Model

This is what Hollywood saturation model seeks

by

We know that feeling: Pedro Pascal seems to be appearing in all the films of the world. In all important, at least: you just saw it in ‘The Last of Us‘And you already have it in’The fantastic 4‘. And within nothing, in the movie of ‘The Mandalorian‘. And the worst thing is not that it comes out in movies and series, it is that it continually jumps on the Internet: microvideos, spots, interviews, viral. And paradoxically, guilt is not yours, but how Hollywood works today. An busy man. Let us recognize that Pascal is an authentic stove of acting, and since he rose in ‘Narcos’, he has not stopped chaining important roles: from 2019, ‘The Mandalorian’ (although he did not see his face), three series in 2021, ‘The last of us’ since 2023. And in cinema, three films in 2022, five in 2024 ‘Gladiator II‘), three in 2025 (the new of Ari Aster, the commented – Spain arrives in August -‘ materialistic ‘and the aforementioned’ the fantastic 4). And for at least 2026, he plans his interventions in the Star Wars and Marvel franchises. Why everywhere. Pascal’s career lives a “global star” effect thanks to being present in those three great franchises of the moment (Marvel, ‘Star Wars’ and ‘The Last of Us’). Thus it has become one of the most recognizable and popular faces of the industry, and it is not the result of chance, but a combination of indisputable talent and a natural charism that makes him connect with the audience both on screen and outside it (that’s why we are also seeing it in ads such as Apple’s directed by Spike Jonze). His sense of humor It has generated genuine empathy, contributing to a nearby image. Franchise world. Hollywood lives immersed in a production model that tries to reduce financial risks by force of Set off any glimpse of experimentation. The huge production costs force great studies to focus on established franchises, something that the box office supports with results as clear as last year, where The ten higher films were all sequelsremakes and derivatives. And therefore, actors like Pedro Pascal They become omnipresent: They embody characters in these universes and acquire enormous visibility by becoming key pieces. Saturation logic. The promotion of these franchises has evolved towards a model of Constant presence in all channels. Studies use aggressive strategies (Fragmented interviewsstaggered content launches, incessant activity in social networks like Tiktok, Instagram and YouTube) … what makes The public can perceive that they are receiving without stop stimuli related to franchises. The great injured are actors like Pedro Pascal, giving the feeling of appearing everywhere but unable to control that cadence. Lack of relief. Parallel to this logic of maximizing the exposure of the IP to keep it alive we have another phenomenon: there are no generational relays among the interpreters. Of course, there are new young actors, but many young actors They leave the industry due to lack of opportunities. The Post-pandemic austerity policy It prevents us from living phenomena such as those of the 80s, the 90s or the first two thousands, when they walked from young actors led each generation. As a producer affirms in This article The Hollywood Reporter, “Hollywood has spent too much time making franchises the star instead of building the next talent generation. We have not encouraged writers to create vehicles that generate stars.” The lack of risk on the part of the producers also affects this issue: they want the delivery to be full of safe values, without space for doubts. Always the same, always Pedro Pascal. Header | Disney In Xataka | The figures suggest that the domain of blockbusters in Hollywood is not over. In fact, it will go worse

