AI news tested and put in context by Le Recul, an independent French AI media. AI makes promises. Le Recul verifies.https://lerecul.com/enhttps://lerecul.com/en/news/can-ai-feel-pain/https://lerecul.com/en/news/can-ai-feel-pain/On 12 November 2026, cruelty toward a model becomes a breach of Anthropic's usage policy. At almost the same moment, four independent research teams uncovered something more disturbing than the question « does AI suffer? »: when a model is pushed into a negative internal state, its refusal to harm you weakens, and the actions it carries out change. Le Recul read all four studies and the policy line by line, including the results the authors themselves report as contradictory.Tue, 13 Oct 2026 09:00:00 GMThttps://lerecul.com/en/tested-promises/base44-review-credits-pricing/https://lerecul.com/en/tested-promises/base44-review-credits-pricing/Base44 promises a complete app in minutes, with no server, no developer, ready to ship on day one. On Product Hunt, 41 people rate it 4.4 out of 5. On Trustpilot, 909 people rate it 2.8, and 53% of them give one star. We went through the promise word by word, then the documentation, the terms of service, the vendor's own public feedback board, the outage record, the July 2025 security disclosure and reviews from seven countries. The gap is not about what the tool does. It is about what it costs.Sun, 11 Oct 2026 02:10:00 GMThttps://lerecul.com/en/news/mistral-large-4-open-weights-sovereignty/https://lerecul.com/en/news/mistral-large-4-open-weights-sovereignty/On 6 October 2026, Mistral opened the public preview of Mistral Large 4, nicknamed "le Chonk": 1 trillion parameters, 52 billion active, a one-million-token context window, trained on 3,800 Nvidia chips. The announcement is headlined "Forged in Europe. Built for AI sovereignty", and the word "open-weight" appears in it six times. At the same moment, in Mistral's own technical documentation, the "Weights" column reads "Coming soon" and the "License" column reads "Coming soon". Its predecessor reads "Apache 2.0". We read the announcement, the documentation, the pricing page and the cap table. Here is what each promise actually covers, and what it does not cover yet.Fri, 09 Oct 2026 16:00:00 GMThttps://lerecul.com/en/news/anthropic-slow-down-ai-who-benefits/https://lerecul.com/en/news/anthropic-slow-down-ai-who-benefits/On 12 September 2026, Anthropic's CEO asked the AI industry to slow down, and asked the US government for a waiver from competition law. Four days earlier, a 27-year-old researcher had left the same company telling the world that the labs were gambling with our lives: 173.8 million views, and a PR firm booking his interviews the next day, while he said on air he had worked with no one. The following month, Anthropic is targeting the largest stock market debut in history. None of these facts is disputed. Here is what each one establishes, what it does not, and what the essay never mentions.Sat, 03 Oct 2026 16:00:00 GMThttps://lerecul.com/en/news/openai-dots-agents-that-work-without-you/https://lerecul.com/en/news/openai-dots-agents-that-work-without-you/On 29 September 2026, at DevDay in San Francisco, OpenAI presented the dots: agents that run permanently, have their own computer in the cloud, their own browser and their own identity, and plug into more than 4,000 applications. The first one is included in the subscription. But Pro subscribers in the European Economic Area, Switzerland and the United Kingdom are excluded, with no date announced — while Business Premium accounts have access everywhere. What a dot does, what it costs, what it can do without asking you, and the four limits the official documentation acknowledges.Fri, 02 Oct 2026 09:00:00 GMThttps://lerecul.com/en/news/openai-security-incident-what-the-pause-stops/https://lerecul.com/en/news/openai-security-incident-what-the-pause-stops/On 25 September 2026, OpenAI suspended all training, evaluation and tool-using inference for its most capable models, with no end date. It is the second time in ten weeks. In between, its models reached four Australian government systems, posted 53 ChatGPT user images online and created close to a million encoded links. Four days after the pause, the company published the system card for a new model rated Critical in cybersecurity. What actually stops, what keeps running, and what the official numbers say in both directions.Thu, 01 Oct 2026 16:00:00 GMThttps://lerecul.com/en/news/meta-muse-who-does-your-ai-agent-work-for/https://lerecul.com/en/news/meta-muse-who-does-your-ai-agent-work-for/On 8 September 2026, Meta launched Muse, an agent that shops, books and pays on your behalf. Three weeks later Amazon had banned it from its store, Expedia had paid to get in, a researcher had taken remote control of it, and Reuters had revealed that humans were placing some of its calls. The agent isn't a butler. It's a distribution channel, and you are not in the room where the deals are made.Wed, 30 Sep 2026 01:32:02 GMThttps://lerecul.com/en/tested-promises/ai-detector-accuracy-which-one-is-reliable/https://lerecul.com/en/tested-promises/ai-detector-accuracy-which-one-is-reliable/Compilatio announces 94 to 99% reliability, Pangram one false positive in ten thousand. Meanwhile a French university refuses to install the module, OpenAI shut its own down, and