Flash ships at 0.75 and 3.75 dollars per million tokens, list doubles on 1 January 2027</td><td>Google (13 August)</td><td>Your 2027 budget is being built on a 2026 promotional rate</td></tr>
<tr class="avi-alt"><td>V4-Pro launched and API prices raised by as much as 1,100 percent, effective 16 August</td><td>DeepSeek (announced 14 August)</td><td>The cheap option repriced itself, and introduced peak and off peak hours</td></tr>
<tr><td>Deal finalised to buy OpenRouter for more than 7 billion dollars</td><td>Stripe (reported 16 August)</td><td>The layer you use to move between models now has an owner</td></tr>
<tr class="avi-alt"><td>Memoranda signed with six financial firms to mobilise over 500 billion dollars</td><td>Nvidia (10 August)</td><td>Compute is now collateral, and the debt sits outside the balance sheet</td></tr>
<tr><td>Purchase commitments near 1.5 trillion dollars, lease commitments a further 1.5 trillion</td><td>Financial Times and Goldman Sachs analyses (14 August)</td><td>Your supplier's future obligations are not on the accounts you read</td></tr>
<tr class="avi-alt"><td>OpenAI frontier models embedded across IBM Consulting, with a dedicated OpenAI practice</td><td>IBM (13 August)</td><td>Your integrator's default model may no longer be a neutral choice</td></tr>
</tbody>
</table>
</div>
<h2>1. Top AI News</h2>
<p class="avi-lede">Google's best price this year expires on 31 December.</p>
<p><em>What happened:</em> on 13 August Google released Gemini 3.7 Flash, three weeks after 3.6 Flash, at an introductory rate of 0.75 dollars per million input tokens and 3.75 dollars per million output tokens. Google's published pricing lists 1.50 and 7.50 dollars from 1 January 2027, exactly double, once the introductory period ends. The model carries a one million token input context with a 64,000 token output limit, accepts text, image, video, audio and PDF, and does function calling, search as a tool and computer use. The capability jump is real: FrontierCode 1.1 rose from 34.4 to 43.6 percent, DeepSWE v1.1 from 49.0 to 65.3 percent, and AutomationBench from 17.0 to 30.4 percent against its predecessor. <em>Why it matters:</em> the number that will land in your 2027 plan is a promotional rate with a published end date, and almost nobody builds the doubling into the business case. <em>Business implications:</em> if a workload only clears its ROI threshold at 0.75 dollars, it does not clear it. It clears at half of its real price for four and a half months. The gap widens with agentic work, where a single task consumes many times the tokens of a chat turn. <em>What to do:</em> recalculate every AI business case at list price, not at the rate you were quoted, and record the vendor's own expiry date in the case. Then decide which workloads you would still run at double. Those are the ones worth building on.</p>
<p class="avi-lede">The cheap model became an expensive one, overnight.</p>
<p><em>What happened:</em> on 14 August DeepSeek launched V4-Pro and announced API price increases taking effect on 16 August, reaching as much as 1,100 percent on some rates. V4-Pro output tokens move to 3.96 dollars per million at peak, from 0.87, more than four times the previous rate. Cache miss input tokens move to 1.32 dollars per million, from 0.435. DeepSeek also introduced peak and off peak pricing, with peak defined as 01:00 to 04:00 and 06:00 to 10:00 UTC and off peak rates set at half of peak, and gave capacity strain as the reason. <em>Why it matters:</em> for eighteen months the strategic assumption behind the low cost tier was that competition only pushes prices down. A capacity constrained provider just demonstrated the other direction, and did it with two days of notice. <em>Business implications:</em> two things break at once. Any unit economics model that treated inference as an asymptotically cheaper input is now wrong, and any workload you moved onto a low cost provider to make the maths work has just had its maths redone for it. Peak and off peak pricing is the part to read twice: it converts a flat cost line into a scheduling decision, and most enterprise batch jobs run when it suits the operations team, not when tokens are cheap. <em>What to do:</em> find every workload where the provider was chosen on price and check what it costs today, not in the quote. Then ask which of your batch and asynchronous jobs could run in an off peak window. That is a scheduling change, not an architecture change, and it is the cheapest saving available to you this quarter.</p>
