--- title: "What AI Applications Do I Use" date: 2026-05-01 description: "The AI tools I actually use — from Claude to Midjourney, NotebookLM to Sora 2. An honest user\'s map of a battlefield that keeps moving." author: "Petru Cojocaru" categories: - name: "AI - a tool" url: "https://cojocarupetru.info/ai-a-tool.md" --- # What AI Applications Do I Use **Charting a Battlefield in Transformation.** In the autumn of 2025, as I write these lines, I find myself in the paradoxical position of the historian who tries to document the present. It is rather like attempting to draw a map of a battlefield whilst the bombardment still echoes in the distance, whilst the front line shifts with every hour that passes.  [![What AI Applications Do I Use](https://cojocarupetru.info/images/AI-o-unealta/ai-o-unealta.jpg "What AI Applications Do I Use")](https://cojocarupetru.info/index.php?option=com_content&view=article&id=349:what-ai-applications-do-i-use&catid=41:ai-a-tool&lang=en-GB) I set out to write about the artificial intelligence tools I use daily, only to discover quickly that this apparent simplicity conceals a vertiginous complexity. Every week brings a new model, a new capability, a new promise. Claude Sonnet 4.5 was released at the end of September, transforming the industry's standards in the field of programming. A mere fortnight later, Claude Haiku 4.5 was demonstrating that near-peak performance can come in a faster and more economical package. In August, GPT-5 had arrived with great fanfare, promising a new era of artificial reasoning. I am overwhelmed. I admit this without embarrassment, the way an honest general acknowledges that his map is becoming obsolete before it is even finished. But this sense of being overwhelmed is not resignation — it is that paradoxical state in which the awareness of one's own limits becomes the starting point for the search for pragmatic solutions. What follows is not an academic treatise on artificial intelligence, nor a comprehensive technical review. It is a letter to you, the reader, written from the trenches of an ordinary user attempting to navigate this landscape of constant change. It is the confession of someone who seeks simple solutions for complex problems, who accepts that he cannot know everything, yet still hopes to understand enough to be of use. Just as Beevor reconstructed the battles of Normandy by intersecting the perspective of the soldier in the trenches with that of the general at headquarters, so too shall I attempt to blend technical detail with human experience, the particular with the universal, enthusiasm with scepticism. --- ## 1. Large Language Models: The Heavy Artillery of the New Era ### Free: The Democratisation of Intelligence **Claude Sonnet 4.5** – When I enter into dialogue with Claude, I have the strange sensation of speaking with someone who has shed their inhibitions without becoming reckless. Released on the last day of September 2025, this model quickly became 'the world's best coding model', as Anthropic boldly claims. I tested that assertion one rainy afternoon, asking it to analyse a complex GitHub repository. It managed to identify problems that other AI assistants had missed, proposing elegant solutions that worked on the first attempt. What fascinates me about Claude is the way it 'thinks' — it seems to have a predilection for unconventional solutions, for lateral perspectives. When I ask for help with an article or a brainstorming session, I do not receive generic, by-the-book responses. I receive perspectives that make me stop and reflect. Recently, according to the official documentation, it can operate autonomously for more than thirty hours on complex tasks, maintaining focus and incremental progress — a remarkable achievement that transforms the concept of 'assistant' into that of 'long-haul colleague'. Its Artefacts are spectacular — those interactive windows in which it constructs complete applications, data visualisations, or formatted documents. It is as though you were to ask a craftsman to make you a table and found yourself watching it take shape before your eyes, every joint perfect, every detail considered. There is, however, a hidden cost to this generosity: the strict message limit for free accounts. It is as though you were given access to a magnificent library, but only for a few hours each week. The frustration is real when you find yourself in the middle of a complex project and receive the notification that you have exhausted your daily allowance. **Gemini 2.5 Pro** (via Google AI Studio) – If Claude is the spiritual conversationalist, Gemini is the encyclopaedic librarian with a photographic memory. Its context window is obscenely large — it