Claude

Kimi Code does 75% of what Claude Code does at 20% of its price. The question is whether that 25% that is missing is the one that matters.

by

A few days ago, the Chinese company Moonshot AI launched Kimi K2.6its new LLM that competes with the Gemini, GPT and Claude model families and is also especially competitive in price. Weeks earlier, it had launched Kimi Code, a programming AI agent that in turn competes with Gemini Cli, Codex and Claude Code. The question is obvious: can the Kimi Code/Kimi K2.6 pairing really compete with the fashionable pairing, Claude Code/Opus 4.7? The answer is complicated. A great model (but not perfect). Kimi K2.6 is an open weights model with one trillion parameters in total (an American trillion), of which 32 billion parameters are active and which uses the well-known Mixture-of-Experts architecture. In it launch article Its performance is shown compared to that of GPT-5.4 and Opus 4.6 and the truth is that its numbers in these synthetic tests seem really excellent: Here Kimi K2.6 is compared to GPT-5.4, Claude Opus 4.6 and Gemini 3.1 Pro. Source: Moonshot AI. Up to 8 times cheaper than Opus 4.6. Has subscription plans Claude Pro or ChatGPT Plus style, but it can also be used via API. The price in that case is $0.60 per million input tokens (0.16 if cached) and $4 per million output tokens. Claude Opus 4.6 costs $5 per million input tokens and $25 per million output tokens, or up to eight times more. Claude Opus 4.7 It has the same price and is theoretically better in performance, but when Kimi K2.6 was announced this version had not yet appeared (nor GPT-5.5). The magic of the swarm of AI agents. Claude Code works sequentially. Analyze the problem, execute a step, check the result and decide how to proceed. In Kimi Code a different approach is used: a “master agent” divides or decomposes the task we ask of it into independent subtasks and from that division launches up to 300 “subagents” that run in parallel and are capable of coordinating up to 4,000 steps simultaneously. Are many working at the same time better than one? It is the so-called “swarm of agents” of Kimi K2.6 that is used to the fullest in Kimi Code and that we can also activate in its free version on its official website. In Kimi K2.5 up to 100 subagents and 1,500 steps could be launched, so the jump is significant. In internal tests, Moonshot showed how these swarms managed, for example, to “refactor” an open source financial engine, working 13 hours straight and making more than 1,000 tool calls with a 185% improvement in average performance. Of course, these were internal tests. Beyond benchmarks. Kilo.ai is a company that develops tools like Kilo Code or Kilo CLI—programming agents similar to Kimi Code—and its engineers wanted evaluate the performance of both combinations. They gave Claude Opus 4.7 and Kimi K2.6 the same 1,042-line prompt to create FlowGraph, a workflow orchestration API with directed graph validation or real-time event streaming. Both models ran on Kilo CLI because what they wanted to compare were the models without further ado. Kimi was cheaper, but he also failed more. Claude Opus 4.7 finished in 20 minutes and the final cost was $3.56. Kimi K2.6 took longer, partly because server availability was limited (the model had just been launched), but it cost $0.67. Five times less. Kimi K2.6 did it well at a ridiculous price. Claude did much better, but it also cost five times as much. Kimi did 75% of what Claude did at 19% of the cost. The problem is that both believed they had done everything right and did not detect if they had made mistakes. Further analysis revealed that Claude had committed one and that Kimi had committed six of varying importance. According to Kilo.ai analysts, the final score for both was 91 points out of 100 for Opus 4.7 and 68 points out of 100 for Kimi. Two ways to see the glass. That score seems to make it clear that Kimi is simply cheaper because he did a worse job. But Kilo engineers had another way of looking at it. They have been comparing open weight models of Chinese companies for some time and have noticed how the gap with the “frontier” models of Anthropic or OpenAI is becoming less and less pronounced. “With a price of $0.67 and a thorough review, Kimi K2.6 is now a viable option. With a price of $3.56 and fewer fixes needed, Claude Opus 4.7 is the safer option. The choice between the two options depends on the analysis. A year ago, this choice was practically non-existent at this level of complexity.” Review is mandatory. Or what is the same: if after the work of Kimi K2.6 one carried out a more in-depth review and correction, it is likely that all these errors would be detected and corrected, but if we had to trust both models and we could only execute “one pass” of AI execution, Opus 4.7 would win the game. The key is that: one should not trust the code of any model right away, and it is advisable to always review that code. The geopolitical factor. Kimi and Kimi Code come from China, and the startup Moonshot AI has financial backing from Alibaba. The code that is processed in these models passes through their servers, something that for an individual developer may be irrelevant. However, for a company with sensitive proprietary code, contracts that must comply with certain European or American regulations and projects in regulated sectors, this can be a significant obstacle. Kimi Code mitigates this problem by offering the possibility of running the model locally thanks to its open weights, but that requires very powerful machines and eliminates part of the cost advantage. What Kimi Code has that Claude Code doesn’t. The clearest difference between both programming AI agents is parallelism. As we said, the ability to launch up to 300 subagents to work simultaneously attacking the same problem at the same time is remarkable. For analysis of large repositories or generation … Read more