<p class="avi-lede">One of the largest integrators just made a model choice on your behalf.</p>
<p><em>What happened:</em> on 13 August IBM announced a strategic partnership with OpenAI to bring its frontier models and products, including GPT-5.6, Codex and ChatGPT Work, into IBM Consulting Advantage, the platform through which IBM Consulting delivers client work. IBM is creating a dedicated OpenAI practice, joining the Elite tier of the OpenAI Partner Network, and putting thousands of consultants and engineers through expert level certification. The stated targets are large scale deployments in financial services, government, telecommunications and retail, across finance, procurement, customer operations and human resources. <em>Why it matters:</em> the model layer was supposed to be the part you could swap. When your delivery partner's platform, tooling and certified people are built around one vendor's models, the swap stops being a technical decision and becomes a commercial renegotiation. <em>Business implications:</em> expect this from every large integrator this year, and note that it is not inherently bad, because depth beats breadth on delivery quality. But it changes what you are buying. A statement of work that says "we will select the appropriate model" now has a default, and the default has a price list that is not yours to negotiate. <em>What to do:</em> ask two questions in writing in your next integrator contract. Which model families is the delivery team certified and tooled for, and what does it cost, in time and money, to move a delivered workload to a different provider. If nobody can answer the second one, you have bought a model decision without pricing it.</p>
<h2>2. Enterprise AI Trend: somebody bought the switch</h2>
<p><em>What happened:</em> on 16 August Bloomberg reported that Stripe had finalised a deal to acquire OpenRouter for more than 7 billion dollars. OpenRouter is the model router: a single access point through which customers reach many different models and pick between them on capability, latency and budget. The price is a 5.4 times markup on the 1.3 billion dollar valuation OpenRouter carried at its 113 million dollar Series B in May 2026, three months earlier, and it is the largest known bet Stripe has made on AI infrastructure. <em>Why it matters:</em> multi model routing is the standard answer to vendor concentration risk, and half the CIOs in this region have a slide that says exactly that. This week the most widely used implementation of that answer was bought by a payments company, at a price that only makes sense if routing is a toll road. <em>Business implications:</em> an abstraction layer with an owner is a supplier, not an escape hatch. The independence you were buying is now subject to someone else's roadmap and pricing decisions, and a payments company has obvious reasons to care about what flows through the meter. None of that makes the product worse tomorrow. It does mean your concentration risk did not disappear, it moved one layer up, where fewer people are looking at it. <em>What to do:</em> if a gateway or router sits in your critical path, treat it as a tier one supplier this week: contract terms, exit terms, data handling, and a written answer to what you do if its pricing changes. And separate the two ideas your architecture is conflating. Being able to call many models is a capability. Being able to leave is a contract.</p>
<h2>3. Data & AI Readiness: your integrations have a shutdown calendar</h2>
<p><em>What happened:</em> yesterday, 17 August, Google shut down three Imagen 4 endpoints in the Gemini API, the standard, ultra and fast variants, deprecated on 15 June. The recommended replacement, Gemini 3.1 Flash Image, is not a drop in swap, and it is more expensive for some workloads: roughly 0.067 dollars per 1,000 images against 0.02 to 0.06 dollars before. Eight days from now, on 26 August, OpenAI retires o3 from ChatGPT at the end of a 90 day sunset announced on 28 May, with existing conversations moving to a GPT-5 model automatically; the API is on a separate track and no API retirement date has been published. <em>Why it matters:</em> nothing here is an outage or an incident. It is a calendar, published months in