We pick and plainly summarize new features, pricing, usage limits, and policy changes across major AI tools — Claude, ChatGPT·Codex, Gemini, and Cursor — from a solo developer and maker’s point of view.
Music video editing often requires cutting clips exactly on the beat, but manually tapping out every beat and adding markers takes time. Claude Code was used to build BeatMarker, a small Adobe Premiere Pro plugin that automates that job. BeatMarker lets the editor choose a WAV or MP3 file, analyze it, and place colored markers on each beat in the clip. The first beat is marked red, the backbeats are marked blue, and beat three is marked yellow. The editor can then cut directly to those markers. The marker positions can also be shifted with arrow controls. The plugin is free and open source, works on Windows and macOS, and supports Premiere Pro 2025 and 2026. It still has limits: for now, it only supports 4/4 time and music with a steady tempo.
Claude now has Claude Design, and Claude Code can also be extended with many front end design skills. For a beginner trying to build a website for a small business, the two tools can look like slightly different versions of the same thing. Many videos praise new features or skills, but they do not clearly explain when to use Claude Design and when to use Claude Code. The main issue is choosing the right Claude tool before starting real website work.
Some Claude accounts have shown a lower minimum for the Team plan, changing from 5 users to 2 users. This started appearing in reports about 10 days ago, but it is not visible for everyone. Some accounts still show 5 users as the minimum. It is unclear whether the 2-user option is a slow rollout to more accounts or a limited test for only some users.
The team behind Tessl, an AI coding-agent tool, ran about 3,300 coding tasks across four Gemini models and shared the results. Gemini 3.1 Pro scored 87.9 and cost $0.66 per task; Gemini 3.5 Flash scored 88.6 and cost $1.05 per task. The score gap is just 0.7 points, yet Flash cost 59% more. The twist: 3.1 Pro's published per-token price is actually higher than Flash's. The cost reversal comes from how each model works through a problem. On average, 3.1 Pro used 26 conversation turns and roughly 650,000 tokens per task, while Flash used 39 turns and about 1.4 million tokens. Flash pulled in far more context and took more steps, which overwhelmed its lower unit price. A second finding: when the team supplied relevant skills from their registry, 3.1 Pro's cost dropped by roughly 23% and its score improved substantially, while Flash models saw little to no benefit from the same addition.
A product demo can be created by prompting Claude. The tool first scans the product page and figures out what the product does. It then writes a narrative and script for the demo. It automatically records the screen and creates voiceover audio and music. Gemini is used to understand the video and produce editing instructions that Claude can use. The final step combines the edited video, voiceover, music, and zoom-in effects into one finished demo. The tool could be adjusted later to make other demo styles, such as avatar videos where a founder appears to explain the product.
A solo maker who says they cannot code yet used Claude to build QSELF, a personal health dashboard. The tool tracks workouts, food, sleep, habits, mood, body weight, and activity in one place. Daily tracking covers calories, protein, carbs, fats, water, sleep stages, habit and mood check-ins, job checklists, naps, and bodyweight trends. Training features include per-set weight and reps, effort level, personal records, strength charts, muscle group volume, recovery indicators, deload detection, and an exercise library. It can log running, cycling, swimming, and other sessions, and it can auto-import data from Amazfit/Zepp watches, including sleep, heart rate, HRV, blood oxygen, and steps. It also uses a LLM to create an AI coaching report from the person’s own data and can compare two tracked variables against each other.
Economist and writer Tyler Cowen uses ChatGPT Pro as a broad tool for learning, research, and judgment. He reads two to three books a day, writes daily, has traveled to about 105 countries, and calls himself an “infovore,” meaning someone who constantly consumes information. ChatGPT Pro helps him ask many kinds of questions faster than normal web search. His examples include critiquing a 1978 macroeconomics paper by Robert J. Barro, finding reading material on why baseball pitchers now throw faster, planning a five-day trip in northern Ghana, checking local safety near a hotel in Sao Paulo, improving how he looks at Mondrian paintings, choosing works to see at the Detroit Institute of Arts, analyzing a short chess game, explaining why early David Burliuk paintings are worth more than later ones, finding concerts in Paris, and learning how to listen to Sibelius’s Symphony No. 7. Cowen sees ChatGPT Pro as close to pocket-calculator reliability for many of these uses. Because he has published online for more than 20 years and has many public podcast conversations, he also feels AI systems can understand his tastes unusually well.