can process documents, audio, and video natively, making it a genuine beast for working with massive quantities of information. I use it as my default 'beta reader'. When I finish an article or a chapter, I send it to Gemini along with a few reference materials. It returns with structured feedback, identifying argumentative gaps, stylistic inconsistencies, and even missed opportunities to develop certain ideas further. It is that friend who reads attentively and is not afraid to tell you uncomfortable truths. Recently I uploaded a two-hour audio recording from a conference and asked it to analyse it alongside the slides that had been presented. It extracted not only a transcript, but also the relationships between the ideas delivered orally and the visual materials, identifying key moments and actionable recommendations. This native multimodal capability is something that other competitors are still trying to replicate. **GPT-4o / GPT-5** – Here the story becomes complicated, because OpenAI has created a peculiar situation. GPT-5 was launched in August 2025 with great fanfare — Sam Altman described it as 'a significant step towards AGI', at least in their theoretical framework. The model promised an 80% lower hallucination rate than its predecessors when using extended reasoning, improved coding capabilities (144% better than GPT-4o, they claim), and the ability to 'think' when the situation demands it. But the reality has been more nuanced. Users complained that GPT-5 felt less intuitive, more rigid in conversation. OpenAI had to restore GPT-4o as an option for paying customers who preferred the better 'conversational feel' of the previous model. I myself frequently gravitate towards GPT-4o for conversation, even if GPT-5 is theoretically superior. What makes ChatGPT irresistible is not so much any individual model, but the complete package: web access, image generation, Advanced Voice Mode (an almost uncanny experience when it shares your screen or camera during multimodal conversations), advanced data analysis, and deep research capabilities. It is a Swiss Army knife of the digital age. ### Paid: When Need Exceeds Gratitude **OpenAI o3** – This is the heavy artillery for when problems become truly difficult. Although GPT-5 Thinking (the extended reasoning variant) achieves marginally better benchmark scores, I have a soft spot for o3. Perhaps because I encountered it first; perhaps because its response structure — automatic tables, systematic organisation — suits my thinking style. When I work on complex research projects, o3 becomes the archaeologist who not only finds the buried information, but also organises it in chronological stratigraphy. Recently I asked it to analyse the impact of semiconductor export restrictions on technology investment strategies. It generated a report structured in clear tables, with verified sources and actionable recommendations. Others find this style 'irritating' — I find it useful. It is a matter of personal preference, rather like choosing between a teacher who writes everything on the board and one who improvises dynamically. **ChatGPT Plus** – At twenty dollars a month, this subscription is what I call 'a legal steal'. Not because OpenAI is generous by nature, but because the volume of functionality on offer far exceeds the price asked. It is as though you had paid for a chamber music concert ticket and discovered that it is in fact a three-day festival. The list is long: the best native image generation in its class, Advanced Voice Mode (live, multimodal conversations in which the AI can 'see' through your camera), Advanced Data Analysis ('Code Interpreter', which can run Python code for complex statistical analyses), Deep Research (one of the best research agents on the market), reasoning models such as o3, Canvas (the lateral space for collaborative editing of code and documents), and web browsing for current information. But perhaps the most valuable element is the least discussed: the ability to create custom GPTs. I think of these mini-assistants as a team of specialists I have recruited and trained for specific tasks. I have a GPT for academic text editing that knows my style and the conventions of the field exactly. I have another for code debugging that is familiar with the technology stack I use. Each one is a context economy — I no longer need to explain the same preferences and constraints every single time. --- ## 2. AI Images: When Pixels Become Dreams ### Free: The Democratisation of the Visual **GPT-4o Native Image Generation** – One spring morning, after reading about the native image generation integration in GPT-4o, I decided to test it. 