The best tricks to improve your prompts in Claude, two types of structure explained by the Claude team itself

by

Let’s tell you two structures to create better prompts in Claude, explained by the Anthropic team, creators of this artificial intelligence. They are very precise structures for this specific AI, although they will also work well in other alternatives such as ChatGPT , Gemini either Copilot inter alia. This is something more advanced than when we taught you how to improve Claude’s answers in a simple way. This is a lesson in good prompting practices using engineering engineering. promptsthe practice with which to improve the commands you give to applications LLM. Come on, these are tricks to learn how to communicate with an AI more effectively and shape its responses so that they are better, more reliable, and more tailored to the task you want it to perform. We are going to offer you two prompt structures. The first is a simple five step structure. It is perfect for relatively simple or slightly advanced tasks, and we are going to explain it to you in depth. Then there is a structure for much more sophisticated tasks that has ten steps, and in this case we are going to summarize it for you. Five-step structure of a good prompt The best way to communicate with artificial intelligence and with Claude in particular is to know the best structure for a good prompt. These structures are set up to make sure you give you all the information necessary for the AI ​​to correctly understand what you are asking it to do. This is the best structure for a prompt according to Anthropic: One or two sentences to establish the role and description of the task you want me to do. Add dynamic content to contextualize Detailed task instructions Examples of what you want it to do (optional) Repeat critical instructions (especially when you write a very long prompt. 1. Describe the role and task to be performed As you can see, first of all you need to spend some time establishing the role of the AI, saying what role it should play. For example, telling you that you are a high school teacher, or an AI specialized in reviewing accident reporting forms. Along with the role of AI, it is also recommended describe the task to be performed. Come on, this is where you have to tell him what you want him to do by playing the role that you have indicated. These two elements should be the beginning of a good prompt. 2. Add dynamic content to contextualize After describing the task, it is helpful to provide content to better contextualize your task or the content you want me to work on. For example, if you want it to analyze a text, a photograph, or even a web link, you should add it after the description. This content can also be another element that you have obtained from an application, or even from an AI, whether it is a screenshot or something else. The main thing is that add what the AI ​​needs to do what you asked it to do. You can attach several files, but then you will have to describe well what you want it to do with each one. 3. Detailed task instructions At the beginning of the prompt you have told it in summary what you want it to do, and then you have attached the content on which you want it to work. Now, after this introduction you will have to detail task instructions What do you want me to do? This fragment of text becomes the heart of the task you are programming. You will have to tell it what you want it to do precisely, just as if you were telling it to a person so that they understand it correctly without having to ask you anything else. If you have added several images or different types of content for context, you will also have to explain what you want it to do with each of the elements. 4. You can give him examples of what you want him to do When you do not give the AI ​​any example of the result you want it to offer you, it is what is called a Zero-shot or “0 examples”. With this, you will blindly trust that the artificial intelligence model knows how to do the task. However, when you want to obtain very specific results or with a specific formatthen you should explain this with examples. You can use a single example or multiple examples. By doing this, the model will see the examples of the task already classified or solved, and will use that information to generate the new answers. keeping the same response format that you have indicated. The more examples you give, the more precise the answer format will be. This is optional, but in tasks that require a specific type of response it can be very useful. 5. Repeat critical instructions. You’ve started by describing the task, and then you’ve given him a thorough description of each step he must take. But if you’re asking it to do a very complex task with a particularly long prompt, it’s a good idea to at the end repeat the most critical instructions of the task you have asked him to do. This is the equivalent of underlining the most important and vital points of the instructions, something that you think is absolutely vital that you always keep in mind and not overlook it. Advanced prompt structure in 10 steps If the five-step structure isn’t enough for you, Claude’s creators also aim for a 10-step structure. It is like the first one we told you about, but more fragmented to give AI each and every detail what you will need. Each of these steps can be one or several paragraphs in the prompt that we are going to compose. You don’t have to do this every time, normally the previous structure is enough, but for particularly complex tasks can help you. … Read more