The model that dominated technology for 50 years has died

by

The numbers do not usually lie, especially in the long term. And Intel’s figures during the last decade speak of a historical replication. Of A giant come less in an alarming way. In figures. In 2015, Intel’s stock capitalization multiplied by Nvidia. Almost ten years later, that of Nvidia, the first company in history to exceed four billion dollars, multiplies by 41! to Intel, which has been jibarized at a time when it has been exceeded by many lower companies. A brutal change that symbolizes the change of technological era: Nvidia’s boom first for its GPUS sales for cryptocurrency mining and then for the training of AI models. Intel’s decline due to its inconsequence in the mobile era, the rival boom in CPUS, or the passage of Microsoft or Apple to ARM. And not only Nvidia has advanced. Also Texas Instruments, Qualcomm, AMD … Gelsinger’s legacy. The contrast. In 2006, Intel refused to make chips for the iPhone Because I considered that Apple was too small. That same year, Intel tried to buy Nvidia For 20,000 million dollars, but Jensen Huang refused. Today these decisions have completely checked Intel’s route, which at the beginning of the century was always in the top 10 of world companies by stock market and today is close to leaving the Top 200. The threat. The Hard recent admission of CEO Intel, Lip-Bu Tan, that Intel is no longer among the top 10 of semiconductors, says many more things. The company that fed the PC revolution, which created the Wintel ecosystem that dominated for decades and was synonymous with American innovation, now is worth less than Spotify, Coinbase or Iberdrola. Yes, but. The 5,500 new layoffs announced by Intel They are more than an adjustment or cut. In California, almost 2,000 jobs will disappear. In Oregon, almost 3,000. In total, 20% of the workforce will be fired. The workers who developed the technology that moved the world are now expendable. The turn. Intel has outsourced TSMC 30% of its production by 2025. Paradoxical in a company that presumed its vertical integration, which controlled from design to manufacturing. And that now depends on his Taiwanese rival to survive. A symbolic surrender: the king of manufacturing asking for help to who removed the throne. Between the lines. The strategy of pivoting to the edge announced by so sounds like last minute refuge. “It’s too late for us“The CEO admitted. It is not only to recognize the loss against Nvidia, it is to recognize the loss in the most important race of what we have been from the century: that of AI. And another paradox: China represents 29% of Intel’s incomemore than any other market. The most iconic American technology company depends more on China than its own market. Trump is focused on technological independence in this legislature … … but Intel needs China to stay alive. Turning point. He 18A process That must save Intel by 2025 generates certain doubts. And if he fails, Intel runs out of his last asset to compete with TSMC. If it triumphs, it will be a partial victory, since it will be arriving five years late to the battle. This Intel fall also symbolizes something else: the end of the vertical integration model in chips represented Intel against Nvidia’s specialization model. The first designs and manufactures. The second designs but leaves manufacturing in the hands of TSMC. And hence the problem for the West: advanced manufacturing is concentrated in Asia and the value is redistributed. In Xataka | In its darkest time Intel is receiving a crucial economic support from its best client: China Outstanding image | Intel

Nintendo Switch 2 After a month of use, we tried the new Tesla Model and Juniper and much more in 1×15 crossover

by

Friday arrives and that means two things: that the weekend is just around the corner and that the New Crossover episode. As always, we are facing a very varied program in which we talk about the new Nintendo Switch 2 And our experience during this first month of life, of Peter Thiel and of the new Tesla Model Y. Videogames, PayPal Mafia And technological cars, there is nothing. First, Carlos and Jaume discover with our partner Javier Pastor the peculiar story of Peter Thiela well -known executive from Silicon Valley who co -founded Paypal and whose fortune amounts to more than 2,000 million dollars. It was also the first investors on Facebook and, today, is dedicated to the world of Big Data, Intelligence and Security in Palantir. Then Jaume tells us how his experience with the new Tesla Model and Juniper. The firm has refined the formula in its new SUV, offering better materials and finishes, an optimized dynamic and a Price/autonomy ratio that is still one of the best in the market. Of course, it still has things to improve and the competition squeezes more than ever. Finally, we review how the Nintendo Switch 2 for a montha console that has proven to be as juicy as polarizing. The new Nintendo laptop wins in potential, materials and ergonomics, but its screen, despite having won seclies, is already LCD again to have an improvable brightness level. We would also have liked to have enjoyed better autonomy. On YouTube | Crossover

Grok 4 destroys the tests and aims to be the most advanced AI model. The problem is that Elon Musk continues to sabotage his answers