'Create an image of a pelican riding a bicycle,' I wrote, without any further specifications. Within a few seconds I received a fully realised image: a recognisable pelican, in a plausible pose, on a bicycle drawn correctly from a technical standpoint. What astonished me was not only the visual quality, but the understanding of context. When I then asked for 'the same scene, but in the style of French Impressionism, at sunset', the transformation was instantaneous and coherent. The model does not merely generate pixels — it understands stylistic, historical, and emotional contexts. But the most revolutionary capability is text rendering. Previous generations of AI models made a mess of letters and words. GPT-4o can render entire pages of legible text within an image. I tested it with requests for conference posters, complex diagrams with labels, even comic pages with speech bubbles. The results are, in the most literal sense, unbelievable. Ironically accessible: although ChatGPT Plus is required for unlimited access, you can run the same system for free through copilot.microsoft.com — a detail that very few people know. **Gemini 2.5 Flash** (nicknamed 'Nano Banana') – If GPT-4o is the painter who creates original works, Nano Banana is the restorer who repairs and modifies with surgical precision. This small, fast version of Gemini specialises in image editing with detail preservation. The process is strangely natural: you upload an image, describe in plain language what you wish to change, and the model does exactly that. 'Change the colour of the sky from blue to sunset orange, but keep all other colours intact.' It works. 'Add a cat in the bottom left corner, but make sure the lighting matches the rest of the scene.' It works. I used Nano Banana recently to modify a series of photographs for an article. I needed to remove distracting elements from the background without losing the details in the foreground. Other AI tools left visible artefacts or subtly altered the tone of the entire image. Nano Banana made the modifications so subtly that even I could not identify the edited areas. ### Paid: Aesthetic Mastery **Midjourney V7** – To speak of Midjourney in 2025 is to speak of a legendary musician who still dominates the concert stage, even though younger rivals have learnt his technical tricks. The truth is that many models are now better at text rendering, at scene complexity, at precise prompt-following. But Midjourney remains the undisputed master of stylistic exploration and the creation of a personalised aesthetic. I have written dozens of posts about Midjourney over the years because, unlike other tools that focus on 'correctness' and 'precision', Midjourney is concerned with something more elusive: beauty. When I ask for a photorealistic image, I receive not merely a technically accurate recreation — I receive a photograph you would want to frame. Recently I experimented with creating a visual identity for a personal project. I began with a vague prompt about 'Mediterranean melancholy at sunset'. Midjourney generated variations that not only met the technical requirements, but each one told a different story about the same emotion. I spent hours exploring the stylistic ramifications, gradually learning the visual language of my own vision. --- ## 3. AI Video: Movement Becomes Reality The AI video space is making the most spectacular leaps in recent months. It is as though we were witnessing the invention of the sound film after decades of silent cinema — every week brings a new capability that would have seemed impossible six months ago. ### Free: The Threshold of Visual Democracy **Kling 2.5 Turbo** – Kling has become my favourite for image-to-video work for pragmatic reasons: regular updates, interesting new features, and enough free starter credits to experiment seriously. The process is simple yet magical: you upload a static image, describe the movement you wish to see, and Kling transforms it into a short video clip. I tested it with a photograph I had taken on a windy day — trees frozen in motion. 