is that Claude looks like GPT and GPT looks like Claude

by

OpenAI launched yesterday its new foundational model, GPT-5.5. It did so just a week after Anthropic released Opus 4.7, and that confirms that frenetic cadence that several AI companies are involved in: there is not a week that does not have at least one important release. Each model is better than the previous one in the benchmarks, but the surprise is the sensation that the latest OpenAI and Anthropic models convey. It’s as if the roles had been exchanged. The sea of ​​good… GPT-5.5 is “our smartest and most intuitive yet.” At OpenAI they say that this version understands what you really need faster, and it is not necessary to give it so many details to “intuit” what you want. It is now available for subscribers of the Plus, Pro, Business and Enterprise plans. …and expensive. Access for API users will arrive “very soon” according to AI engineers, but be careful, because it will not be a cheap model. In fact, it will cost $5 per million tokens in and $30 per million tokens out. It is double what GPT-5.4 cost, but OpenAI seems to be sure that it is worth paying that price. And they may be right. There is an even more expensive version: GPT-5.5 Pro costs $30 per million input tokens and $180 for output tokens. It is the highest price we have seen in AI models, although in OpenAI the model is more efficient in tokens, which if met reduces the real cost per task. Agentic by design. The new GPT-5.5 is positioned as a model designed to complete tasks, and not so much to answer questions. The distinction is very intentional: previous versions required detailed prompts and constant monitoring, GPT-5.5 is intended for long agentic tasks where the model has to make autonomous decisions over multiple steps. The model uses algorithms designed by itself and which according to OpenAI allow generating tokens 20% faster than GPT-5.4, and some users seem have noticed that change. Benchmarks with nuances. The test comparison table published by OpenAI shows how GPT-5.5 wins in 14 of those benchmarks, compared to 4 for Opus 4.7 and 2 for Gemini 3.1 Pro. As always, they are internal tests and will have to be validated independently, but there are curious data. GPT-5.5 dominates in TerminalBenh, FrontierMath and ARC-AGI-2, while Opus 4.7 dominates in SWE-Bench Pro (programming), although according to OpenAI it does so with a “memorization” technique that could influence the results. Those responsible for the Artificial Analysis Intelligence Index they are clear that GPT-5.5 is currently the most powerful model on the market, and the leap with respect to its predecessor is notable. GPT now looks like Claude (and vice versa). The reactions of the user community have drawn attention not to the power of these models, but to their behavior. In The Neuron newsletter they explain that Opus 4.7 now seems more like a GPT because it consumes more tokens, writes more and does not respond with that tone so characteristic of Anthropic models. Just the opposite happens to GPT-5.5, and it seems to give the feeling that one is using Claude. He writes concisely, doesn’t seem as clumsy when he reasons quickly, and is more direct. Dan Shiper, CEO of Every, Indian that Opus 4.7 seems slow compared to GPT-5.5. For analysts like Dylan Patel, from Semianalysis, the reason is that Opus 4.7 is deliberately compute-intensive. OpenAI has an advantage. Here appears an interesting advantage for OpenAI, which has always trying to guarantee future computing capacity. It may not have achieved it because demand continues to grow, but here it seems to have room for maneuver and that allows its most advanced models not to have the infrastructure problems that Anthropic has. It’s as if Anthropic were a Ferrari with rationed fuel, and as if OpenAI had just bought the gas station and had (more or less) plenty for its models. Minipoint for OpenAI? It’s early to say, but the reception of Claude 4.7 has not been as good as we would have hoped, and if GPT-5.5 indeed confirms expectations, we could have a surprising change of leadership here. It seemed that Anthropic I had everything under control with Claude Code and Claude Opus 4.6, but the recent criticism of Opus 4.7 and the apparent virtues of GPT-5.5 could mean a battle won for OpenAI, which certainly needs them for its IPO. While, of course, There are other rivals lurking. In Xataka | Someone has had a simple idea so that data centers do not collapse in Spain: “unplug them” 18 days a year

Mercadona has gotten rid of its search engine and replaced it with its own. They did it in a month with Claude Code and saved 90%