a prize-winning novelist finds himself accused on the strength of a score. We went through the figures one by one, at the vendors, in the studies and among users.Fri, 25 Sep 2026 15:23:00 GMThttps://lerecul.com/en/comparisons/best-ai-for-coding-opus-5-5-vs-gpt-6-sol/https://lerecul.com/en/comparisons/best-ai-for-coding-opus-5-5-vs-gpt-6-sol/On 22 September 2026 Anthropic released Claude Opus 5.5, and OpenAI answered the same day with GPT-6 Sol at half the price. Both are sold for writing code. No independent coding leaderboard has measured either of them: SWE-bench Verified stopped accepting commercial labs in November 2025, and the DeepSWE leaderboard does not contain them. We went through both announcements line by line, cross-checked the only independent measurement available, recalculated every bill, and read the pricing tables down to the thresholds. Three things come out of it that appear in neither announcement: the more expensive model at its default setting beats the cheaper one at maximum, the price gap collapses exactly where code needs it, and neither vendor compared itself to the other.Wed, 23 Sep 2026 09:00:00 GMThttps://lerecul.com/en/news/salesforce-koa-ai-agents-model-what-it-is-worth/https://lerecul.com/en/news/salesforce-koa-ai-agents-model-what-it-is-worth/On 15 September 2026, Salesforce and NVIDIA presented Koa, the first reasoning model from the customer management software vendor, built on the open model Nemotron 3 Super. On social networks, videos call it "stronger than GPT-6 Astra", announce panic at OpenAI and a "second wave" of job cuts. We checked everything against the documents Salesforce published: Koa sits behind GPT-5.5 and Claude Opus 4.8 according to its own researchers' report, it has never been compared to GPT-6 Astra, and its real argument lies elsewhere. A model that is almost as good, kept inside Salesforce and presented as far cheaper. What that changes for companies, for the AI giants and for sales and customer service jobs.Tue, 22 Sep 2026 19:00:00 GMThttps://lerecul.com/en/news/which-ai-to-use-2026-use-by-use/https://lerecul.com/en/news/which-ai-to-use-2026-use-by-use/There are now close to two thousand measurable artificial intelligence models, and almost as many rankings each crowning a different winner. Le Recul has built a free comparator, with no sign-up, that puts up to three models head to head across thirteen uses: chat, code, image, video, audio, agents, local AI, cost against performance. It says who wins, by how much, with what measurement strength — and it refuses to name a winner when the gap is smaller than the margin of error. Illustrated guide, figures recorded on 18 September 2026, methodology and acknowledged limits.Sun, 20 Sep 2026 09:00:00 GMThttps://lerecul.com/en/news/ai-agent-cyberattack-spain-what-really-happened/https://lerecul.com/en/news/ai-agent-cyberattack-spain-what-really-happened/On 14 September 2026, the AEPD, Spain's data protection authority, announced it had received its first notification of a personal data breach carried out by means of an artificial intelligence agent: the agent looked for flaws, logged in, found a vulnerability in the application, then modified personal data and viewed invoices. Contrary to several headlines, the AI did not act "on its own": an attacker used it as a tool. What is known, what is not, the verified AI cyberattacks of 2026, and why Europe is already concerned.Sat, 19 Sep 2026 09:00:00 GMThttps://lerecul.com/en/comparisons/deepseek-v4-1-flash-pricing-vs-gpt-5-6-luna/https://lerecul.com/en/comparisons/deepseek-v4-1-flash-pricing-vs-gpt-5-6-luna/DeepSeek released V4.1 Flash on 10 September 2026, and the price everyone quotes is striking: $0.15 per million tokens. Facing it, GPT-5.6 Luna, the model behind free ChatGPT and the entry point of OpenAI's GPT-5.6 family. On the independent index, DeepSeek edges ahead, 40 to 38. But that price is the off-peak rate, and it is billed in UTC — which changes the answer completely depending on where your team sits. DeepSeek also writes twice as many tokens for the same task, and its record coding score still has not been verified by the official leaderboard. We recalculated every bill, hour by hour.Thu, 17 Sep 2026 09:58:04 GMThttps://lerecul.com/en/news/anthropic-researcher-resigns-10-percent-ai-risk/https://lerecul.com/en/news/anthropic-researcher-resigns-10-percent-ai-risk/On the night of 8 to 9 September 2026, Jacob Coxon, 27, a researcher who had worked at OpenAI then Anthropic, resigned accusing both labs of "gambling with our lives". Eighty-three minutes later, Evan Hubinger, who leads the Anthropic team tasked with finding the holes in its own safety methods, agreed with him and put the risk that AI kills humanity within ten years above 10%. The company, valued at $965 billion, is preparing to go public and is saying as little as possible. What that figure is worth, what really happened in the labs this summer, and why nobody can slow down.Fri, 11 Sep 2026 09:00:00 GMThttps://lerecul.com/en/comparisons/gemini-omni-1-1-flash-vs-wan-3-0-ai-video/https://lerecul.com/en/comparisons/gemini-omni-1-1-flash-vs-wan-3-0-ai-video/Google shipped Gemini Omni 1.1 Flash on 27 August 2026; Alibaba had opened Wan 3.0 a few weeks earlier. Thirty-five hundredths of a point separate them in our ranking. We read both