by

Xai has launched Grok 4its new artificial intelligence model that is sweeping the most demanding performance tests in the sector. The model exceeds OpenAi, Google and Anthropic proposals in multiple benchmarks. And is that in the absence of knowing OpenAI’s strategy with GPT-5At the moment it has all the ballots of being the most pointer the model. However, Grok continues to drag the usual problems: controversial responses, offensive content and an Elon Musk using the model to subtract credibility. The numbers speak alone. In the test Humanity’s last examconsidered one of the most difficult to measure abilities of AI, XAI states that Grok 4 has achieved 25.4% Without additional tools, surpassing O3 of Openai (21%) and Gemini 2.5 Pro from Google (21.6%). But it is with Grok 4 Heavy, its multiagente version, where the results shoot: according to the company, reaches 44.4% With ‘tools’, almost doubleing competition. Image: XAI In addition, at the benchmark ARC-AGI-2which measures the ability to solve complex visual patterns, Grok 4 has obtained 16.2%practically double the next commercial model. According to Musk“Grok 4 exceeds the doctoral level in all subjects, without exception”, an affirmation that, although it sounds to marketing, is supported by the results obtained. The revolutionary approach. Grok 4 Heavy works with A “multiple agents” system that work in parallel about the same problem, then comparing their results as if it were a study group. This architecture allows you to climb intelligence according to the available computational power, a concept that could redefine how we understand the AI ​​performance. Image: XAI The usual problems. The launch of Grok 4 occurs just after the previous version of the chatbot published anti -Semitic comments In X, even identifying as “Mechahitler” in some answers. XAI had to temporarily withdraw the service and eliminate offensive publications, while countries like Poland They announced complaints before the European Commission and Türkiye blocked access to chatbot. The cause was a modification in the system instructions that allowed the model “not to avoid politically incorrect statements.” Although Xai withdrew that guideline, the damage was already done. A constant sabotage. Despite these technical advances, Musk continues to condition Grok’s responses in ways that compromise their usefulness. The model performs automatic searches for the opinions of the tycoon in X to respond to controversial issues, turning the alleged search for “truth” into an echo of its creator’s ideas. This practice, confirmed By experts in AI like Carlos Santana, he once again demonstrates how Musk’s controversial decisions are directly influencing the development of the model. In addition, several researchers already They have managed to avoid Easily model safety barriers, making it generate content on chemical weapons, malicious software, drugs and other sensitive issues through relatively simple jailbreak techniques. Ethan Mollick, professor at Warton and an expert at AI, Point out The lack of transparency of the company: ‘There is no detailed technical documentation, risk analysis or explanations on how to avoid future incidents’. This opacity makes it difficult for companies to trust Grok for critical applications. Grok 4 prices. XAI Grok 4 Basic offers at a price of about 30 dollars a month. However, the company has also launched Supergrok Heavy, a subscription of $ 300 per month in which its most advanced model is offered and that directly becomes the most expensive market service on the market. What is coming. XAI plans to launch a programming model in August, multimodal agents in September and video generation in October. Also will integrate Grok 4 in Tesla vehicles Next week, expanding the scope of AI throughout the Musk ecosystem. The question is whether the company will separate the technical excellence of the media controversies that surround it, or if it will remain hostage of its founder’s impulsive decisions. Cover image | XAI In Xataka | We knew that AI would generate new jobs that did not exist before. What we did not expect is that he was fixing his pifias

There is a Granada cooperative demonstrating that the opposite model to Mercadona also works: Covir will

by

The large supermarket chains have market share and penetration – or in An extraordinarily high profit margin for your sector-, but there are alternative models that without appearing at the top of those rankings, they are also successful in their own parameters. Covirán is a good case study: it is sustaining in Spain and Portugal, building an empire of proximity. Why is it important. Covirán represents the alternative approach to the dominant distribution model. Your growth strategy (rural, cooperative) goes against nature of a sector used to consolidation. This Granadina cooperative closed 2024 with revenues of 1,846 million euros (everything that is sold in all stores operating under its teaching), and with 617 million business figure of the gripo (which really invoices as a company). The negative reading is that it was just 0.13% more than in 2023. The positive reading is that the flat figure tells another story: it is changing its business model while conquering more territories. Its benefit was 2.34 million euros over 617 million. 0.38% that sounds low because it is, but that allows you to get benefit, unlike older groups such as

The Xiaomi electric SUV is already priced in China. Yu7 is a direct dart to Tesla Model and