'Make the leaves move in the wind, add subtle branch movement.' The result was so convincing that friends who saw the video clip asked when I had made the recording, refusing to believe it had been generated from a static image. **Sora 2** – OpenAI launched Sora 2 with great ceremony and a 'social network'-style platform that, frankly, most serious users detest. But beyond the irritating interface lies the best text-to-video model in its class. What sets it apart is the capacity to generate its own synchronised audio and to function with diverse prompting approaches. I generated a sequence featuring a busy urban street. Sora 2 not only created the plausible visual movement of pedestrians and vehicles — it added traffic noise, distant conversations, even the sound of a tram bell passing by. The audio-video synchronisation was imperfect, but its very existence was wonderful. For now it is free, but OpenAI has announced that this 'will not last'. Enjoy the temporary generosity. ### Paid: Professional Tools **Krea** – When you wish to experiment with multiple models without juggling different platforms, Krea becomes the elegant solution. It is an aggregator that brings together third-party video models (plus image and 3D generation) into a single, clean interface. You can quickly compare results for the same prompt across different engines, thereby learning the strengths of each. **Runway Gen-4** – The recent Gen-4 model is solid (although Gen-4 Turbo is less impressive — proof that speed and quality remain in competition). But the real value of Runway lies in their all-in-one platform for creators, with dozens of robust tools for professional filmmakers: frame-by-frame editing, colour grading, special effects. It is the difference between having a war chariot and having an entire arms factory. **Veo 3.1** – The 3.1 update has transformed Veo into a top-tier model for realistic video clips with native audio effects and speech. I recently generated a sequence featuring an interview — the AI-generated person spoke naturally, with near-perfect lip synchronisation and convincing intonation. It is not yet at the level of deceiving someone who watches closely, but it is disturbingly close. --- ## 4. AI Music: The Symphony of Algorithms This space is surprisingly stable compared to the chaos in video. Perhaps because music, unlike images or text, is less susceptible to objective 'correctness' — either it moves you emotionally or it does not. ### Free and Paid: The Democratisation of Sound **Suno, Udio, Riffusion FUZZ** – The three are largely interchangeable for casual users. Differences exist — Suno perhaps has slightly superior audio quality following recent feature releases, Udio excels at certain specific genres, Riffusion FUZZ is exceptional for experimental music — but all offer generous free credits. I recently composed a soundtrack for a video project using Suno. My prompt was vague: 'electronic melancholy with jazz elements, slow rhythm, for a scene of urban rain.' Within two minutes I had three complete variations, each with a different interpretation of the same atmosphere. One had a subtle saxophone in the background. Another added synthesisers with deep reverb. The third was more minimalist, concentrating on electric piano and bass. The quality is good enough for musical background in personal projects, podcasts, and YouTube videos. It is not yet at the level of competing with professional human production, but for 99% of an individual creator's needs, it is more than adequate. Once you have chosen your favourite, you can always upgrade for additional features. I occasionally pay for Suno when I have larger projects, but the free credits are sufficient for sporadic use. --- ## 5. Research: Charting Knowledge ### Free: Algorithmic Librarians **NotebookLM** – I have been a fan of NotebookLM since March 2024, and the product continues to improve. It is that ideal research assistant who not only finds information, but understands the relationships between ideas. You upload documents, articles, notes — any sources — and NotebookLM creates a personalised workspace in which you can explore the connections. My favourite feature remains the audio podcast generation. You upload your research materials, and NotebookLM creates a two-way conversation between two AI 'hosts' who discuss the content, ask questions, and make unexpected connections. It is as though you had two university lecturers discussing your subject whilst you listen and learn. Recently they introduced video presentations and a revised Studio panel, transforming the tool from a simple notes aggregator into a complete knowledge synthesis platform. **Learn About** (Google) – If NotebookLM is the librarian who works with the books you bring, Learn About is the explorer who ventures into the world to find new territories. It can navigate the web and is built for open-ended exploration of any subject. It is the difference between directed research (I know what I am looking for) and pure serendipity (I want to discover). **Perplexity** – I do not use Perplexity very often, but its 'Pro' search occupies a valuable intermediate position between simple web browsing (superficial, but quick) and intensive 'Deep Research' tools (thorough, but time-consuming). When I need a quick but cited answer, Perplexity delivers. ### Paid: Diving into the Depths **Deep Research** (Google or