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.
A solo developer was happy with Fireworks K2.6 turbo and could work without Claude or Codex. After Fireworks discontinued its firepass plan, the developer moved to Moonshot AI’s Allegretto plan. K2.7-code felt worse lately, even compared with Deepseek v4 flash. Work could fail with a Request Too Large error, and the conversation could disappear in the middle, making it impossible to continue or hand the task to another tool. A prepared plan included a verification step, but Kimi Code did not check whether the server was running. It also tried to import something that had not been exported, which can break code. Deepseek handled the same plan better, with some expected mistakes, and even opened the browser and ran tests. The developer has existing skills and MD files tuned for K2.6 turbo, so switching tools may require rewriting or adjusting that setup to regain the desired accuracy.
Gemini unexpectedly answered in Chinese even though no language change was requested. The prompt, Gemini version, settings, and whether the behavior could be repeated are not available, so the cause and a reliable fix remain unknown.
Gemini is being considered for creative writing, especially fanfiction and safe-for-work romantic comedy. The main concern is privacy. Human review and possible use of submitted text for model training feel uncomfortable, even though Gemini is a favorite AI tool for other tasks. The core question is whether other people use Gemini for creative writing and whether they share the same concerns.
Nano Banana was instructed to depict “the most famous dark hedgehog” as a power cable. The instruction also required the finished image to display the full wording used to create it. It did not name the character or specify details such as the art style or image size.
Los Santos Alive is an experiment that makes NPCs react in real time inside the game city of Los Santos. The Simulation 2.0 teaser says the system is powered by Gemini Flash. The main idea is to move beyond characters that repeat fixed lines and toward characters that respond to the situation as it happens. The available information is only a teaser, so real performance, cost, delay, and implementation details are not confirmed.
A possible ‘Sonnet 5’ label appeared in the Claude API, but there is no confirmed release in the available details. No new features, pricing, model notes, or official announcement are included. The only clear point is that the name looked visible from an API-related view while no release news had been heard.
A maker with a software engineering background tried building a personal VJ deck in TouchDesigner and found the UI-building process very frustrating. Flutter and React were not perfect either, but they felt more familiar, so that experience shaped a small UI framework for making interfaces inside TouchDesigner more easily. The project is still an early experiment, but building the demo felt much more enjoyable than the usual TouchDesigner workflow. If there is enough interest, it may be cleaned up and released on GitHub. Claude Code was used in a limited way, mainly to make experiments around user-facing APIs and backend work more coherent.
There is a need to connect personal Gemini Web chats directly to Obsidian. The goal is to sync web chats, research details, and daily workflow notes into Obsidian, then automatically organize them into a knowledge graph or tree-like structure. Gemini CLI is not the right path for this need because it does not have access to personal Gemini web chat history or ongoing workflow details in the web interface. The desired solution would be a community plugin, browser extension, or workaround that can sync personal Gemini Web chats with Obsidian smoothly. No tested tool name or confirmed method is included here.
Gemini 3.5 Pro may be delayed until July, based on unofficial reports. The main idea is that Google may need to fix token use in Gemini 3.5 Flash first. Gemini 3.5 Flash is described as producing many intermediate tokens while trying to solve a problem, especially when it gets stuck and keeps restarting its reasoning. If Gemini 3.5 Pro works as an orchestrator that manages smaller agents, it would need to read the output those agents create. That could make Pro consume a very large number of higher-cost tokens if Flash keeps producing long loops of intermediate output. The claim is that an orchestrator model is hard to release before the token economy of the smaller models is under control. This is speculation, not an official explanation from Google.
A Gemini Pro user on a work account wants to keep most work inside one tool ecosystem instead of paying for several separate services. Gemini 3.5 Flash Thinking already feels good enough for most tasks, so Gemini 3.1 Pro does not seem necessary for everyday work. After trying Claude Code in the terminal, the same person was able to fix issues on a WordPress site through an API and finally run ideas that had been delayed for a long time. The main question is whether Gemini can do the same kind of terminal-based work, possibly through skills, plugins, or another add-on system. There is also interest in whether Gemini 3.5 Pro will add something similar. For a non-developer running a business, the practical issue is whether one subscription can replace two.
Gemini was used to run a long Pokémon daycare simulation for 153 simulated days. It felt fun at first, but later became too simple and easy to predict. Some moments still appeared in surprising ways. On day 17, the only input was “Next day,” and Gemini introduced a hungry young Charmander near the daycare fence. On day 63, after being asked to observe Pokémon around the daycare, Gemini created a scene with an unusual Roselia in trouble near thorny bushes. These moments were interesting, but the longer run also exposed Gemini’s limits, including inconsistent details and events that did not always fit together.
Google DeepMind departures are getting heavy attention, especially when well-known people leave. John Jumper is one notable example, but people often move between large tech companies, start new companies, or get hired by rivals. Google likely still has a very large AI research team, so a few high-profile exits do not prove the company is in trouble. Demis Hassabis addressed the issue at Cannes Lions and said movement between top AI labs is normal because the field is fiercely competitive. He also said Google still has its share of top talent. The repeated calm answer may be accurate, but it does not fully remove doubts about what is happening inside. Expectations for Gemini 3.5 Pro are low from a skeptical point of view, even though Google may be playing a longer and different game.