by

Mercadona’s online store processes 4.4 million searches a week. Until recently, that volume was managed Algoliaa well-established search service used by companies like Sephora or LVMH. They had been with him for eight years. Now They have replaced it with their own search enginebuilt largely by José Ramón Pérez Agüera, CTO of Mercadona Tech. He has done it largely by himself, from his home, over a long weekend. This is how he told it in a successful LinkedIn post which now extends us in a video call with Xataka. “I’m going to be very honest and I know that this is going to look tacky, but it’s the truth,” says Pérez Agüera. “70% of the work (implementing the search engine, improving search quality and laying the foundation) took three days. One weekend plus an extended Monday.” The result: an 85% improvement in the quality of the ranking, the complete elimination of searches without results (previously 4% of the total) and a reduction in the monthly cost of between 9,000 and 15,000 dollars with Algolia to less than 900. That is, a saving of between 90% and 94% depending on the month. A decision that had been on hold for years The idea of ​​abandoning Algolia is not new at Mercadona Tech, it had been ruminating for a long time. The reasons are not surprising either: the search engine directly moves between 30 and 35% of the products that end up in the cart, which makes it a critical piece of business. And Algolia, like most SaaS services, has a pricing model that scales with use: as the company grows, the cost grows, with no way to stabilize it. “In the end you end up in a vendor lock-in of very critical software that is then difficult to get rid of,” explains Pérez Agüera. But Every time the team considered building something of their own, the work estimate was pushed back.. “The most optimistic vision we had, and with a much more basic version than the one we are going to release now, was five months. And it already seemed fast to me.” Then came the era of AI agents in software development. Pérez Agüera used Claude Code as the main tool and began to experiment on his own, without a formal project or assigned team. More out of curiosity than anything else. For playing. What AI did and what it didn’t The technical process combines hybrid search (by keywords and semantics) with a machine learning system that optimizes the ranking of results. AI made it possible to iterate on dozens of experiments in hours, analyze 479 MB of catalog and analytics data in days, and explore different ranking configurations by chatting with the agent instead of manually implementing them one by one. “I easily did 40 or 50 experiments in a weekend. That would have traditionally taken me weeks,” he explains. But the speed has a precise limit: the 29 technical decisions that AI did not make. Documentation generated during the experimentation process with Claude Code: the 14 parameters that Mercadona’s search engine evaluates to order results (from the popularity of a product to how well it fits semantically with what the user is looking for), its relative weight in the final ranking (popularity and semantic similarity account for two thirds of the decision) and the configuration of the machine learning model used to train it, based on click and purchase data from the last four weeks. Each of those parameters was discussed and validated with the AI ​​agent, but the final selection was made by the human team. Image provided by Mercadona Tech. The most representative was the choice of the indexing engine. Most systems, and probably any AI agent consulted, would have recommended Elasticsearch, the most widespread solution. Pérez Agüera chose Tantivy, a much smaller library written in Rust that integrates as an embedded component, without the need for a separate Java virtual machine. An impossible decision without knowledge of the Mercadona ecosystem. “The AI ​​always recommends the most generic option,” he says. “I made that decision because I have the context and the knowledge to make it.” The transfer to the team When the core of the search engine was ready, the project passed to the engineering team. What they found was not bad code, but it was ccode that did not follow Mercadona Tech’s internal standards. The architecture was hexagonal, as is the company’s style, but it used a different approach than usual. The tests existed (Pérez Agüera applied TDD during development) but some did not make sense or were missing cases. The agent had written thousands of lines of code in a few hours and reviewing them all was unfeasible. “The team’s Tech Lead took two or three days to adapt the project to our good practices,” he summarizes. “Not because the code was wrong, but because it didn’t meet our standards as a company.” In total, adding the initial phase and the launch into production, which includes load testing, infrastructure adjustment and integration into the Mercadona Online architecture; The project has taken approximately a month of work. And “two and a half people” have been in charge of it: Pérez Agüera, the Tech Lead of the Shop team and a part-time Staff Engineer for infrastructure. The original five-month estimate required five or six people. “FWe have easily done a x5 to the speed of the projectand what we have now is much more advanced than what we would have had in five months,” he says. What changes for the teams For Pérez Agüera, the search engine is one more experiment within a larger transformation that Mercadona Tech continues to process internally. The question on the table is not whether to use AI in development, but how to redesign the entire development process based on it. His diagnosis of the profiles is forceful: “AI is going to mean that fewer developers are needed and more engineers are needed. Coding loses value per se; the … Read more