OpenAI) – Both companies now offer 'Deep Research' tools that intelligently analyse dozens of sources over the course of ten to fifteen minutes. It is as though you had a doctoral research assistant who reads everything they can find that is relevant and produces a structured report for you. I recently used OpenAI's Deep Research to understand the impact of new cryptocurrency regulations. I received a fifteen-page report with analysis from multiple perspectives, quotations from official documents, expert opinions, and possible development scenarios. The time saved was enormous compared to manual research. --- ## 6. Productivity: Tools for Everyday Life ### Free: Daily Efficiency **Comet (Perplexity)** – I have not yet put Comet through its full paces — it is too new for a definitive assessment. But the concept of a browser with native AI integration is handy for many tasks: automatic summarisation of web pages, contextual translation, extracting key information without reading lengthy articles. It is as though your browser becomes a persistent assistant working in the background. **Google AI Studio** – It is still remarkable to me that Google makes so much available for free through this platform. You can access the latest frontier Gemini models (2.5 Pro), image generation, voice generation, coding tools, and live calls with screen or camera sharing. It is an excellent way to experiment with cutting-edge technologies without any financial outlay. I use Google AI Studio when I want to test complex prompts or compare behaviour across different models. The interface is technical — it is not designed for the casual consumer — but for those who wish to understand how these systems work 'from the inside', it is invaluable. **Napkin AI** – For people like me who struggle with design, Napkin AI is a lifesaver. It automatically transforms walls of text into compelling visuals: diagrams, infographics, data visualisations. You write a paragraph explaining a complex process, press a button, and receive three nicely formatted diagram options. It is as though you had a graphic designer working at the speed of thought. ### Paid: Orchestrated Power **Superagent (Genspark)** – This is the closest thing to a real 'AI agent' I have used to date. It can orchestrate dozens of tools to complete complex, multi-step tasks, often without constant human intervention. I recently gave it a complicated task: 'Analyse the performance of our competitors over the last quarter, identify gaps in our content strategy, and propose an editorial plan for the next two months.' Superagent used web scraping tools, sentiment analysis, content generation, and strategic planning, delivering a complete report in approximately thirty minutes. You receive monthly free credits to try it out, but you will probably need the paid plan for larger tasks. The investment is worthwhile if you work on projects where time is more valuable than money. --- ## Conclusion: The Map and the Territory In *Stalingrad*, Antony Beevor writes about the moment when German soldiers realised that their military map no longer corresponded to the reality of the terrain. The city they believed they understood had become an unpredictable labyrinth, constantly changing under bombardment and street fighting. Every morning brought a new landscape. That is how I feel about the artificial intelligence field in October 2025. The map I began to draw at the start of this article is already out of date. Perhaps a major new launch has taken place since then. Perhaps one of the tools I wrote about with enthusiasm has changed its pricing policy or been surpassed by a rival. The territory moves faster than I can write about it. But this does not invalidate the exercise. Quite the contrary. In the midst of this dizzying transformation, what matters is not having the perfect map — it is learning to navigate without one. I began this text by acknowledging that I am overwhelmed. Now, having charted the tools I use, I understand the nature of that overwhelm more clearly. It is not a matter of being incapable of keeping pace — no one can. It is a matter of choosing where to invest limited attention. **The simple solutions I have found:** **First**: Strategic specialisation. I do not need to use every tool. I have chosen a few for each major category — conversation (Claude, GPT), images (Midjourney for aesthetics, GPT-4o for utility), video (Kling for rapid experiments), research (NotebookLM for depth). Instead of chasing every new release, I let the tools prove their long-term value to me. **Second**: Accepting imperfection. When GPT-5 was launched with great fanfare but proved less intuitive than GPT-4o, I learnt that 'the newest' does