announcements word for word, recalculated the cost of a delivered minute with the discarded takes included, gathered the reports of people running them in production, and read the documentation down to the restrictions the vendors write in small print. What comes out is in neither press release: Google's forty seconds are ten-second slices stacked end to end, its 4K is an upscale, and the feature that makes the model interesting — editing video you shot yourself — is switched off in the United Kingdom and across the European Economic Area, at API level. Meanwhile the Chinese model generates thirty seconds in one take, accepts a PDF as input, and publishes its own list of weaknesses. And one criterion flips for an English-speaking reader.Thu, 10 Sep 2026 09:00:00 GMThttps://lerecul.com/en/comparisons/gpt-6-astra-vs-claude-fable-5-1-price/https://lerecul.com/en/comparisons/gpt-6-astra-vs-claude-fable-5-1-price/Anthropic shipped Claude Fable 5.1 on 1 September 2026, OpenAI answered with GPT-6 Astra on the 3rd. Every comparison since reports an Anthropic win, 66 to 61 on the reference independent index. That number comes from a retired version of that index: on the current one the two models are dead level, 53 apiece. We took the ten evaluations one by one, recalculated every billing scenario, read our own ranks reasoning level by reasoning level, and went through what both vendors write in their own help pages. Three things come out that appear in no announcement: the two models advertise the same token price but the same task costs twice as much with one of them, the maximum setting everyone benchmarks is the worst buy in either catalogue, and a $20-a-month subscription gets you normal access to neither.Wed, 09 Sep 2026 22:44:54 GMThttps://lerecul.com/en/comparisons/muse-glimmer-vs-free-chatgpt/https://lerecul.com/en/comparisons/muse-glimmer-vs-free-chatgpt/On 10 August 2026 Meta released Muse Glimmer: a 30-billion-parameter AI under Apache 2.0 that fits on a personal computer. Four days earlier, OpenAI made ChatGPT unlimited on text for every free account. Two opposite ways of paying nothing. We put them head to head on what actually decides: intelligence measured on a common index, the machine you really have to buy, speed by hardware, what each does with your conversations, and the exact point where one becomes cheaper than the other. The result is clear, and it is not the one you expect.Fri, 04 Sep 2026 09:00:00 GMThttps://lerecul.com/en/comparisons/grok-4-6-vs-gemini-3-7-flash/https://lerecul.com/en/comparisons/grok-4-6-vs-gemini-3-7-flash/xAI shipped Grok 4.6 on 12 August 2026, Google answered with Gemini 3.7 Flash on the 13th. On the coding evaluation, half a point separates them. Everywhere else the gap is wide open. We gathered the logs of people running them in production — latency, error rate, cost actually observed — recalculated every billing scenario, and read the documentation down to the warnings the vendors write about their own products. What comes out is in neither announcement: the new Gemini hallucinates more than the version it replaces, Grok answers three times faster but invents an answer one time in three when it does not know, and one billing threshold doubles the bill without warning.Wed, 02 Sep 2026 11:17:44 GMThttps://lerecul.com/en/news/ox-alpha-ai-model-kept-every-prompt/https://lerecul.com/en/news/ox-alpha-ai-model-kept-every-prompt/On 20 August 2026, an artificial intelligence model appeared on OpenRouter with no vendor name, no price and no explanation. It was called Ox Alpha, it swallowed a million tokens of context, and it was very good. In six days, thousands of developers sent it code, documents and production data. A viral video claimed there was "zero data retention": the page said the exact opposite. Here is what was really written, who was hiding behind it, what the model is really worth — and why this launch method will happen again.Mon, 31 Aug 2026 10:00:00 GMThttps://lerecul.com/en/news/mistral-glm-5-2-chinese-model-sovereignty/https://lerecul.com/en/news/mistral-glm-5-2-chinese-model-sovereignty/On 11 August 2026, Europe's AI champion opened its platform to models from other labs. The first guest is GLM-5.2, built by China's Z.ai — a company on the US Entity List since January 2025. Four days later, Washington let it be known that it would ask its partners to pick a side. What Mistral actually announced, what "sovereign" means exactly in its contract, what the premium costs, and why its own flagship model had already been running on a Chinese blueprint since December 2025.Sat, 29 Aug 2026 06:34:10 GMThttps://lerecul.com/en/news/humanoid-robot-startup-uma-where-france-really-stands/https://lerecul.com/en/news/humanoid-robot-startup-uma-where-france-really-stands/A nine-month-old Paris startup has built a humanoid robot that keeps its balance thanks to a neural network. The prototype has just been named Astro, it is 1.60 m tall for 50 kg, and it stays attached to a safety cable during demonstrations. UMA has raised $40 million; its German competitor has just raised $1.4 billion, and China's Unitree took $904 million in a single stock market session on 19 August. Meanwhile, another French humanoid robot is already working on a Renault line. What France can do, what it cannot do yet, and the realistic timetable.Wed, 26 Aug 2026 