A Claude artifact, such as a calculator, small game, or simple webpage, normally stays inside the chat. To send it around, it needs a public link or a separate web host. Claude’s Publish button is the fastest option on Free, Pro, and Max plans, and people can open the link without a Claude account. The downsides are practical: unpublishing cannot be undone, so there is no easy way to take the page down, fix it, and restore the same link. It is also awkward to keep one link pointing to the newest version while still editing. The feature is controlled by Anthropic, so its behavior can change. Team and Enterprise plans cannot publish artifacts publicly. Netlify Drop and GitHub Pages are free and stable options, but Netlify Drop expects a folder rather than one HTML file, and GitHub Pages requires setting up a repository. nippy.host is presented as another option where HTML can be pasted, kept as index.html, and published at a chosen address such as my-thing.nippy.site.
Claude and similar AI tools can create a working prototype very quickly. That does not mean the same thing as running a real SaaS product for paying users. Real systems face problems such as a database outage, many people using the service at once, slow parts of the app, and rising cloud costs. A build-by-prompt approach can get something working, but it may not prepare the product for failures and edge cases. Software engineering is not only writing code; it also includes choosing the right structure, understanding trade-offs, planning for scale, and controlling costs. AI can speed up coding, but major design choices still need human judgment and experience.
Claude Code’s 5-hour usage window can shrink very quickly in long coding sessions because of how its to-do tracking may interact with caching. When Claude Code uses TodoWrite, it adds a reminder block for the task list near the front of the conversation. Each time an item is marked done, added, or changed, that block can be written again. Because the block sits in the part of the prompt that is normally cached, changing it can break the prompt cache. The next request may then process the whole long context again, sometimes hundreds of thousands of tokens, even when the visible answer is tiny. In a long session with many task updates, this can repeat many times and burn through usage quickly. Related reactions describe sudden jumps in Claude Max limits, Claude desktop usage after an update, and Claude Code sessions where usage moved far faster than expected. The wider lesson also applies to Codex and other coding tools: large codebases and broad context can become expensive unless the tool reads only the files and lines it truly needs.
Google is beginning to roll out a new agent feature within its Search AI Mode. These agents are built to persistently track specific information for the user and remember data across different search sessions.
Codex users are seeing confusing differences between the desktop app and the CLI. On the Mac app, even a low or light usage setting appeared to burn through the 5-hour limit after only 2 or 3 prompts, while the same kind of work in the CLI did not seem to reduce the limit as quickly. Codebase reading tools, memory setup, and MCP were present in both places, so the difference cannot be clearly blamed on one helper tool from the available information. The same discussion connects to wider uncertainty about Codex resets, whether limits are counted over 5 hours or by week, and whether writing-heavy workflows should spend Codex quota at all. Some developers are also comparing this with Claude Code through Microsoft Foundry, where usage can be paid from credits instead of a normal subscription cap.
Gemini can now create meeting notes inside Google Meet. The feature is available to Google AI Pro and Google AI Ultra subscribers. It can reduce the need to write notes by hand during a call. The confirmed details are the supported app, Google Meet, and the paid subscriber groups that can use it.
An anesthesiologist with no programming, computer science, or AI background started using Claude in March 2026 and is building an anesthesia and emergency resuscitation simulator as a side project. The goal is not to copy expensive hospital equipment perfectly, but to make a free or low-cost simulator that many people can use. The current version uses two web pages on the same network. One page works as the patient monitor, and the other works as the instructor control screen. It can run on a two-monitor setup now, and the longer-term plan is to support two separate devices, such as iPads. The waveform display still has accuracy problems that need work. A future goal is a free repository where other people can create, load, and share their own training scenarios.
Croco is a desktop app for keeping side projects in one place. It helps track where each project lives, what was last being worked on, and what still needs to be done. It includes 35 project templates for web apps, backend projects, command-line tools, Minecraft mods, Discord bots, and more. The app can open a project in an IDE, handle commit and push actions, and create GitHub repos automatically. Each project can have its own todos, notes, and activity log, so work is easier to resume later. It also supports project-specific AI chat with Claude, Codex, Gemini, and Ollama. Storage can use SQLite or JSON, and the app includes 20 themes plus keyboard shortcuts. It is free to download, still in beta, and currently works only on Windows.
This F1 title odds simulator uses real standings and race results to estimate each driver’s chance of winning the season. It runs a Monte Carlo simulation for the remaining races. Driver settings can be changed, and the odds update in real time. Historical race results since 2020 are also available. The service updates automatically every Monday. It is free, open source, has no ads, and does not require signup. The odds may be wrong, so they should not be used for serious betting decisions.