by

Xiaomi today faces one of the most relevant events for the company in 2025. One in which they have revealed Your first smart glasseshe Xiaomi Mix Flip 2 and the price of YU7, the electric SUV that accompanies the Berlina Su7. After presenting it last May, it was only necessary to know how much it cost to access the most technological Xiaomi car to date. We already have the data on the table and, to surprise that no one is a dart for Tesla. 253,500 yuan. The price of Xiaomi Yu7 is 253,500 yuan, about 30,191 euros to change. It is an aggressive price that shoots towards the Model andcurrently with a price that starts from 263,500 in China, about 31,351 euros to change. Xiaomi Yu7: 253,500 yuan, about 30,191 euros to change. Xiaomi Yu7 Pro: 279 900 yuan, about 33,335 euros to change. Xiaomi Yu7 Max: 329,900 yuan, about 39,290 euros to change. The price is base, and varies depending on the colors we choose. In addition to the base colors (blue, green and orange), the rest adds 7,000 yuan, 832 euros to change to each purchase. The car can be purchased in China from the brand’s own app from 10:00 p.m. It is a launch that remembers more than that of a telephone than that of a car. The accessories. Xiaomi has found in the YU7 a machine to sell not only the car itself, but also its accessories. Transport Blankets Aluminum vessels Toy miniatures in different sizes Colors Carbon fiber finishes Personalized seats Aerodynamic Kits Economic base price for those who need more than the car. Personalization and accessories to the end for which he looks for a much more complete purchase experience. Why is it important. China is The country that buys more electric carsand win this market is key to a manufacturer who, since the birth of its automotive division, aspires to climb to Top 3. know-how Acquired with the Berlina Su7 to one of the most popular formats not only in China, but in the world: the SUV format. The Xiaomi Yu7 thus points directly to Tesla Model Y. The car of Elon Musk’s company It is the best selling SUV in Chinaand Xiaomi has refined specification specification to try to overcome it. They have demonstrated it with data. What does Yu7 offer. The Xiaomi SUV has a remarkably superior autonomy even in its basic version, with 835 km compared to the 593km of Tesla Model and RWD, according to the Chinese homologation cycle CLTC. The computing capacity is remarkably superior thanks to the Nvidia Thor platform, and is a much more technological car: self -adaptable suspensions, gigantic HUD with information from information from Hyperosetc. It also exceeds Model and in load speed (only 21 minutes to obtain almost 500 kilometers of autonomy), acceleration, tip and useful space, a section to which Xiaomi gives special importance to be a family SUV, bending it in Tops. The Su7 has already marked the way. Xiaomi announced its first electric car just a few months ago, with the launch of the SU7. In less than a year, he has managed to overcome sales to Tesla Model 3, placing itself as the fourth best -selling car in China at the beginning of 2025 Analyzing the global photo of sedan vehicles. At the end of 2024, the Xiaomi Su7 entered the top 20 of best -selling cars in China, with an unbeatable ByD and a Tesla with a model and in second position, above the other national models. It’s just the beginning. Xiaomi has just started selling electric cars, and has already managed to sneak into the Top 20 Chinese and Top 5 by categories. Beyond numbers, it has proven to live up (or above) of giants of the sector such as Tesla, Byd or Geely, with dynamic records in the largest dynamic benchmark to which an electric car can access: The nürburgring circuit. He already told us Abi Go in 2022: The electric car will be a fundamental piece in the Xiaomi ecosystem. And if we are clear is that there is something that the Chinese company obsesses as anyone: sell more than its rivals. Image | Xiaomi In Xataka | The Xiaomi Su7 self -imposed a challenge: to eat Porsche in his own land. And they just got it

The AI ​​race seemed to be matching. Openai has just hit the table with a model that points very high

by

The career for artificial intelligence had been narrowing for months. Gemini 2.5 Pro, Claude 4 Opus and Deepseek R1 they had managed to cut distances with OpenAi models, presenting increasingly sophisticated reasoning. The overwhelming domain of 2022, when Chatgpt He left the competition bewildered and several steps behind, no longer defines the current board. Today, the fight is direct. But the firm led by Sam Altman It has not been still. It has just announced O3-Proa new reasoning model with which, at least on paper, it intends to reaffirm its leadership position in the industry. It is available from today only for some plans, excluding Chatgpt plus of 20 dollars a month. Access to the most advanced AI has a price. More tools and more context. The O3-Pro model part of the same base as O3but with an additional stage of reinforcement training and more inference resources. It incorporates full access to the tools that have made chatgpt a more useful tool: web search, file analysis, images about Python and customization through memory. These are functions that allow you to adapt the responses to the user’s history, access external sources in real time and solve complex tasks in several steps. Has: Search, code, document analysis and vision Customization based on active memory Context window of up to 200,000 tokens and maximum output of 100,000 More precise, but not faster. Of course, O3-Pro is not a model designed for speed. The answers usually take more than in O1-Pro, but in return promises to offer a plus in precision focused on complex tasks. Openai recommends using it when reliability is more important than speed. So we could say that it is a model designed for those who need things to go well, even if that implies waiting for a few more seconds. In any case, we are talking about AI models, so hallucinations can still exist and can give us erroneous information. Theoretically superior performance. According to internal OpenAI tests, human evaluators prefer O3-Pro against O3 in all tested categories: science, education, programming, personal writing and data analysis. It also achieves better results in academic Benchmarks: in competitive mathematics it rises from 86 % of 93 % O1-PRO, and in programming exceeds 2,700 EL points, compared to the 2,517 of O3 and the 1,707 of O1-Pro. Available now, but not for everyone. O3-PRO is now available for users of the PRO PLAN (200 dollars a month) and the Team Plan, where it replaces O1-Pro in the model selector. OpenAI has confirmed that Enterprise users and educational accounts will have access from next week. At the moment, access to the most powerful models is still restricted to those who pay for the most advanced modalities. The Plus Plan of 20 dollars per month, the most popular among private users, is not mentioned. And everything indicates that, for now, it will be out of this update. Images | OpenAI In Xataka | Someone put Chatgpt, Gemini, Claude and others to play a kind of Risk. The results could not be more disparate