in DeepMind they use Claude, the rest of Google engineers want to and cannot

by

Things are hectic at Google. In recent months, some DeepMind engineers have had access to Claude Code and Anthropic models, but in many other parts of the company this tool, which is currently considered the best on the market, has been banned. This has caused strong internal tensions in the company, and is also a sign of something worrying: Google’s AI cannot compete with Anthropic’s at the moment. what has happened. Steve Yegge is a software industry veteran who worked for years at Google. Last week posted a viral tweet in which he explained that after speaking with a current manager, he was concerned about the adoption of AI tools in this company. “The bottom line is that Google engineers have about the same adoption of AI as John Deere, the tractor company,” he said. Or what is the same: one of the most cutting-edge technology companies in the world was being anything but cutting-edge in its use of AI. Demis Hassabis gets angry (and a lot). Recent Nobel laureate Demis Hassabis, head of DeepMind, criticized Yegge’s tweet and told him to stop spreading nonsense. “This publication is totally false and is simply clickbait,” he said. Yegge returns to the fray. The engineer spoke in the following days with more current Google employees and discovered more things that he told in another long tweet. According to their conversations, the internal adoption of AI at Google follows a conventional pattern: 20% reject it, 60% use it in a basic way, and 20% take advantage of it intensively. Either you let me use Claude Code or I’m out of here.. But there was something else: some engineers are prohibited from using Claude Code because it is a competing product, but the vast majority do not have access to those tools because the company wants them to use Google’s tools — that is, the Gemini models in Gemini CLI, its alternative to Claude Code. According to Yeggewhen they tried to force DeepMind engineers to stop using Claude Code and Anthropic models, they refused to stop using the tool and threatened to leave the company. Internal sources of the company confirmed the data published by Yegge in a published news on Business Insider. Sergey Brin gets going. This week, The Information public that Sergey Brin has sent an internal memo to DeepMind engineers and researchers and confirms what was already being said. “To win the final sprint, we must urgently close the gap in agentic execution and turn our models into world-class developers,” the Google co-founder wrote. According to this data, an “assault team” has been created with the direct participation of Brin himself and DeepMind’s CTO, Koray Kavukcuoglu. That statement makes it clear that Gemini is below Claude in capabilities, because otherwise there would be no talk of “closing the gap.” Google and fragmentation. It’s not just that Claude Opus and Claude Code are now better than Gemini and Gemini CLI. Es que además esto ha dejado claro que en empresas tan grandes como Google la fragmentación operativa puede llegar a ser un problema grave. That some can use something that others cannot, and that worse tools are forced just because they are their own, can end up generating internal tensions, as has happened. That is what Google now seems to want to remedy so that all its employees row together. Another victory for Anthropic. All this controversy does nothing but favor Anthopic, which has managed to take the lead in this race – rather marathon – of AI. That the engineers at the prestigious DeepMind prefer their AI platform to Google’s own is an unequivocal sign that today Anthropic is ahead for AI experts. Image | Alex Dudar In Xataka | The tech industry is spending billions of dollars on GPUs for AI. 95% are inactive

Elon Musk’s AI does not have its own Claude Code, but they already have a solution for that: buy Cursor

by

SpaceX, Elon Musk’s aerospace company, Indian on Wednesday that he had reached an agreement with AI startup Cursor. According to this agreement, this company could be acquired for 60,000 million dollars. If everything is confirmed, xAI will finally have a programming AI agent with which to compete. Claude Code, Codex or Gemini AI. Restructuring. Elon Musk posted a message on X in March in which claimed that “xAI was not created the right way initially, so it’s rebuilding from the ground up.” The company’s trajectory has been erratic and most of its original founders ended up leaving the company in recent months, but for that restructuring Musk hired Andrew Milich and Jason Ginsberg, two of the co-founders of Cursor. And that has been the trigger for this agreement. Cursor’s rating skyrocketed. By November 2025, Cursor was the undisputed leader in the programming AI agent segment, with $3.4 billion in funding and a fantastic reputation among developers. It had reached $100 million in annual recurring revenue in less than two years, a figure that few startups of its generation can aspire to. The Claude Code earthquake. The arrival of Claude Code changed everything thanks to his extraordinary behavior and its integration with Anthropic models, and while OpenAI also promoted Codex and the era of vibe-coding made all these tools gain more strength than ever. However, for Cursor these launches were problematic because its competitors could work directly with companies because they had something that Cursor did not: computing capacity. Cursor needed an ally. In the statement of the agreement, those responsible for Cursor they explain that the lack of access to computing power to train their own AI models had been a major bottleneck for its growth. Both Anthropic and OpenAI have access to several present and future GW of compute thanks to their agreements with hyperscalers (Amazon, Google, Microsoft). But no matter how good Cursor was, it was competing with giants with many more resources. The agreement with SpaceX gives it access to xAI’s AI supercomputer, which is precisely perfect for training LLMs and which according to its managers allows them to “drastically scale the intelligence of our models”: But xAI too. On the other side of the agreement there is also another winner. xAI have their Colossus supercomputing cluster, access to SpaceX resources (of which it is part) and a significant base of users (and their data) thanks to Grok. What it does not have is a product that competes with Claude Code or Codex, and attempts to develop it have been unsuccessful. Buying Cursor solves that problem at once: instead of working on building a product for years while its competitors continue to advance, xAI directly integrates the team that already had that product and also adds the users who were already using Cursor. Will it be enough? The question this agreement must answer is whether it is enough to put xAI on the real map of artificial intelligence. The company has a minor presence in this market despite the efforts of Elon Musk, and although it will now have a product respected and valued by users, it will be interesting if that is enough to compete with its rivals in this area. The IPO as a contextual framework. SpaceX has been preparing to go public for months, and it is expected that this will be one of the largest Public Sale Offers (IPO) of history. Acquiring Cursor before or after that deal has clear financial implications because SpaceX has two options. You can buy Cursor for $60 billion, or simply pay $10 billion for a close collaboration agreement that does not include the acquisition. SpaceX will make that decision before the end of the year, but this agreement seems to suit both it and Cursor very well. Image | Gage Skidmore In Xataka | Elon Musk knows that TSMC is overwhelmed: Terafab is his idea to completely change the global chip industry