Anthropic costs can rise sharply when a company passes 150 seats. After that point, the company must move to the Enterprise tier, and seats no longer include usage. Every token is then charged at standard API rates. At the current pace, the yearly bill would rise from $400,000 to $1.4 million, about 3.5 times higher. One Claude Code user spent $4,000 in three days by accident, and some top users on the support team are reaching $800 per month. For engineering work, AI coding agents may still be worth the price because they can save more time and money than cheaper tools. For other roles, the value is less clear when people build unused apps or duplicate skills that already exist. Showing each person their own spending makes the cost real, and spending limits are expected next.
Google AI Pro was canceled, AI Plus was restored, and Claude Pro became the main web-based AI chat tool. Gemini had been attractive because Gemini 2.5 Pro showed a strong comeback, and Google’s research background made its future look promising. Gemini also had generous usage limits, deep links with Google services, fast replies, and strong multimodal ability. It worked well for daily questions, photo-based questions, and hard science problems. For someone using a Google Pixel and Google Cloud, Google AI Pro also added Google Drive storage and made phone access to AI convenient. The weak point was the Gemini app and web version. They lacked useful everyday features such as selecting and quoting text and branching conversations, and they loaded slowly. A recent interface change still felt out of step with Material Design and brought bugs in areas such as mobile table display and code blocks.
AI tool use is spreading inside a late-stage startup, but the day-to-day setup is messy rather than organized. The company is under pressure to keep growing 15% every quarter, and workers who have lived through reorganizations and layoffs treat AI as a way to protect their jobs. AI has also made the company’s physical product more important in the networking space, which pushed one internal manager into a broader AI rollout role. The main problem is that many people built their own unofficial tool setups without IT oversight. The current fix is custom training, Google Drive folders full of markdown notes, and reusable template skills made with Claude. Sensitive access details, such as server connections, Cloud Run deployment, and database access, are being kept in a 1Password vault so people can work in a more secure setup. The hardest part is teaching about 40 people who are mainly motivated by job security while also keeping normal work moving.
jdbg is a debugger interface built so AI coding agents can inspect a running Java program directly. The problem it targets is familiar in agent-assisted coding: agents often guess from error traces, add many temporary print statements, or repeat the same failing command without learning what is happening inside the program. jdbg wraps the JDK’s built-in jdb and makes it easier for agents to control. It supports persistent debug sessions, breakpoints, stepping, local variables, stack frames, watches, threads, and locks. It includes MCP tools and setup paths for Claude Code, Codex, and OpenCode. It is designed to work without shell-script glue or temporary files, with Windows as the first target while still aiming to be cross-platform. The main idea is that an agent should use runtime facts from the JVM instead of guessing the state of a Java app from logs alone.
Vidilearn is an open-source command-line tool that extracts YouTube transcripts, subtitles, chapters, articles, and structured metadata on a local computer. It does not require API keys, so it aims to reduce recurring costs for content collection. When modern JavaScript-heavy websites block simpler extraction tools, it can use Playwright as a fallback to open pages and collect content more reliably. The extracted content can feed AI agents, RAG pipelines, automation systems, Codex or Gemini CLI workflows, and semantic search stacks. Its features include YouTube transcript extraction, article cleaning, structured metadata, local embedding generation, and MCP server mode. Shared benchmark numbers include a 94.2% RAG hit rate, 92.1% precision, a 0.931 F1 score, and a 10.0 cost-adjusted score. Claude is described as still slightly better in absolute accuracy, while Vidilearn gets close at near-zero operating cost. It can be installed with `npm i vidilearn`, and the GitHub repository is Alfo-Tech-Lab/vidilearn.
A MacBook can pause long work when the lid is closed. That can interrupt a long compile, a large download, or an overnight AI coding task. Hold My Lid is presented as an app for keeping the Mac awake while tools such as Claude Code, Codex, OpenCode, Cursor, Gemini CLI, and Aider are still working. It claims to use lifecycle hooks and process detection, so it checks whether an AI coding agent is actively doing work rather than only checking whether an app is open. It also includes battery controls, such as stopping at a chosen battery level, running only while plugged in, and respecting Low Power Mode. It can turn off the display while the work continues, and it can send a sound and banner notification when the agents finish.
A larger coding job can be broken into high-level beads first. These beads work like a backlog of planned tasks. Superpowers is then used with Opus 4.8 on X-High or Max to brainstorm implementation plans for as many beads as needed. Once the implementation plan is ready, the work is not started in that same session. The session is exited, the context is cleared, and the setup is switched to Ultracode. Ultracode is then asked to create a dynamic workflow that can carry out the implementation plan. The main idea is to separate planning from execution so a long Claude Code session does not become overloaded with earlier discussion.