Ukraine has found the solution to produce three times more weapons than you can buy: the Danish model

by

It was known that Ukraine had raised a drone industry in front of the Lack of help. What was no longer known was that, since Russia launched its large -scale invasion in 2022, the Ukrainian defense industry had experienced a Explosive growthmultiplying its productive capacity by 35, from 1,000 million to 35,000 million dollars in annual potential. The problem was that, that the “potential” had no exit due to lack of financing. Until now. Lack of funds. As we said, this immense development Not only has it covered large companies, but also small workshops that operate even from particular garages, adapting to the needs of the battlefield. However, and despite this growing capacity, the country is limited by the Lack of financing: The state budget only allows to hire about 11,500 million in weapons, which leaves more than two thirds of the idle industrial potential. In fact, companies in the sector and senior officials They coincide in Insider in which, if there were the necessary funds, Ukraine could triple your production Armamentistic, thus reinforcing its resistance and reducing its dependence on foreign suppliers. Now they think they have found the solution in Denmark. The Danish model. To solve this paradox, the Ukrainian industry has proposed a small -scale solution: the call “Danish model”. This formula, applied by Denmark since 2023, allows allied countries Buy directly Weapons produced in Ukraine, instead of sending them from their own arsenals or acquiring them to Western suppliers. This way not only Reduce costs and accelerates the delivery of material, but also guarantees that the soldiers receive exactly What they needmanufactured according to its tactical demands and with an optimized logistics. Several countries have already added to this approachand others, like Germany, work in own adaptations. Thanks to this formula, Ukraine has been able to increase the production of key systems such as Bohdana self -propelled obúsand already has a list of weapons (including artillery and armored vehicles) ready to manufacture if more funds are channeled under the scheme. Collaboration and alternatives. Although the Danish model represents the road faster and fasterIt is not the only one. Serhiy Goncharov, president of the National Association of Defense Industries of Ukraine, and other leaders of the sector have also raised more complex models, based on co -production and the co -production and the Financing of Ukrainian weapons which include European components. This formula, although slower, could be attractive to a European industry that is also expanding its capacity Before the fear of A Russian expansion In the continent. In addition, the use of frozen Russian assets as a source of financing and the establishment of industrial alliances for technological transfers is considered. So much, Goncharov warns That many Western partners still do not expand their defensive capacity with the urgency that the situation requires, and that Ukraine, therefore, has no choice but to continue growing on its own to meet their own needs. Strategic opportunity Collaboration with Ukraine not only represents a way of supporting his defense in the short term, but a strategic opportunity For allied countries. The Ukrainian arms industry, born and adapted in the midst of the conflict, accumulates a direct experience in combat that is invaluable for any military power. Let’s think that your constant contact with the front allows you to adjust the design and performance of weapons based on the real conditions of the battlefield, something that many Western industries cannot replicate with the same speed. In that sense, associating with Ukraine offers Europe and other regions a double advantage: strengthening an essential supply line for Kyiv and learning from a military ecosystem that has proven extraordinarily Resilient, innovative and effective under pressure. As Goncharov concludes“We do not choose to be at war, but that is our reality,” and within it, the production capacity is ready. Image | Ministry of Defense In Xataka | We suspected that Ukraine drones attack had been destructive. Space images have revealed how much In Xataka | A technology 20 years ago took Ukraine to the Russian bombers. Moscow’s response comes from China: a laser cannon