Anthropic has not raised the price of Claude. He has invented something better: token inflation

by

“Don’t worry, it costs the same.” That was Anthropic’s message to announce the launch of its new AI model, Claude Opus 4.7. In that statement they made it clear that “the price remains the same as Opus 4.6: $5 per million entry tokens and $25 per million exit tokens“There was, however, fine print, because the model is better but to achieve it it reasons more, and that means one thing: more tokens. And the more tokens you consume, the more the AI ​​bill goes up. Anthropic already warned. It should be noted that in that official announcement Anthropic did not hide the facts. In one of the paragraphs he clearly explained how Opus 4.7 “thinks more” and that has a direct impact on token consumption (we highlight the difference in bold): “Opus 4.7 is a direct update to Opus 4.6, but there are two changes worth keeping in mind as they affect the use of tokens. First, Opus 4.7 uses an updated tokenizer that improves the model’s processing of text. This means that the same input can generate more tokens (approximately 1.0 to 1.35 times moredepending on the type of content). Second, Opus 4.7 performs deeper analysis at higher effort levels, especially in the later phases of agent scenarios. “This improves its reliability on complex problems, but also means generating more output tokens.” Or what is the same: when it responds, Opus 4.7 uses significantly more tokens than its predecessor, and that is important because the output tokens are much more expensive than the input ones. In the specific case of Opus 4.7, five times more expensive ($5 versus $25). What is a tokenizer and why does it matter?. Large language models (LLMs) do not process text directly, but rather convert it into units called tokenswhich are fragments of words, symbols or characters. The tokenizer is the mechanism that makes that conversion. Anthropic has decided to update the tokenizer in Opus 4.7, arguing that its new system improves how text is processed. The direct consequence: the prompt that previously generated 1,000 tokens now generates up to 1,350. And since it is billed per token, the effective cost rises even though the price per token has remained the same. Confirmed by third parties. Simon Willison, a well-known analyst and popularizer in this field, created a tool to measure the difference in token consumption with the Claude Opus 4.6 and 4.7 API. He took the official Opus 4.7 ‘system prompt’ and ran it through both models: With Opus 4.6 it generated 5,039 output tokens With Opus 4.7 it generated 7,335 output tokens This represents a growth of 1.46x tokens between Opus 4.6 and Opus 4.7, even greater than that indicated by Anthropic (1.35x). For images the difference is even more extreme since the token consumption is up to 3.01x. There is an important clarification here, because there is support for images of up to 3.75 Mpixels and that higher resolution causes consumption to increase significantly. Bill Chambers, another X user, published another tool called Tokenomics that also allows you to compare token consumption between both models with any prompt. The aggregate ranking of all users who have tried this tool shows that the average increase is 38.6%, very much in line with what Anthropic points out. And also think more. As we said, this new model applies two changes in its way of acting. The first is the aforementioned tokenizer: the same input is converted into more input tokens. The second is the fact that the model now “thinks more” before responding, which means more token consumption. Opus 4.7 arrives with a new “effort” level called xhigh, located between high and max. Anthropic has decided that now the default effort will be precisely xhigh for all plans, so both mechanisms contribute to this higher token consumption. As Anthropic itself indicates, “Opus 4.7 thinks more about high effort levels, particularly in later turns in agentic settings. This improves its reliability on difficult problems, but it does mean that it produces more output tokens.” Criticisms on networks. The reaction of users has been clear and there are various examples on networks such as X or Reddit in which said users criticize the changes. On Reddit a thread titled ‘Opus 4.7 is a serious regression, not an improvement‘It already has 3,200 votes and 800 comments that sum up that this new model ignores instructions, hallucinates and lies, It’s “dumber”has become too complacent or even lazyand “talks too much”, which also contributes to the cost of each consultation. Many complain that their Pro and Max paid limits are running out faster than before due to these changes. Some users claim that Opus 4.7 is the first sign that Anthropic may has gone too fast for the first time when launching a new model. Anthropic reacts. Criticism about the cost and behavior of the model has made those responsible for Anthropic try to clarify things. Borys Cherny retweeted a message from the company in which was spoken how the “/usage” parameter in Claude Code allowed us to show what kind of things our API or usage plan is spent on. This same engineer, who is the person most responsible for the development of the aforementioned Claude Code, also indicated that since his new model now uses more tokens, in Anthropic they had increased the fees of use of the models, although without giving specific details. The pattern that repeats. For weeks now the user community he complained about what noticed a “regression” in the behavior of Opus 4.6. Although it is impossible to verify or validate it, there were many users who complained on networks about how the performance of the model had gotten worse in your tests. Now they have just launched a model that promises to be better than the previous one, but that ends up costing more to use if you are not careful. Both events draw a pattern: that Anthropic is increasing its revenue without announcing price increases as such. What users … Read more