Claude and similar AI tools can turn an idea into something that works in a day or two. Turning that working idea into a real product takes much longer. Product work needs many rounds of checks, testing on different devices, and review from other people. A first version can feel exciting, but it still needs repeated fixing before real users can rely on it. The hard part of vibe coding is putting aside the fun rush and doing the boring work until the product holds up.
llm-council-mcp-server is a local tool for getting more than one AI-style view on a hard question before turning the result into one answer. It works as an MCP server inside coding agents such as Claude Code, Codex CLI, and Antigravity. The free internal council mode does not need an OpenRouter key; the current agent imitates five viewpoints: a pragmatist, an architect, a skeptic, a clean-code reviewer, and a product or UX thinker. The full multi-model mode uses OpenRouter and follows three steps: separate answers from different models, anonymous review of those answers, and a final synthesis by a chairman model. It can run from PyPI through uvx, so there is no need to clone the repository. A browser setup screen lets users change the API key, models, chairman model, and temperature settings. The docs and setup screen are available in English and German, and the tool is still early and mainly focused on Mac for now.
hypequery is a code-based analytics tool for TypeScript developers building on ClickHouse. Metrics and dimensions can be defined once and reused across dashboards, apps, and APIs. It does not require adopting a separate analytics platform or maintaining large YAML configuration files. The definitions compile into ClickHouse SQL, so there is no extra service, proxy, or runtime to deploy. It is designed for multi-tenant products, with query-layer safeguards that block cross-tenant data access and helpers for connecting to an existing authentication system. A dataset is declared as a fixed list of allowed dimensions and measures, so it can also act as a safe catalog for AI agents. The package includes an MCP server that can give a large language model a typed catalog to query. A serve() entry point can expose a dataset as an HTTP endpoint, letting the same type-safe definitions power dashboards and APIs.
A ChatGPT Pro account used for two years was deleted without warning. The account was used for work and household help, with no explicit requests or explicit language claimed. There was no email notice before the deletion. OpenAI support first suggested changing settings, then asked for another appeal, but that appeal was rejected because a similar case was already open. Later replies pointed back to an article and another appeal form. Even after the issue was supposedly escalated to a human, the replies did not give a clear reason or fix. Similar cases appeared on Reddit, so the practical warning is to avoid keeping important work only inside ChatGPT and to download anything needed later.
Researchers from Boston Children’s Hospital, Harvard University, and OpenAI rechecked 376 rare disease cases that had stayed unsolved after earlier expert review. The cases included children with neurodevelopmental conditions, people with rare neuromuscular disease, children and teens with early psychosis, and pediatric sudden unexpected death cases. OpenAI o3 Deep Research connected symptoms, family data, genomic variant tables, and scientific papers to suggest possible causes for experts to review. After medical review, extra testing, and clinical lab confirmation, physicians confirmed 18 diagnoses, adding a 4.8% diagnostic yield. The results were 10 of 100 neurodevelopmental cases, 4 of 61 neuromuscular cases, 2 of 200 pediatric sudden death cases, and 2 of 15 early psychosis cases. The model did not diagnose patients or make medical decisions. It produced leads that qualified experts checked through normal clinical processes. The study did not measure time saved, cost, doctor workload, false positives, or changes in care.
Claude was used as a coding and research helper to build a 135 million parameter looped language model from the ground up. The model was trained on 4.6 billion tokens from FineWeb, with experiments inspired by the Parcae paper on making looped language models more stable. Five comparison tests took about two weeks to debug, but at this small size, the more complex stability methods did not beat the simple baseline. The training was still completed, followed by supervised fine-tuning, and both the base model and tuned model were released on Hugging Face. Training used two H100 graphics cards on Modal plus free Lightning H200 time, took about three hours, and cost about $51 in total. Claude helped check the paper implementation, find optimizer routing mistakes, write the Modal training setup, and get the model shipped.
Long-time Claude accounts may lose their old subscription price when moving to a lower plan. A Max 5x subscriber upgraded to Max 20x, then asked for a refund and a return to Max 5x less than a day later after running into a usage block. Anthropic handled the request by canceling the Max 20x plan because it said it could not directly process a downgrade. After that cancellation, the old 100 euro monthly price was gone. Restarting Max 5x now costs 137 euros per month for European customers. Anthropic’s explanation was that the old 100 euro rate was an older localized price, while a canceled and restarted plan uses the current price based on the dollar rate plus local taxes. A short upgrade and rollback can therefore turn into a permanent price increase.