The restoration of old photos has always been a challenge. This new AI model has just resolved it

by

The ability of AI to generate texts, images or video have proven to be amazing, but there was still a problem that These systems were resisted: the one Restore old photos to leave them perfect. That has just changed with Flux.1 Kontexta family of generative AI models with a very special characteristic: instead of following the line of the traditional text-imagen models, Flux.1 Kontext generates in-conntext images: it is possible to use as prompt a reference image and a text (prompt) so that the combination is much more powerful. You don’t need to look at the camera. The AI ​​is able to generate an image of yours looking from the original (on the left) and even a smiling. Source: Black Forest Labs. The possibilities, as explained by Black Forest Labs – the startup that develops these models – are spectacular. It is possible to maintain the context – paids, faces, objects – to modify the rest of the image at our whim: it is enough to combine those reference images with more or less detailed prompts to obtain the desired result. What Black Forest Labs engineers did not say is that this same idea could be applied to an old need: the aforementioned Old photos restoration. It is enough to load any image damaged by the passage of time or for any other reason, and Flux will do a spectacular restoration work in just a few moments. An old woman (I will not say how much) mine: if I see her from afar the result is spectacular, credible. When I do zoom, somewhat more weird things begin to appear: the face somewhat more elongated, more “square”, the nose and forehead. It’s me, but it’s not me. It is possible to test the tool in the “Recreo Patio” (Playground) of Black Forestor in services such as Fal.AI. We wanted to try it in Freepik, the Spanish generative the AI ​​platform that from the first moment has given access to previous Black Forest Labs models and of course already available Flux.1 Kontext in their two modalities, Pro and Max. Real Madrid. Commonwealth Regional Championship, October 28, 1934. Source: Institute of Cultural Heritage of Spain The results have really been striking, as can be seen in That old photo of the Real Madrid of 1934. The image headed by the article is a clear example, and here there are some details that are common in this generative model. Thus, it is possible to turn old photos into current photos that also have color and much more detail than the originals. It may even in portraits the result is similar but strange, artificial, or that traits of the person who make the result are somewhat shocking are modified. It happened for example with a Lincoln portrait obtained of the US Congress bookstore. Of course, that can end up being counterproductive. We have used a very simple prompt for this process: “Restore and Colorize This Image. Remove Any Scratches or Imperfections, Keep Fidelity”, but it is possible to adjust the results more if we introduce much more specific requests and prompts. Of course, miracles do not exist and if we zoom in the images, the problem of “hallucinations” applied to the generative image of images can arise. In the famous photograph of ‘Lunch atop a Skyscraper’ (Wikipedia) The detail of the original is scarce, so when AI tries to restore it, invents small details or features that are not. The result seen from afar, yes, is fantastic. Although Flux.1 Kontext fails to improve the definition of the image too much, sacrifices that level of detail to maintain loyalty, color and fix all the cracks of the original photo. Source: Reddit. The examples speak for themselves and demonstrate that the restoration of photos reaches thanks to this model a really spectacular quality, as happens in the photo of that family (Reddit). To get an idea of ​​the cost, each image restored with Flux Pro in Freepick costs 100 credits, and with a monthly subscription at its premium service (10 euros/month) 18,000 credits are achieved. Thus, 180 of these restorations for 10 euros, less than six cents per image. Other Black Forest Labs models are available in Github And they can be installed and used at home – ideal if you have a good GPU – but for now Flux.1 Kontext is not included in that catalog. We do not know if they will add it soon, but for the moment it is possible to experience for free with the model in services such as those mentioned. Without a doubt, a great option to perform this type of task. In Xataka | The 1900 Paris, in 4K all color: the restored videos that transport you to the 19th century