not automatically mean 'the best for me'. I frequently return to older tools when they serve me better. **Third**: Separating the level of enthusiasm from practical relevance. The AI industry is full of hype. Companies announcing 'revolutions' that are in fact incremental improvements. I have learnt to ignore the superlatives in press releases and to test real utility for myself. **Fourth**: The community as compass. I follow what creators I respect are using, what developers whose projects I trust are recommending. Collective wisdom is more stable than individual enthusiasm. **Fifth**: Recognising that being overwhelmed is the normal state, not the exception. Once you accept that you will never know everything, the pressure to know everything disappears. What remains is only curiosity and pragmatism. The perspectives I have tried to integrate here are multiple and contradictory. The enthusiasm of the early adopter who wants to test everything. The scepticism of the user burnt by unfulfilled promises. The pragmatism of the professional who seeks tools that work, not merely tools that sound impressive. The philosophical fascination with what creativity means when machines can generate art. The anxiety about what is lost when we outsource thinking. Each perspective is valid. Each adds a necessary nuance. In the autumn of 2025, I use these tools daily. They write code for me, generate images for me, research subjects for me, edit my texts. But the most valuable thing they have given me is not increased efficiency or expanded capability. It is a sharper awareness of my own thinking. When I know I can delegate the search for information, I must ask myself: which questions are worth posing? When I know I can generate hundreds of visual variations, I must ask myself: what exactly do I want to say through an image? When I know I can ask the AI to write in my place, I must ask myself: what is so personal that it cannot be delegated? These tools do not take my thinking away — they oblige me to think more clearly about what is worth thinking about. And perhaps this is the central paradox of the AI era: at the very moment when machines become more capable of thought, we must become more precise about what distinctively human thinking means. We can no longer get away with vague answers. Machines are very good at vagueness now. What remains is the unrepeatable specificity of the individual perspective. The question that only you can ask. The connection that only you can see. The value that only you can articulate. This is the territory worth defending. Not against the machines — but alongside them, using them to clarify what we do best. The autumn of 2025 will become outdated. My map will be superseded. But the process of charting — of honest acknowledgement of limits, of pragmatic search for solutions, of integrating diverse perspectives — this process remains valid. It is the only way I know of navigating a constantly changing territory: with curiosity, with humility, with openness to surprise, and with clarity about what truly matters. --- ## Bibliography and Online Resources ### Primary Sources Consulted **Official documentation:** - Anthropic. 'Introducing Claude Sonnet 4.5.' 29 September 2025. https://www.anthropic.com/news/claude-sonnet-4-5 - Anthropic. 'Introducing Claude Haiku 4.5.' 15 October 2025. https://www.anthropic.com/news/claude-haiku-4-5 - Anthropic. 'What's new in Claude 4.5 – Claude Docs.' https://docs.claude.com/en/docs/about-claude/models/whats-new-claude-4-5 - OpenAI. 'Introducing GPT-5.' Accessed 17 October 2025. https://openai.com/index/introducing-gpt-5/ - Amazon Web Services. 'Introducing Claude Sonnet 4.5 in Amazon Bedrock.' 29 September 2025. https://aws.amazon.com/blogs/aws/introducing-claude-sonnet-4-5-in-amazon-bedrock-anthropics-most-intelligent-model-best-for-coding-and-complex-agents/ **Independent analyses and tests:** - Willison, Simon. 'Claude Sonnet 4.5 is probably the "best coding model in the world" (at least for now).' 29 September 2025. https://simonwillison.net/2025/Sep/29/claude-sonnet-4-5/ - Leanware. 'Claude Sonnet 4.5: Features, Benchmarks & Pricing Guide (2025).' October 2025. https://www.leanware.co/insights/claude-sonnet-4-5-overview - Botpress. 'Everything you should know about GPT-5 [September 2025].' https://botpress.com/blog/everything-you-should-know-about-gpt-5 - Chatbase. 'How to access and use GPT-5 [2025].' https://www.chatbase.co/blog/gpt-5 **Press articles:** - CNBC. 'Anthropic launches Claude Sonnet 4.5, its latest AI model that's "more of a colleague".' 29 September 2025. https://www.cnbc.com/2025/09/29/anthropic-claude-ai-sonnet-4-5.html - CNBC. 'Anthropic launches Claude Haiku 4.5, a smaller, cheaper AI model.' 