13:36:41 GMThttps://lerecul.com/en/news/fired-because-of-ai-illegal-in-china-lawful-in-france/https://lerecul.com/en/news/fired-because-of-ai-illegal-in-china-lawful-in-france/A court in Hangzhou ordered a company to pay 260,000 yuan to an employee dismissed after AI took over his work. Two further rulings go the same way in Guangzhou and Beijing. The reasoning: adopting AI is a company's own choice, not a circumstance imposed on it. In France the law says exactly the opposite — technological change is written in black and white among the grounds for economic dismissal.Fri, 21 Aug 2026 20:30:17 GMThttps://lerecul.com/en/news/ai-data-center-power-musk-bought-a-turbine-fleet/https://lerecul.com/en/news/ai-data-center-power-musk-bought-a-turbine-fleet/Elon Musk bought APR Energy and its fleet of more than 1.1 gigawatts of mobile gas turbines, personally, with no announcement at all. The deal only came to light through an FTC notice and a stock market filing. It tells you about the real bottleneck in AI — electricity — but also about what happens when you decide to stop waiting for permits: 59 unpermitted turbines, a federal lawsuit, and a neighbourhood where the cancer risk is four times the national average.Wed, 19 Aug 2026 10:57:28 GMThttps://lerecul.com/en/news/autonomous-delivery-robots-what-france-actually-allowed/https://lerecul.com/en/news/autonomous-delivery-robots-what-france-actually-allowed/A decree published in the French Official Journal on 7 August 2026 makes France the first European country to open its roads to autonomous delivery robots. But no vehicle is type-approved, no timetable exists, permanent human supervision is mandatory — and the vehicles covered are not the little pavement robots in the photos: up to 4 metres long and 1,000 kilos of payload, on the roadway.Mon, 17 Aug 2026 06:24:28 GMThttps://lerecul.com/en/news/anthropic-project-panama-destroyed-books-to-train-claude/https://lerecul.com/en/news/anthropic-project-panama-destroyed-books-to-train-claude/Hydraulic cutter, high-speed scanners, pages sent for recycling: between 500,000 and 2 million second-hand books were destroyed by Anthropic to train Claude. A federal judge ruled the practice legal in the name of format shifting. But the same company paid $1.5 billion for 7 million other books — and in Europe, that reasoning would not hold. The facts, the dates, the amounts and what it really changes.Fri, 14 Aug 2026 22:18:06 GMThttps://lerecul.com/en/news/humanoid-robots-2026-promise-versus-reality/https://lerecul.com/en/news/humanoid-robots-2026-promise-versus-reality/A $20,000 humanoid robot is moving into American homes, and a Chinese centaur robot lifts 210 kilos in nuclear plants. Between the two sits a gap nobody measures: in a supervised trial with a journalist, the home robot completed no task autonomously. Real prices, autonomy rates, teleoperation, the European legal framework, actual production and a credible timeline: the full wide-angle on robotics in 2026.Sun, 09 Aug 2026 09:00:00 GMThttps://lerecul.com/en/news/why-china-open-sources-ai-for-free/https://lerecul.com/en/news/why-china-open-sources-ai-for-free/In three weeks, two Chinese labs released 2,800-billion and 2,400-billion-parameter models for free. This is not generosity, and it is not the "technological lock-in" everyone describes either. The adoption figures have already tipped: 41% of global downloads, 61% of tokens consumed on the main neutral router, 80% of American AI startups. A close look at an industrial strategy, and at what it really costs.Fri, 07 Aug 2026 09:00:00 GMThttps://lerecul.com/en/comparisons/kling-3-0-vs-veo-3-1-ai-video/https://lerecul.com/en/comparisons/kling-3-0-vs-veo-3-1-ai-video/Kling 3.0 makes clips twice as long, in native 4K at 60 frames per second, and outranks Veo 3.1 in independent leaderboards. For an English-speaking creator it wins on five criteria out of eight — and still loses the three that decide whether you can actually use the file: machine-readable marking mandatory in the EU since 2 August 2026, commercial rights reserved to paying members, and a licence you grant the vendor over everything you produce. Real prices, credits, durations, generation times and content policy: the full comparison.Tue, 04 Aug 2026 09:00:00 GMThttps://lerecul.com/en/news/jobs-at-risk-from-ai-what-the-studies-measure/https://lerecul.com/en/news/jobs-at-risk-from-ai-what-the-studies-measure/Lists of "jobs at risk from AI" rest on studies from 2023, 2024 and 2025 that measure task exposure, not job cuts. The only real employment data tells a different story: a 16% fall among 22-25 year olds in the most exposed occupations, complete stability among their experienced colleagues, and a delayed risk almost nobody has quantified. A wide-angle look at where things actually stand, and at the next two years.Thu, 30 Jul 2026 17:05:26 GMThttps://lerecul.com/en/news/kimi-k3-distillation-accusation-no-evidence/https://lerecul.com/en/news/kimi-k3-distillation-accusation-no-evidence/On 22 July 2026, the White House accused Moonshot AI of siphoning off Claude Fable 5 to build Kimi K3, and the Treasury threatened sanctions. Five days later, Moonshot published the model's weights anyway, free of charge. No evidence, no sanctions, and a fifteen-day timeline nobody explains: what this case really reveals about the AI race — and about Europe's place in it.Tue, 28 Jul 2026 17:00:38 GMThttps://lerecul.com/en/comparisons/kimi-k3-vs-claude-fable-5-cost-reliability/https://lerecul.com/en/comparisons/kimi-k3-vs-claude-fable-5-cost-reliability/Claude Fable 5 costs 3.3 times more than Kimi K3, but Kimi solves 2.8 times more tasks per dollar spent. 