advance, and it will still break working systems in organisations that have no list of which model versions they depend on. <em>Business implications:</em> model IDs have quietly become a dependency class with no owner. Your CMDB tracks servers, licences and certificates with renewal dates. Almost none of them track the model endpoints your applications call, who owns each integration, and when the vendor has said it will stop answering. The failure mode is familiar: a working feature stops working on a Monday, the vendor published the date in June, and nobody was subscribed to the notice. <em>What to do:</em> build the inventory this month, and keep it to four columns: model ID, the application that calls it, the named owner, and the vendor's published end of life date. Add one control on top, a subscription to each provider's deprecation notices routed to a person rather than a shared mailbox. This is an afternoon of work and it removes a whole category of avoidable incident.</p>
<h2>4. Business Process Spotlight: compute became collateral</h2>
<p><em>What happened:</em> on 10 August Nvidia signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than 500 billion dollars of third party capital for AI compute infrastructure. The structure matters more than the number: outside investors fund data centres built on Nvidia hardware, the debt is issued privately and through special purpose entities, compute capacity serves as collateral, and most of it stays off Nvidia's own balance sheet. Jensen Huang said Nvidia may still support up to 25 percent of a given financing. In the same week, a Financial Times analysis put the combined purchase commitments of Alphabet, Microsoft, Amazon, Nvidia, Oracle and Meta at close to 1.5 trillion dollars, largely for compute, chips, capacity and energy, while Goldman Sachs analysts estimated hyperscaler lease commitments at around 1.5 trillion dollars, up from roughly 200 billion five years ago, including about 1 trillion of uncommenced leases that do not yet appear in financial statements. Goldman also expects big technology companies to fund more than a third of AI investment with debt in 2027. <em>Why it matters:</em> your supplier due diligence process reads a balance sheet. The obligations that will determine whether these suppliers hold their prices are largely not on one. <em>Business implications:</em> this is not a prediction that anything fails. It is an observation that the financing of the capability you are buying has become leveraged and harder to see, and leverage transmits. When debt service tightens it arrives at your desk as list price increases, discount programmes that quietly end, and minimum commitments attached to renewal. This week already showed two of the three. <em>What to do:</em> change one thing in your procurement template. For any AI supplier above a materiality threshold, require the commitments and lease disclosures, not just the headline financials, and have someone who reads accounts for a living look at them. Then cap annual increases in the contract, and keep at least one workload genuinely portable so the cap has teeth.</p>
<h2>5. Nordic Technology: two regimes, one group, a year apart</h2>
<p><em>What happened:</em> the AI Act's transparency duties and the Commission's enforcement powers went live on 2 August across the EU, and Ireland arrived on time: the Regulation of Artificial Intelligence Act was signed into law on 21 July, the AI Office of Ireland was established as the State's central authority with Paul Byrne appointed as its first chief executive on 30 July, and the office was operational for the 2 August date. Norway did not. As an EEA state Norway needs the AI Act incorporated through the EEA Joint Committee before its national law can take effect, that decision is still pending, the government's own target of late summer 2026 has passed, and the Act is now expected to apply in Norway from around August 2027, with a bill going to the Storting in spring 2027. Nkom, the Norwegian Communications Authority, has been designated as the national supervisory authority. <em>Why it matters:</em> almost every Nordic group of any size runs Norwegian entities alongside EU ones, and for roughly the next year the same AI system sits under two different regimes inside one organisation. <em>Business implications:</em> the tempting reading is that Norwegian operations