15 October 2025. https://www.cnbc.com/2025/10/15/anthropic-claude-haiku-4-5-ai.html - CNBC. 'OpenAI launches new GPT-5 model for all ChatGPT users.' 7 August 2025. https://www.cnbc.com/2025/08/07/openai-launches-gpt-5-model-for-all-chatgpt-users.html - CNBC. 'OpenAI's GPT-5 is gaining where it matters most: Enterprise.' 14 August 2025. https://www.cnbc.com/2025/08/14/gpt-5-openai-ai-enterprise.html - VentureBeat. 'Anthropic is giving away its powerful Claude Haiku 4.5 AI for free to take on OpenAI.' 15 October 2025. https://venturebeat.com/ai/anthropic-is-giving-away-its-powerful-claude-haiku-4-5-ai-for-free-to-take - InfoQ. 'Claude Sonnet 4.5 Tops SWE-Bench Verified, Extends Coding Focus beyond 30 Hours.' 11 October 2025. https://www.infoq.com/news/2025/10/claude-sonnet-4-5/ - MarkTechPost. 'Anthropic Launches Claude Haiku 4.5: Small AI Model that Delivers Sonnet-4-Level Coding Performance.' 15 October 2025. https://www.marktechpost.com/2025/10/15/anthropic-launches-claude-haiku-4-5-small-ai-model-that-delivers-sonnet-4-level-coding-performance-at-one-third-the-cost-and-more-than-twice-the-speed/ **Romanian-language sources:** - StartupCafe. 'OpenAI (ChatGPT) a lansat modelul de inteligență artificială GPT-5.' 8 August 2025. https://startupcafe.ro/openai-chatgpt-model-inteligenta-artificiala-gpt-5-85014 - Mediafax. 'OpenAI lansează cel mai avansat model de inteligență artificială de până acum, GPT-5.' 7 August 2025. https://www.mediafax.ro/tehnologie/openai-lanseaza-cel-mai-avansat-model-de-inteligenta-artificiala-de-pana-acum-gpt-5-23592687 - Playtech. 'Viitorul inteligenței artificiale: GPT-4.5 și GPT-5 se anunță în premieră.' 21 February 2025. https://playtech.ro/2025/viitorul-inteligentei-artificiale-gpt-4-5-si-gpt-5-se-anunta-in-premiera/ - Aleph News. 'OpenAI se pregătește să lanseze GPT-5 în august.' 25 July 2025. https://alephnews.ro/tehnologie/openai-se-pregateste-sa-lanseze-gpt-5-in-august-cu-ce-difera-acesta-fata-de-modelele-anterioare/ ### Linguistic Verification Tools Used - **Dicționarul ortografic, ortoepic și morfologic al limbii române (DOOM)** [The Romanian Language Orthographic, Orthoepic and Morphological Dictionary]: https://doom.lingv.ro/ - **Dicționarul explicativ al limbii române (DEX Online)** [The Romanian Explanatory Dictionary Online]: https://dexonline.ro/ ### Methodological Notes All technical information about models (benchmark scores, specific capabilities, pricing) has been verified against the official documentation of the producing companies. Release dates and product characteristics have been confirmed through multiple independent sources. Personal perspectives and usage experiences reflect the author's practice in the period from March 2024 to October 2025. Subjective evaluations (for example, 'the best for...') are clearly marked as such and reflect specific use cases, not universal objective claims. --- **Article date to this point**: 17 October 2025 --- --- ## May 2026 Update Six months later, the battlefield has shifted *Added: 17 May 2026* --- Re-reading what I wrote in October, I have a strange feeling — rather like reading the campaign diary of a general who believed he knew where the front line was. I am not ashamed of what I wrote then. It was true. It was everything I knew. But in six months, the industry has advanced so rapidly that some paragraphs seem written from a different century. I need to return to them with a pencil in hand. --- ### What Has Changed at Anthropic: Sonnet 4.6 and the Lesson About Pricing In February 2026, Anthropic launched Claude Sonnet 4.6, and for the first time since I have been following this industry, I felt I truly understood what was happening beneath the surface of the marketing announcements. This model is not a simple iteration. It is a strategic move, thought through with patience, like the move of a chess player who has waited five turns before opening the diagonal. Sonnet 4.6 reaches Opus-class performance at a price five times lower — three dollars per million input tokens versus fifteen dollars per million for Opus. And for teams running AI agents with millions of calls per day, this difference is not incremental. It is transformational. What struck me personally, as a user who has been following the evolution of Claude since Sonnet 4.5 appeared, is that users in Claude Code preferred Sonnet 4.6 over Sonnet 4.5 at a rate of 70%, and — what strikes me as extraordinary — they preferred it even over Opus 4.5, the flagship model from November 2025, in 59% of cases. They reported fewer hallucinations, fewer false success assertions, and better continuity on multi-step tasks. Sonnet 4.6 brings capabilities that previously required an Opus-class