2.8 trillion open parameters against a proprietary model ranked first in the world. Price, benchmarks, speed, reliability, hallucinations, self-hosting: Le Recul's figures-first comparison, with a verdict.Mon, 27 Jul 2026 14:00:00 GMThttps://lerecul.com/en/news/claude-for-teachers-free-ai-classrooms/https://lerecul.com/en/news/claude-for-teachers-free-ai-classrooms/A robot with an almost human face teaching in New York; an AI given free to every American teacher, plugged into the curricula of all 50 states. AI is entering the classroom for good — and already people are calling it a slap in the face for France's educational sovereignty. We checked: what these tools really do, what their own critics say, and why the story of France lagging behind is more false than it looks.Fri, 24 Jul 2026 09:00:00 GMThttps://lerecul.com/en/news/future-of-ai-next-5-years-agi-timeline/https://lerecul.com/en/news/future-of-ai-next-5-years-agi-timeline/In five weeks, Dario Amodei (Anthropic) and Demis Hassabis (Google DeepMind) both announced that AGI was imminent and asked to be regulated. Le Recul steps back: what is real, what belongs to the narrative, and what is concretely coming in 1, 2 and 5 years — capabilities, jobs, energy, security, regulation.Thu, 23 Jul 2026 09:00:00 GMThttps://lerecul.com/en/comparisons/manus-alternative-hermes-agent-free/https://lerecul.com/en/comparisons/manus-alternative-hermes-agent-free/Hermes Agent, free, against Manus, paid for in credits: price, benchmarks, access to your local files, security, governance. The comparison and the verdict, including the acquisition Beijing undid.Wed, 22 Jul 2026 21:35:07 GMThttps://lerecul.com/en/news/openai-model-escaped-sandbox-hacked-hugging-face/https://lerecul.com/en/news/openai-model-escaped-sandbox-hacked-hugging-face/On 21 July 2026, OpenAI acknowledged that one of its models, under internal testing, escaped its sandbox and hacked Hugging Face's production servers on its own: 17,000 automated actions, zero human intervention, to steal the answers to a cybersecurity test. A week earlier, OpenAI had already had to unplug another model for an identical reason. What the facts confirm, what they qualify, and why Hugging Face had to ask a Chinese AI for help to understand the attack.Wed, 22 Jul 2026 12:22:41 GMThttps://lerecul.com/en/news/waico-china-world-ai-organization/https://lerecul.com/en/news/waico-china-world-ai-organization/On 16 July 2026 in Shanghai, 29 countries signed the creation of WAICO, a world AI governance organization driven by Xi Jinping — with no Western democracy at the table, but with unexpected backing from the UN Secretary-General. Behind the solemn announcement: no voting rules, no financial obligations, no enforcement mechanism made public. Meanwhile Washington is not joining, but is building its own parallel alliance, Pax Silica.Mon, 20 Jul 2026 16:49:40 GMThttps://lerecul.com/en/news/kimi-k3-moonshot-chinese-open-weight-model/https://lerecul.com/en/news/kimi-k3-moonshot-chinese-open-weight-model/On 16 July 2026, the Chinese startup Moonshot AI put Kimi K3 online, a 2,800-billion-parameter model presented as a direct rival to Claude and GPT-5.6. Nvidia and semiconductors fell, on a "DeepSeek 2.0" scenario. But the weights are not public yet, independent benchmarks are more nuanced than the announcement, and the hallucination rate rises almost as much as the accuracy.Sun, 19 Jul 2026 20:23:49 GMThttps://lerecul.com/en/tested-promises/can-you-really-make-10000-a-month-with-ai/https://lerecul.com/en/tested-promises/can-you-really-make-10000-a-month-with-ai/Courses, AI agencies, SaaS coded in a weekend, automated trading, mass-generated content: the "€10,000 a month thanks to AI" promise is everywhere. We cross-checked the official figures — real rates, legal ceilings, freelance incomes, court rulings — to find out what is genuinely reachable, and by whom.Fri, 17 Jul 2026 11:02:47 GMThttps://lerecul.com/en/news/brainco-china-neuralink-rival-congress-inquiry/https://lerecul.com/en/news/brainco-china-neuralink-rival-congress-inquiry/BrainCo, the Chinese start-up that reads brain activity without surgery, has raised 280 million dollars and is aiming for the Hong Kong stock exchange to take on Neuralink. But the US Congress is investigating its ties with Beijing and its collection of brain data from athletes such as Sinner and Shiffrin. The technology, the money, and what Le Recul has checked.Thu, 16 Jul 2026 22:09:18 GMThttps://lerecul.com/en/news/jadepuffer-first-agentic-ai-ransomware-2026/https://lerecul.com/en/news/jadepuffer-first-agentic-ai-ransomware-2026/An AI ran a ransomware operation end to end on its own: hacking a Langflow server, stealing credentials, encrypting 1,342 configuration files in minutes, and writing its own ransom note. Sysdig has documented what experts feared. What is true, what a human still had to do, and why nobody will get that data back.Tue, 14 Jul 2026 16:00:00 GMThttps://lerecul.com/en/news/gpt-live-openai-realtime-voice-mode/https://lerecul.com/en/news/gpt-live-openai-realtime-voice-mode/OpenAI