have a year of relief, and it is wrong twice over. Your customers, your EU group functions and your enterprise buyers will apply the EU standard to the whole group regardless of where a system is hosted, and any Norwegian deployment built to the lower bar becomes remediation work in 2027, priced at 2027 rates and competing with everything else in that year's plan. Meanwhile the supervisory relationship you will eventually have is with Nkom, which is not the authority most Nordic compliance functions have been cultivating. <em>What to do:</em> set one internal standard at the EU level and apply it group wide, then keep a short register of where Norwegian law is genuinely different so you are not guessing in 2027. If you are procuring an AI system for Norwegian operations this autumn, write the AI Act obligations into the contract now. The supplier will accept them today as a differentiator. In eighteen months they will be a change request.</p>
<div class="avi-box">
<strong class="avi-k">Avalerion's Take.</strong> Every number you were quoted this week had a date on it, and most people only wrote down the number. Read the week as a price war and you will draw the wrong conclusion, because prices went up as often as down. Google's newest model is genuinely cheap, until 1 January. DeepSeek's was genuinely cheap, until Sunday. The router that was supposed to protect you from both now belongs to a payments company. The largest integrator in the room has picked a model family and is certifying thousands of people in it. And the infrastructure underneath all of it is being financed with off balance sheet debt against compute as collateral, which is a sentence that should make any CFO read a contract twice. The pattern is not that AI is getting cheaper or dearer. It is that the commercial terms of AI have become the volatile part of the system, while everyone in the market is still doing capability bake offs. You cannot forecast a price list. You can decide which of your workloads survive a doubling, which of your integrations have an owner and an end of life date, and which of your suppliers you could actually leave. Those three answers are worth more to your 2027 plan than any benchmark published this week, and all three are structural. <strong>Architecture before acceleration.</strong>
</div>
<h2>Do this next</h2>
<p>Three questions this week put on the table, thirty minutes with us: which of your AI business cases still work at list price rather than the promotional rate, whether anyone owns the inventory of model endpoints your applications depend on and their published end of life dates, and what it would actually cost you to leave your primary model provider. We will map all three with you. No pitch, no deck.</p>
<p><a class="avi-cta" href="https://www.avalerions.com/contact?utm_source=archive&utm_medium=web&utm_campaign=edition_011">Book an AI Readiness Session</a></p>
<p style="font-size:15px"><strong>Recommended reading:</strong> <em>What AI readiness actually means for Nordic companies</em> and <em>Why most enterprise AI projects still fail</em>, at <a href="https://www.avalerions.com/insights?utm_source=archive&utm_medium=web&utm_campaign=edition_011" style="color:#0f2740">avalerions.com/insights</a></p>
<hr class="avi-rule">
<div class="avi-faq">
<h2>Frequently asked questions</h2>
<h3>How much does Gemini 3.7 Flash cost, and when does the price change?</h3>
<p>Google released it on 13 August 2026 at an introductory rate of 0.75 dollars per million input tokens and 3.75 dollars per million output tokens. Google's published pricing lists 1.50 and 7.50 dollars per million from 1 January 2027, double the introductory rate. Any 2027 business case built on the launch price is understating its token cost by half.</p>
<h3>Why did DeepSeek raise its API prices in August 2026?</h3>
<p>DeepSeek announced the increases on 14 August 2026 alongside the launch of V4-Pro, effective 16 August, citing capacity strain. Some rates rose by as much as 1,100 percent. V4-Pro output moved to 3.96 dollars per million tokens at peak from 0.87, and cache miss input to 1.32 dollars from 0.435. It also introduced peak hours of 01:00 to 04:00 and 06:00 to 10:00 UTC, with off peak rates at half of peak.</p>
<h3>What does Stripe buying OpenRouter mean for multi model strategies?</h3>