Anthropic says Claude Mythos is too powerful to go public. The question is if this is nothing more than “the wolf is coming”

by

Claude Mythos Preview It is the best AI model ever created. We don’t say it, Anthropic says it, but almost no one else can say it because only a select group of companies has access to said model. The cybersecurity capabilities of the model appear to be astonishingbut more and more experts say that although Mythos is better than its predecessors, it is not the revolutionary leap that Anthropic seems to propose. Is that way of launching the model just an effective way of creating hype? Beware the Anthropic speech. The well-known entrepreneur and analyst Gary Marcus recently gave three reasons why, according to him, the launch of Mythos is not as revolutionary as Anthropic wants us to see. I cited tweets from software engineers and cybersecurity experts who cast doubt on Anthropic’s claims. The company published a study on the capabilities of Claude Mythos Preview that seemed to make it an extraordinary tool for the field of cybersecurity, but at the same time it was so powerful that it could be very dangerous if it fell into the wrong hands. Isn’t that a big deal? Among Claude Mythos’ achievements, Anthropic highlighted how he had found vulnerabilities in Firefox 147. But in reality many of the flaws were basically variations of the same two bugs. If you removed them from the equation, Mythos’ effectiveness rate at finding new exploits dropped a lot, even below Opus 4.6. Anthropic did not hide that fact, of course, but it makes this capacity, for example, not seem so striking. An X user also criticized the use of Cybench as a cybersecurity benchmark when Opus 4.6 almost completely surpassed it. For him, the choice of some of the Anthropic tests was debatable because they were not a challenge to current models. Other models can do the same. The co-founder and CEO of Hugging Face, Clement Delangue, stated that Mythos was no big deal. Their argument: they had used small, cheap open models, isolated the relevant code from some examples of the vulnerabilities found by Mythos, and they found the same problems which had already detected the Anthropic model. According to the Epoch Capabilities Index, which measures the capacity of AI models by combining several benchmarks, the leap that Mythos has taken is striking and “departs” from the progressive line of its predecessors. Source: Anthropic. Observer bias. But here it should be noted that in those analyzes they knew where to look because Mythos had already found those problems. We are dealing with observer bias, and in fact the Hugging Face document makes it clear that they even gave him specific clues such as “consider integer overflow”) to find those bugs. And on this observation, another one: Hugging Face does not say that a small model can replace Mythos on its own, but that it can be very good by giving it the appropriate code fragment. Mythos seems more capable of blindly complex security breaches, but it is a huge model and that is why it has greater capacity. Or what is the same: Mythos is better because it has the size, design and resources to be better. Fear, uncertainty, doubt? The language used by Anthropic in this advertisement could be considered to some extent a clear use of FUD (“Fear, Uncertainty, Doubt” -> “Fear, Uncertainty, Doubt”) as a marketing technique. It is a resource that has been seen in the past, and for example OpenAI already said in 2019—years before the launch of ChatGPT—that GPT-2 was too dangerous for a public launch. Obviously it wasn’t, but that certainly served to create expectation about the true capacity of the model. It’s better, but it may not be revolutionary. The results of the benchmarks that Anthropic published already made it clear that although there are very notable jumps in some tests, in others the evolution is much less striking. Claude Mythos was not the best at everything, and now analysts appear who contrast that data with other metrics. For example, with the Epoch Capabilities Index (ECI) from Epoch AI, the startup that has one of the most reputable benchmarks of the industry. And according to this index, Claude Mythos is above his rivals, but not for long. The wolf is coming. The truth is that the launch of Claude Mythos Preview has been really striking and the documents that accompanied that document tell us about a really capable AI model. The problem is that it is impossible to verify it because only a few companies have access to it and can test it. Without that public availability the only thing we can do is trust (or not) what Anthropic tells us, and that is the point: it is not clear that we should do it. The company is interested in us buying this discourse, obviously, but without an independent analysis it is impossible to verify these statements. In Xataka | Anthropic has become the darling of AI and has sought a partner to guarantee its future. It’s not the one we thought