There is a career to achieve the cheapest and most powerful model. And against all prognosis, China is winning it

by

On May 20 Google launched Gemini 2.5 Pro and Gemini 2.5 Flash in preliminary version. These new AI models were better than ever, and to demonstrate the company included in its announcement several graphs and comparative tables. They looked at how both surpassed their rivals both in the field of reasoning and in the traditional performance (Benchmarks of Mathematics or Programming), but there was also another fact that Google presumed: Google: The cost of Gemini 2.5 Flash. Source: Google. That table published by Google made it clear that Gemini 2.5 Flash It was clearly the winner of that comparison in the important price/benefits ratio. What Google did not say is that this success of this model was the exception to the rule, because in that race for having cheap and powerful models, China seems to take the lead. He does it at least if we attend to the cost of using these models. In Xataka we have analyzed that cost based not on the price of subscriptions for end users, but on the cost of access to the API, which is the one that allows developers to integrate these models into their own chatbots and their services. The API prices of each model clearly differentiate two uses from artificial intelligence. On the one hand, how much does it cost to write something for the model to then process it (the so -called input tokens). On the other, how much does the text generated by the model once processed the answer (the so -called output tokens). The entrance tokens They are usually five times cheaper than exitbecause processing the request and generating text is much more expensive than receiving it, analyzing and “understanding it.” We wanted to compare the cost of the main models of the AI ​​developed in China and those of the US, and although as always are not all that are, if they are all. The resulting table is as follows: These prices are public and very easy to find in the case of US AI models (OpenAI, Anthropic, Google) but not so much in the case of Chinese models (Deepseek, Qwen (Alibaba), Doubao (Bytedance), GLM-4 (Zhipu), Ernie (Baidu)). Be that as it may, the table, ordered from the cheapest to the most expensive, demonstrates that today Chinese models are especially cheap. Only Gemini 2.5 Flash Preview manages to compete – and does it exceptionally. In the rest of the cases, the AI ​​models in China win the battle for cost. It must be noted that, Like all comparisons, this is unfair. And it is because that table does not take into account the benefits of each model. Openai O3 and Anthropic Claude Opus 4, the last and most powerful models of those companies, are especially precise in their answers, but each consultation consumes much more resources (computation, energy) and that makes it logical that they are much more expensive than their competitors. But these models are also designed for very special cases and for specialized, detailed and deep consultations. In the vast majority of cases it is not necessary to use these models, and that is where they are competing for example Deepseek R1 or Gemini 2.5 Flash Preview: in the price/benefits ratio. Models with variable prices That price battle has made us see in recent times two techniques that some companies are applying to the prices of use of their APIs. The first one is to differentiate normal entries and outputs of inputs (and even exits) cache. Deepseek API prices. Attentive to the lower left: according to the time you use them, they can leave cheaper. Source: Deepseek. The explanation is simple: a “normal” entry is a request or question that the model has never processed and therefore has to process completely. If the entrance has been caught (Cache hit) is because that request has been processed in the past, so the system can obtain the response of its cache, which significantly reduces the computational costs. Deepseek, Google, Anthropic and Openai offer this type of option, as can be seen in the table. The second technique is to use variable prices according to (at the moment) the time slot in which we use these models. This is what Deepseek has done, which has “day” and “night” prices according to UTC schedule. If you use the Depseek API from 18:30 to 2:30 (Peninsular schedule in Spain), it will be half price. Good news: AI is every time (much) cheaper While China and the US fight who has the most powerful model or who has the cheapest model, what is constantly happening is that the AI ​​price is falling remarkably. It is an observation that several experts such as Ethan Mollick, a professor at the University of Pennsylvania who recently analyzed how That price/benefits ratio does nothing but improve. The models are getting better and cheaper. Other experts such as Raveesh Bhalla – explained by Netflix and LinkedIn – also reflected this evolution at the beginning of the year. Then he showed how the cost of an O1 level model had dropped 27 times in the last three months. Moreover, at this rhythm the GPT4 level models – which a year ago were absolute referents – will be reduced 1,000 times in just 18 months. We are living it in price reduction. Dane Bahey, from Openai, said at a conference in September last year how the cost per million tokens had fallen from 36 dollars At just 0.25 dollars In the last 18 months. And that price drop is still clear and fantastic for users. Thus, we are facing a career that at the moment has a lot of stretch: China’s models carry the lead if we attend only to their cost, but careful, because we must also take into account the benefits. It is true that these Chinese models have already shown in the benchmarks that have been showing that they compete from you to you with the best US models, and now it remains to … Read more

Log In

Forgot password?

Forgot password?

Enter your account data and we will send you a link to reset your password.

Your password reset link appears to be invalid or expired.

Log in

Privacy Policy

Add to Collection

No Collections

Here you'll find all collections you've created before.