<p>Bloomberg reported on 16 August 2026 that Stripe had finalised the acquisition for more than 7 billion dollars, a 5.4 times markup on the 1.3 billion dollar valuation OpenRouter carried at its May 2026 Series B. Model routing remains available, but the independence it represented now sits with a commercial owner. Treat a router in your critical path as a tier one supplier with explicit exit terms.</p>
<h3>Which AI models are being shut down in August 2026?</h3>
<p>Google shut down three Imagen 4 endpoints in the Gemini API, the standard, ultra and fast variants, on 17 August 2026, having deprecated them on 15 June; the recommended replacement, Gemini 3.1 Flash Image, is not a drop in swap and costs roughly 0.067 dollars per 1,000 images against 0.02 to 0.06 before. OpenAI retires o3 from ChatGPT on 26 August 2026 after a 90 day sunset announced on 28 May, with no API retirement date published.</p>
<h3>Does the EU AI Act apply in Norway yet?</h3>
<p>Not yet. The AI Act's transparency duties and Commission enforcement powers took effect across the EU on 2 August 2026, but Norway, as an EEA state, requires incorporation through the EEA Joint Committee first, and that decision is still pending. The government's target of late summer 2026 has passed and application in Norway is now expected from around August 2027, with a bill to the Storting in spring 2027. Nkom has been designated as the national supervisory authority. Ireland, by contrast, signed its Regulation of Artificial Intelligence Act on 21 July 2026 and had the AI Office of Ireland operational for 2 August.</p>
</div>
<hr class="avi-rule">
<p style="font-size:13px;color:#888">You're reading <strong>Avalerion Intelligence</strong>, AI, Data & Enterprise Transformation for the Nordics. Published by Avalerion Consulting AB, Malmö and Stockholm. <a href="https://www.avalerions.com/newsletter?utm_source=archive&utm_medium=web&utm_campaign=edition_011" style="color:#0f2740">Subscribe by email</a>.</p>
<p class="avi-sources"><strong>Sources for this edition:</strong> Google DeepMind and Google's published Gemini API pricing, with VentureBeat, MLQ News, DataCamp, Trending Topics and Memeburn (13 August 2026 release of Gemini 3.7 Flash, introductory 0.75 and 3.75 dollars per million tokens through 31 December 2026 and listed 1.50 and 7.50 dollars from 1 January 2027, one million token input context, 64,000 token output limit, multimodal input, function calling, search as a tool and computer use, FrontierCode 1.1 34.4 to 43.6 percent, DeepSWE v1.1 49.0 to 65.3 percent, AutomationBench 17.0 to 30.4 percent). Caixin Global (14 August 2026, DeepSeek launches V4-Pro and raises API prices by as much as 1,100 percent), with Engadget, InfoWorld and Yahoo Finance (effective 16 August, V4-Pro output to 3.96 dollars per million at peak from 0.87, cache miss input to 1.32 dollars from 0.435, peak hours 01:00 to 04:00 and 06:00 to 10:00 UTC, off peak at half of peak, capacity strain as the stated reason). IBM newsroom, 13 August 2026, with TechCrunch, ibm.com/think and The AI Insider (GPT-5.6, Codex and ChatGPT Work into IBM Consulting Advantage, dedicated OpenAI practice, Elite tier of the OpenAI Partner Network, thousands of consultants certified, target sectors financial services, government, telecommunications and retail). Bloomberg, 16 August 2026, with Fortune, SiliconANGLE and Dataconomy (Stripe finalises deal to acquire OpenRouter for over 7 billion dollars, 5.4 times markup on the 1.3 billion dollar valuation from OpenRouter's 113 million dollar Series B in May 2026). Google's Gemini API deprecations documentation, with Kingy AI, byteiota and Vorp Labs (Imagen 4 standard, ultra and fast endpoints deprecated 15 June 2026 and shut down 17 August 2026, Gemini 3.1 Flash Image as recommended replacement, not a drop in swap, roughly 0.067 dollars per 1,000 images against 0.02 to 0.06 before). OpenAI Help Center model release notes and developer deprecations pages, with gHacks and KuCoin (o3 retired from ChatGPT 26 August 2026 after a 90 day sunset announced 28 May 2026, existing conversations moved to a GPT-5 model, no API retirement date announced). CNBC, 10 August 2026, and Fortune, 11 August 2026 (Nvidia's memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than 500 billion dollars of third party capital, compute as collateral, private offerings and special purpose entity bonds, Jensen Huang on backing up to 25 percent of a given financing). CNBC, 14 