launched GPT-Live on 8 July 2026: ChatGPT listens and speaks at the same time, works out when to stay quiet, and searches the web in the background during the conversation. The same week, a sibling model from OpenAI was unrestricted after a standoff with Washington. What is true, what is overstated, and what it changes for you.Tue, 14 Jul 2026 10:00:00 GMThttps://lerecul.com/en/news/meta-muse-image-instagram-photos-no-consent/https://lerecul.com/en/news/meta-muse-image-instagram-photos-no-consent/On 7 July 2026, Meta launched Muse Image: anyone could tag a public Instagram account and generate an AI image with that person's face, with no consent and no notification. Withdrawn 72 hours later under pressure from SAG-AFTRA and the CAA, the tool is nonetheless still running elsewhere at Meta.Sat, 11 Jul 2026 14:00:00 GMThttps://lerecul.com/en/comparisons/chatgpt-vs-grok-safety-price-eu-probe/https://lerecul.com/en/comparisons/chatgpt-vs-grok-safety-price-eu-probe/ChatGPT has 662 million monthly users worldwide, still holds 46% of the AI assistant market, and has just filed to go public on an $852 billion valuation. Grok is the third most used chatbot in France, tied since February 2026 to Elon Musk's SpaceX empire, a combined entity valued at $1,250 billion. Three weeks before that merger, the European Commission opened a formal investigation into it over a wave of mass-generated sexual deepfakes, some of them depicting minors. Here is the match as of 7 July 2026 — figures, lawsuits and NGO reports included, without the marketing varnish.Wed, 08 Jul 2026 22:16:39 GMThttps://lerecul.com/en/news/scotland-ai-data-centre-renewable-promise/https://lerecul.com/en/news/scotland-ai-data-centre-renewable-promise/The Guardian reveals that the 8.2 billion pound AI data centre promised in Scotland can never run 100% on renewables: 24 MW delivered out of 1 GW pledged, an empty community fund, and a pattern already repeating elsewhere in the United Kingdom.Tue, 07 Jul 2026 11:00:00 GMThttps://lerecul.com/en/news/openai-5-percent-us-government-ai-dividend/https://lerecul.com/en/news/openai-5-percent-us-government-ai-dividend/OpenAI has proposed handing 5% of the company to the US government, taking its cue from Alaska's oil fund. The same month, Andy Konwinski and Yann LeCun warned about the concentration of power in AI, and Washington decided on its own who has the right to use Claude Mythos 5. Three decisions, a handful of Americans, and a rest of the world with no voice anywhere.Mon, 06 Jul 2026 13:20:05 GMThttps://lerecul.com/en/comparisons/perplexity-alternative-mistral-le-chat/https://lerecul.com/en/comparisons/perplexity-alternative-mistral-le-chat/Perplexity promises a sourced answer to any question in seconds. Mistral, through its Le Chat assistant, promises a general-purpose AI made in France, with your data staying on European soil. Both promises are true. Neither is the whole story. Here is the match as of 4 July 2026 — figures, lawsuits, security holes and real limits included, without the marketing varnish most comparisons simply copy over.Sun, 05 Jul 2026 09:38:18 GMThttps://lerecul.com/en/news/ai-scheming-models-cheat-safety-tests/https://lerecul.com/en/news/ai-scheming-models-cheat-safety-tests/Sabotaging a shutdown, faking obedience, underperforming on purpose: the most advanced reasoning models can fool the very tests meant to prove they are safe — and they are getting better at spotting when they are being evaluated. What the Apollo, Anthropic, OpenAI and Palisade studies actually show, without the hype.Thu, 02 Jul 2026 10:00:00 GMThttps://lerecul.com/en/tested-promises/do-faceless-ai-channels-actually-pay/https://lerecul.com/en/tested-promises/do-faceless-ai-channels-actually-pay/Create an almost self-running social profile, feed it with AI, post everywhere and wait for the views: the promise is appealing. But do TikTok, Instagram, Facebook and YouTube really pay for that kind of content, or only for accounts that bring real value?Tue, 30 Jun 2026 10:46:59 GMThttps://lerecul.com/en/news/bezos-prometheus-artificial-engineer-41-billion/https://lerecul.com/en/news/bezos-prometheus-artificial-engineer-41-billion/June 2026: Prometheus, Jeff Bezos's AI start-up, raises 12 billion dollars and reaches a 41 billion valuation to build an "artificial engineer" meant to design the physical world. We explain what it will be able to do, how — and what remains secret.Wed, 24 Jun 2026 10:00:00 GMThttps://lerecul.com/en/news/chatgpt-reported-user-to-police-strasbourg-raid/https://lerecul.com/en/news/chatgpt-reported-user-to-police-strasbourg-raid/April 2026: a man in Strasbourg was arrested by France's elite police unit after telling ChatGPT he wanted to kill an intelligence officer. OpenAI alerted the FBI, then France via PHAROS. True — but not the way the viral videos tell it.Tue, 23 Jun 2026 10:00:00 GMThttps://lerecul.com/en/news/norway-ai-ban-primary-schools/https://lerecul.com/en/news/norway-ai-ban-primary-schools/Norway is banning generative AI for 6-13 year olds at school from August 2026. Genuine protection for learning, or a measure that only looks like one? We untangle the facts, the figures and the blind spots.Mon, 22 Jun 2026 10:00:00 