model — including on real tasks in the domain of office work — at the price of a Sonnet model. My father, the mechanic, would have said that someone found a lorry engine in a Dacia body. [Translator's note: The Dacia is a Romanian-made car, historically associated with economy and modest specification.] I do not know whether the comparison is a felicitous one, but I know that I understand it. I tested Sonnet 4.6 on the same type of projects I was describing in October — repository analyses, editorial brainstorming, structuring of complex articles. What surprised me was that the model 'understands context better before modifying code and consolidates common logic rather than duplicating it.' On long sessions, you feel as though you are interacting with someone who has noted what you said two hours ago — not with an assistant who resets with every message. There is also Claude Opus 4.6, launched on 5 February 2026, with new functionality for agent teams, and Claude Opus 4.7, launched on 16 April 2026, described as the most capable model available to the public, with a major leap in agentic programming capabilities compared to Opus 4.6. I remain with Sonnet 4.6 for daily use. It is like the difference between buying the most expensive tool in the shop and discovering that the one on the middle shelf does the job just as well — sometimes better — and costs five times less. --- ### What Has Changed at OpenAI: A Race With No Finishing Line If in October I was speaking of GPT-5 as the great novelty, the picture now looks different. On 5 February 2026, OpenAI launched GPT-5.3-Codex, positioned as a direct rival to Opus 4.6 and specialised in code generation, navigating repositories, and simultaneous debugging. OpenAI describes it as 'the first model that contributed instrumentally to its own creation.' But the story did not stop there. On 23 April 2026, GPT-5.5 appeared, codenamed 'Spud' — the first OpenAI model built from scratch for agentic workflows, not an incremental update of the previous architecture. It was launched with GPT-5.5 Thinking and GPT-5.5 Pro variants, with no access for users on the free plan. I admit that I have a problem with this pace. It is not fascination — it is a kind of fatigue. GPT-5.5 wins clearly on tasks with ultra-long contexts, with 74% on the MRCR v2 benchmark versus 32% for Claude Opus 4.7. If you work with massive codebases or documentary corpora of hundreds of thousands of tokens, GPT-5.5 is in a different category. But for what I do daily — writing, editorial structuring, code analysis on medium-sized projects — I do not feel that the leap justifies changing my habits. --- ### Gemini 3 and the Composure of One Who Does Not Shout First Google did not shout the loudest during this period, but it advanced consistently. Gemini 3 performed well enough in the autumn of 2025 to trigger an internal 'code red' at OpenAI, hastening the launch of GPT-5.2. Marc Benioff of Salesforce publicly switched from ChatGPT to Gemini 3 after just two hours of testing. Now, in May 2026, Gemini 3.1 Pro is being compared directly with GPT-5.5 on coding, reasoning, and agentic capabilities. Its strength remains its integration within the Google ecosystem and applications with high visual density. I continue to use it as a 'beta reader' for longer articles. It has not disappointed me. But nor has it surprised me. --- ### What Has Not Changed: The Need to Choose With Discernment As I write these lines, I think that in October I was speaking about the feeling of being overwhelmed. It has not disappeared. It has stratified. I now know better what I do not know — and that is, probably, the only form of progress I can honestly claim. The battlefield I was charting then has expanded. The front line has shifted. But the question that mattered then — how do you choose the right tools without losing yourself in the noise of weekly launches — is just as valid in May 2026. Perhaps more so. Anthropic has also launched the 'advisor' tool in beta, which allows a Sonnet or Haiku model to act as executor, whilst Opus provides guidance on demand. The strategy combines Opus-level intelligence with Sonnet-level costs — an approach that summarises, in fact, the direction of the entire industry: more power, at a price that more people can afford. My father repaired engines with tools that were decades old. Not because better tools did not exist, but because he knew those ones. He knew exactly how much to tighten a bolt without stripping the thread. That is my only fixed reference point in all of this race: it does not matter which model is 'the newest'. What matters is which one you know well enough not to over-tighten. --- **Update**: 17 May 2026 --- ##     ---