August 2026, citing a Financial Times analysis putting combined purchase commitments of Alphabet, Microsoft, Amazon, Nvidia, Oracle and Meta at close to 1.5 trillion dollars and Goldman Sachs estimates of roughly 1.5 trillion dollars of hyperscaler lease commitments, up from about 200 billion five years ago, including about 1 trillion uncommenced, with Goldman Sachs on big technology funding more than a third of AI investment with debt in 2027. Ireland's Department of Enterprise, Tourism and Employment (publication of the Regulation of Artificial Intelligence Bill 2026, and the 30 July 2026 establishment of the AI Office of Ireland with Paul Byrne appointed chief executive), with William Fry and RMOK Legal on the Act signed into law on 21 July 2026. Simonsen Vogt Wiig, CMS and White & Case AI Watch on Norway (EEA Joint Committee incorporation still pending, the late summer 2026 target missed, application now expected from around August 2027, a bill to the Storting anticipated in spring 2027, Nkom designated as national supervisory authority). European Commission press release IP/26/1714 and digital-strategy.ec.europa.eu (AI Act transparency rules and enforcement powers from 2 August 2026).</p>
</div>
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [
{
"@type": "Question",
"name": "How much does Gemini 3.7 Flash cost, and when does the price change?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Google released it on 13 August 2026 at an introductory rate of 0.75 dollars per million input tokens and 3.75 dollars per million output tokens. Google's published pricing lists 1.50 and 7.50 dollars per million from 1 January 2027, double the introductory rate. Any 2027 business case built on the launch price is understating its token cost by half."
}
},
{
"@type": "Question",
"name": "Why did DeepSeek raise its API prices in August 2026?",
"acceptedAnswer": {
"@type": "Answer",
"text": "DeepSeek announced the increases on 14 August 2026 alongside the launch of V4-Pro, effective 16 August, citing capacity strain. Some rates rose by as much as 1,100 percent. V4-Pro output moved to 3.96 dollars per million tokens at peak from 0.87, and cache miss input to 1.32 dollars from 0.435. It also introduced peak hours of 01:00 to 04:00 and 06:00 to 10:00 UTC, with off peak rates at half of peak."
}
},
{
"@type": "Question",
"name": "What does Stripe buying OpenRouter mean for multi model strategies?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Bloomberg reported on 16 August 2026 that Stripe had finalised the acquisition for more than 7 billion dollars, a 5.4 times markup on the 1.3 billion dollar valuation OpenRouter carried at its May 2026 Series B. Model routing remains available, but the independence it represented now sits with a commercial owner. Treat a router in your critical path as a tier one supplier with explicit exit terms."
}
},
{
"@type": "Question",
"name": "Which AI models are being shut down in August 2026?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Google shut down three Imagen 4 endpoints in the Gemini API, the standard, ultra and fast variants, on 17 August 2026, having deprecated them on 15 June; the recommended replacement, Gemini 3.1 Flash Image, is not a drop in swap and costs roughly 0.067 dollars per 1,000 images against 0.02 to 0.06 before. OpenAI retires o3 from ChatGPT on 26 August 2026 after a 90 day sunset announced on 28 May, with no API retirement date published."
}
},
{
"@type": "Question",
"name": "Does the EU AI Act apply in Norway yet?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Not yet. The AI Act's transparency duties and Commission enforcement powers took effect across the EU on 2 August 2026, but Norway, as an EEA state, requires incorporation through the EEA Joint Committee first, and that decision is still pending. The government's target of late summer 2026 has passed and application in Norway is now expected from around August 2027, with a bill to the Storting in spring 2027. Nkom has been designated as the national supervisory authority. Ireland, by contrast, signed its Regulation of Artificial Intelligence Act on 21 July 2026 and had the AI Office of Ireland operational for 2 August."
}
}
]
}
</script>
Big Ideas,
Real Impact.
Driven by curiosity and built on purpose, this is where bold thinking meets thoughtful execution. Let’s create something meaningful together.