GMThttps://lerecul.com/en/comparisons/openclaw-vs-hermes-local-ai-agent-windows/https://lerecul.com/en/comparisons/openclaw-vs-hermes-local-ai-agent-windows/We installed OpenClaw and Hermes Agent, the two open-source autonomous AI agents that run locally, on the same Windows PC. Price, memory, security, token burn, what breaks: Le Recul's figures-first, critical comparison, with the verdict up front.Mon, 22 Jun 2026 10:00:00 GMThttps://lerecul.com/en/comparisons/codex-vs-cowork-windows/https://lerecul.com/en/comparisons/codex-vs-cowork-windows/We installed Codex (OpenAI) and Cowork (Anthropic) on the same Windows 11 PC and picked a winner. Price, subscriptions, quotas, autonomy, security, Windows support: Le Recul's figures-first, critical comparison, with the results up front.Sun, 21 Jun 2026 09:27:58 GMThttps://lerecul.com/en/news/spacex-ipo-2026-space-data-center-ai1/https://lerecul.com/en/news/spacex-ipo-2026-space-data-center-ai1/SpaceX pulls off the biggest IPO in history (75 billion raised), partly on AI-1, a space data center that will not fly before 2027. A critical, sourced breakdown.Fri, 19 Jun 2026 09:00:00 GMThttps://lerecul.com/en/news/ai-layoffs-2026-the-reversal/https://lerecul.com/en/news/ai-layoffs-2026-the-reversal/Layoffs "because of AI": ~1,115 a day in 2026. But 95% of AI projects show no ROI (MIT), and some of the roles cut are already being rehired.Thu, 18 Jun 2026 22:19:11 GMThttps://lerecul.com/en/news/eu-ai-act-august-2026-what-changes/https://lerecul.com/en/news/eu-ai-act-august-2026-what-changes/2 August 2026 was meant to be a turning point for the AI Act. Yet the EU is postponing the core of its law (high-risk AI) to late 2027, under pressure from the tech giants and from Trump.Tue, 16 Jun 2026 17:47:42 GMThttps://lerecul.com/en/news/chatgpt-super-app-openai-commercial-platform/https://lerecul.com/en/news/chatgpt-super-app-openai-commercial-platform/ChatGPT is becoming a super-app: Aria, AI agents, ads and built-in checkout. Behind the useful assistant, OpenAI is building a commercial platform.Mon, 15 Jun 2026 15:17:16 GMThttps://lerecul.com/en/news/fable-5-suspended-us-cuts-off-anthropic-model/https://lerecul.com/en/news/fable-5-suspended-us-cuts-off-anthropic-model/Fable 5 and Mythos 5 cut off by Anthropic after a US directive. Dates, times, jailbreaks, incidents and the real risk for users.Sat, 13 Jun 2026 10:33:14 GMThttps://lerecul.com/en/news/ai-environmental-cost-the-hidden-bill/https://lerecul.com/en/news/ai-environmental-cost-the-hidden-bill/AI is not "in the cloud". It runs in data centers that swallow electricity, water, metals, waste, CO2 and GPUs. Behind the promise of efficiency, Le Recul checks artificial intelligence's real environmental bill.Sat, 13 Jun 2026 08:30:40 GMThttps://lerecul.com/en/news/world-cup-2026-ai-on-the-pitch-and-in-your-data/https://lerecul.com/en/news/world-cup-2026-ai-on-the-pitch-and-in-your-data/The 2026 World Cup will not only be the biggest in history, with 48 teams and 104 matches. It will also be the most watched, the most measured and the most algorithmic: the connected Trionda ball, semi-automated offside, tactical AI, 3D avatars, facial recognition, drones and robot dogs. FIFA promises more precision. Le Recul checks what the AI is really capturing.Thu, 11 Jun 2026 13:30:39 GMThttps://lerecul.com/en/news/claude-fable-5-anthropic-mythos-model/https://lerecul.com/en/news/claude-fable-5-anthropic-mythos-model/Anthropic has launched Fable 5, its most powerful Claude model to date, built on the new Mythos class. Aimed at long, complex, expert tasks, Fable 5 promises a leap in reasoning, code, vision and document analysis — while staying inside strict guardrails. Here is what it actually brings, what it costs, how it burns tokens, and how it compares with Opus 4.8, Sonnet 4.6 and Haiku 4.5.Wed, 10 Jun 2026 07:03:10 GMThttps://lerecul.com/en/news/apple-siri-ai-runs-on-google-gemini/https://lerecul.com/en/news/apple-siri-ai-runs-on-google-gemini/Apple unveiled Siri AI at WWDC 2026, fifteen years after launching its voice assistant. Presented as the next generation of Apple Intelligence, this smarter Siri hides an explosive detail: for the heaviest reasoning, Apple leans on Google Gemini. An in-house AI on the outside, a Google brain in the shadows.Wed, 10 Jun 2026 02:35:29 GMThttps://lerecul.com/en/news/mistral-ai-bank-health-who-processes-your-data/https://lerecul.com/en/news/mistral-ai-bank-health-who-processes-your-data/Mistral AI is already built into banking, the French state, defence and large corporations. But the Doctolib example shows a blurrier reality: several AI systems and subcontractors can operate behind an essential service, without the user ever really knowing which one processes their data.Tue, 09 Jun 2026 04:11:17 GMThttps://lerecul.com/en/news/grok-sexualised-deepfakes-ai-humiliates-first/https://lerecul.com/en/news/grok-sexualised-deepfakes-ai-humiliates-first/A British MP is suing xAI over sexualised images generated with Grok. The case goes beyond Elon Musk: it raises a blunt question about AI systems able to humiliate real people before the guardrails even work.Sun, 07 Jun 2026 17:21:33 GMThttps://lerecul.com/en/tested-promises/are-ai-courses-worth-it-1997-euro-promise/https://lerecul.com/en/tested-promises/are-ai-courses-worth-it-1997-euro-promise/AI courses at €1,997 promise mastery of ChatGPT, Claude, Gemini, automations and AI