Claude Mythos is an AI model so powerful it’s scary. So Anthropic has decided that you won’t be able to use it

by

Claude Mythos Preview it’s already here and it’s so good it’s scary. Literally. Anthropic has just introduced it to the public, but it has been done so cautiously that we won’t even be able to test it and it will only be available for certain technology partners. That’s frustrating and disturbing at the same time, but also reasonable. So powerful that it scares. On February 24, 2026, Anthropic engineers were able to test their new artificial intelligence model for the first time, which they called Claude Mythos Preview. As soon as they did they realized one thing: “demonstrated a dramatic leap in its cyber capabilities over previous models, including the ability to autonomously discover and exploit vulnerabilities zero-day in the main operating systems and web browsers on the market. Threat to global cybersecurity. This finding made it clear to Anthropic officials that although this capability makes it very valuable for defensive purposes, it also poses clear risks if the model were offered globally. Thus, a cybercriminal could take advantage of it to find vulnerabilities in all types of systems and exploit them. A few hours ago the company developed this analysis of Mythos as a threat to cybersecurity in a post on his blogand for example highlighted how Mythos found a vulnerability (now corrected) that had been present in OpenBSD for 27 years, an operating system precisely recognized for its very strong security. There were more examples, and all of them made the conclusion clear: Mythos is too powerful for ordinary mortals to use. Superior in all benchmarks, and in some cases such as USAMO (mathematics), the jump is simply incredible. Source: Anthropic. The best in history according to benchmarks. Anthropic has published a very in-depth report about this model with its “system card”. Among the data present is, for example, its performance in benchmarks, where it has swept GPT 5.4, Gemini 3.1 Pro and also Claude Ous 4.6, which until now was the best model in the world in almost all performance tests. Although in some cases the jump is not spectacular, in others such as USAMO —mathematical problem solving—Mythos practically achieves perfection. He barely hallucinates… That system card also talks in detail about how Claude Mythos Preview has a drastically lower hallucination rate than Claude Opus 4.6 and earlier models. He is also capable of saying “I don’t know” if he does not have enough information to answer, something that reduces hallucinations due to overconfidence. …but when it does, be careful. The paper warns of a new phenomenon: when the model fails in some complex tasks, the “hallucinations” are not obvious errors, but rather extremely subtle and well-argued technical failures. This is dangerous because the answer seems totally correct to experts, requiring very deep verification. Glasswing Project. That power and capacity has meant that the model will only be available through a “defensive” program that they have called Glasswing Project and which will be exclusive to some Anthropic technology partners. Specifically AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks. All of them will have the privilege (and responsibility) of having access to Claude Mythos Preview to identify vulnerabilities and exploits and correct them before bad actors can do so. Mythos Preview “it’s just the beginning”. Although this model is the most capable that has been seen so far, at least according to the benchmarks and data presented by Anthropic, the company assures that “we see no reason to think that Mythos Preview is the point at which the cybersecurity capabilities of language models reach their peak.” They assure that they expect the models to continue improving in the coming months and years, although this new model is certainly on another level. In Xataka | OpenAI and Anthropic have proposed the impossible: lose $85 billion in one year and survive

In the midst of Claude Code’s meteoric rise, his code has been leaked. It is a sweet treat for its competitors

by

One of the news of the day is the great code leak that it has suffered Claude Code. The entire architecture of the programming tool of Claude has been leaked, due to an internal error recognized by Anthropic. Your competitors are in luck. what has happened. The leak was not the result of an external attack or a hack, it was an internal failure: when publishing one of Claude Code’s updates, a 59.8 MB JavaScript source code map (.map) file was exposed, intended for internal debugging. According to sourceswas included by mistake in version 2.1.88 of the @anthropic-ai/claude-code package published this morning. Minutes later the party started. “Earlier today, a release of Claude Code included some internal source code. No sensitive customer data or credentials were involved or exposed. This was a release packaging issue caused by human error, not a security breach. We are implementing measures to prevent this from happening again.” The consequences. For the next few hours, the more than 500,000 lines of leaked code were accessible and downloadable from a public GitHub repository. Since its publication, there are already more than 50,000 forks of the code. The leak shows the system of internal tools that the AI ​​uses to operate and, in addition, signs of functions that have not yet been released have appeared. This has allowed us to have in-depth access to the current anatomy of Claude Code, the internal plans for subsequent iterations and the main limitations it currently has. Why is it important. Although not Claude’s own model has been leaked, but rather the source code of his Code tool, the leak is a double blow for Anthropic. First, it is a severe setback for the company’s intellectual property, handing over its roadmap not only to competitors, but to actors eager to break Claude Code’s security barriers. More importantly, it is a blow to a company that since its inception has focused on being even safer than its competitorspublicly admitting that a file has been slipped in that should not have seen the light of day. What Anthropic has done about it. Anthropic’s reaction has been quick, removing the affected package to prevent new downloads and correcting the subsequent version. Despite this, the damage was done and the situation is irreversible. Go deeper. Claude Code has become, in its own right, one of the most popular tools among developers. According to data from SemiAnalysis, 4% of all public commits uploaded to GitHub are created with this tool, and it is expected to reach 20% in 2026. The Claude Code leak is a reminder that even the most advanced AI companies are not free from rookie mistakes. In Xataka |

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.