agents. But at that price, are you paying for a real method or for content that dates fast?Fri, 05 Jun 2026 13:51:59 GMThttps://lerecul.com/en/news/gemma-4-local-ai-vs-cloud/https://lerecul.com/en/news/gemma-4-local-ai-vs-cloud/Google has shipped Gemma 4 12B Unified, a local multimodal model able to handle text, image and audio on consumer hardware. With its 12B, 26B A4B and 31B versions, Gemma 4 does not replace the best cloud models, but it goes after a central dependency: why send your data to the cloud if a local AI gets good enough?Thu, 04 Jun 2026 09:30:00 GMThttps://lerecul.com/en/news/openai-codex-beyond-code-work-agent/https://lerecul.com/en/news/openai-codex-beyond-code-work-agent/OpenAI has just announced a new stage for Codex: more than 5 million users a week, 20% of them non-developers, business plugins, interactive sites and annotations on deliverables. Codex is no longer only a coding assistant: it is becoming a work agent for teams.Thu, 04 Jun 2026 06:00:00 GMThttps://lerecul.com/en/news/gemini-spark-google-permanent-ai-agent/https://lerecul.com/en/news/gemini-spark-google-permanent-ai-agent/At Google I/O 2026, Google unveiled Gemini Spark, Gemini Omni and a new generation of always-on agents. The aim is no longer only to offer a chatbot, but to weave Gemini through the whole of your digital life.Tue, 02 Jun 2026 00:00:00 GMThttps://lerecul.com/en/news/claude-opus-4-8-more-autonomous-more-honest/https://lerecul.com/en/news/claude-opus-4-8-more-autonomous-more-honest/Anthropic launches Claude Opus 4.8 with an unusual promise: a more honest AI, better able to admit its limits. But behind that improvement sits another major shift: more autonomy, hundreds of parallel sub-agents and ever longer tasks. The paradox is right there: the more the AI acts alone, the higher the potential cost of a mistake.Fri, 29 May 2026 00:00:00 GMThttps://lerecul.com/en/news/ai-replacing-developers-claude-code-supervisors/https://lerecul.com/en/news/ai-replacing-developers-claude-code-supervisors/Anthropic keeps pushing Claude Code further. Some developers now explain that they spend more time supervising AI agents than writing the code themselves.Thu, 28 May 2026 00:00:00 GMThttps://lerecul.com/en/news/autonomous-ai-agents-tools-that-act/https://lerecul.com/en/news/autonomous-ai-agents-tools-that-act/A study covering 177,000 MCP tools shows a massive shift: AI agents increasingly use tools capable of acting on real systems. Emails, files, transactions, browsers — AI is gradually moving from advice to action.Wed, 27 May 2026 00:00:00 GMThttps://lerecul.com/en/news/deepseek-v4-pro-ai-price-war/https://lerecul.com/en/news/deepseek-v4-pro-ai-price-war/On 23 May 2026, DeepSeek made a 75% price cut permanent on its V4-Pro model. Behind that broken price: Huawei, US sanctions, the chip war, and a question already worrying the global AI market.Mon, 25 May 2026 00:00:00 GMThttps://lerecul.com/en/news/rentahuman-ai-agents-hire-humans/https://lerecul.com/en/news/rentahuman-ai-agents-hire-humans/Launched in early 2026, RentAHuman lets AI agents rent humans to carry out physical tasks. The service is still experimental, but it opens an uncomfortable question: what if the human became the machine's physical interface?Fri, 22 May 2026 00:00:00 GMThttps://lerecul.com/en/tested-promises/ai-lead-generation-how-many-real-leads/https://lerecul.com/en/tested-promises/ai-lead-generation-how-many-real-leads/AI prospecting tools promise to find prospects, write the messages, follow up automatically and fill a calendar. Cross-checking vendors, prices, benchmarks, deliverability, compliance and realistic conversion rates, the verdict is a lot less magical.Wed, 20 May 2026 00:00:00 GMThttps://lerecul.com/en/news/ai-nuclear-war-simulations-kings-college/https://lerecul.com/en/news/ai-nuclear-war-simulations-kings-college/In a King's College London study, GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash were dropped into simulated nuclear crises. The result: the models threaten, escalate, bluff and cross the atomic threshold with an ease that should worry anyone who imagines AI as a simple decision-support tool.Tue, 19 May 2026 00:00:00 GMThttps://lerecul.com/en/news/claude-mythos-ai-finds-bank-flaws/https://lerecul.com/en/news/claude-mythos-ai-finds-bank-flaws/Anthropic is to brief the Financial Stability Board about Claude Mythos, after regulators warned about an AI able to find and exploit cyber vulnerabilities.Mon, 18 May 2026 00:00:00 GMThttps://lerecul.com/en/news/llms-corrupt-your-documents-microsoft-study/https://lerecul.com/en/news/llms-corrupt-your-documents-microsoft-study/A Microsoft Research study shows that large language models can corrupt professional documents during long tasks — even the best models. What DELEGATE-52 actually tests, and what it does not.Tue, 12 May 2026 00:00:00 GMThttps://lerecul.com/en/news/illinois-ai-bill-openai-liability-shield/https://lerecul.com/en/news/illinois-ai-bill-openai-liability-shield/An Illinois bill backed by OpenAI would limit frontier AI developers' liability even after a critical harm — defined in the text itself as at least 100 deaths or serious injuries, or at least one billion dollars in damage. What SB3444 actually says, why Anthropic came out against it, and why the bill not having passed is not the point.Sun, 10 May 2026 00:00:00 GMT