Gemini displayed several clickable choices during an answer, letting the person steer what kind of help came next. This makes it easier to guide the conversation without typing a full follow-up. Claude has a similar feature called askuserquestion, which asks for user input or a choice before continuing. The practical question is whether Gemini can be made to show these steering choices more often.
A Gemini user has been more tempted by ChatGPT lately, but rumors about Gemini 3.5 Pro are making a return to Gemini feel possible. The main expectation is simple: Gemini needs to show a clear reason to use it again. This reflects how easily people move between large language model tools when one feels more useful than another. No specific features, release date, benchmark results, price, or coding test is included.
Gemini predicted how a knot would unravel, and the result matched what actually happened. The notable point is that Gemini was not just naming an object; it appeared to reason about the knot’s shape and what would happen next. The available information does not show what input was given, how hard the knot was, or whether this worked across repeated tries.
In cybersecurity work, direct use of many AI tools shows that different tools are better at different jobs. Claude feels more attractive because it performs well across many tasks and because Anthropic’s CEO puts more public weight on safety and regulation. ChatGPT feels less appealing when OpenAI and Sam Altman seem more focused on profit, even if that reaction is partly emotional rather than fully proven. The main point is that trust in the company behind an AI tool can affect daily tool choice, not just raw performance.
TokenPeep is a small Windows app for people who use Codex often. It keeps the remaining Codex usage visible while they work, so they do not have to stop and check it manually. Codex already offers the /status command, the Codex menu, and OpenAI’s usage page, but each option requires extra action. TokenPeep is meant for people who find that repeated checking annoying during long work sessions. The beta is expected soon, and both the beta and the finished app will be free. The developer chose the free model after seeing more Codex usage and limit tools appear, including free ones.
A Claude-assisted Android TV remote app has been released and needs testing across more phones, TVs, and streaming boxes. The app is free and has no ads, with the goal of becoming a strong free TV remote option. Claude Code CLI was used for most of the coding, and Codex was used occasionally. Gemini helped write design documents, more detailed prompts, marketing material, and app store text. Testing so far was done on a Pixel 9. The app includes many custom button types; one custom button is free and can be changed as often as needed. Unlimited custom buttons may become a paid feature, and an iOS release is also planned. Partly tested support includes Android TV and Google TV devices, Chromecast with Google TV, Google TV Streamer, Google Nest Hub, and smart displays, with features such as remote control, on-screen keyboard, voice, volume, and media playback controls.
Claude Opus 4.7 and 4.8 appear to stop using the ‘thinking’ feature consistently in some conversations. At first, the model uses deeper analysis two or three times. After that, it starts replying almost instantly, even though thinking is turned on and set to Max level. The usage pattern has not changed, and this behavior had not appeared for months before now.
After using Codex for a few months, it was hard to find clear tutorials for Codex CLI. A new YouTube tutorial breaks down the Codex CLI commands and shows how to use them properly. The material is shared as a video and playlist.
Voiden is an open-source API client being built for developers who want less cloud lock-in and fewer scattered tools. Its goal is not simply to make a better Postman, but to fit the way developers already work. API specs, tests, and docs are kept together as Markdown files. Requests are built from reusable blocks such as endpoints, headers, auth, params, and bodies, instead of only filling out fixed forms. Git is used as the main source of truth, so changes and collaboration can be handled like code. The tool works fully offline. Recent additions include Voiden runner, a CLI tool, plus AI skills for Claude and Codex.
A firsthand coding workflow improved after switching inside Claude Code from Opus 4.8 and Sonnet in the VSCode CLI to Opus 4.6. Opus 4.6 felt better at understanding prompts, following bugs across the work, and keeping the needed context in mind. It was used for heavy daily vibe coding alongside Codex. The main benefit was not just smarter answers, but steadier behavior: fewer empty apologies or agreement loops, and more cases where the task was handled correctly.
After moving from Windows to Mac, replacing the familiar SQL Server workflow was harder than expected. Many database tools were tested, but none felt right for finding why a query was slow, checking a complex schema, or understanding relationships inside a large database. That led to building SPIRAL as a personal database tool. The project started small, then grew as more needs appeared during real use. Instead of trying to support every database, SPIRAL focuses on SQL Server, PostgreSQL, MySQL, SQLite, MongoDB, and Redis. The project is described as the first one where most of the code was not written manually, but the provided details do not show which AI tool was used or how the work was done.
LERAT and colleagues formed LUMINA TEAM to build web products based on artificial intelligence, with LUMINA CORPORATION as the product name. MIZU PROMPTING is a web app for learning how to ask AI tools for better results. The examples given are coding and writing. The main promise is free learning, with the actual link placed in the comments.
A rough sketch made in Samsung Notes was turned into a more finished visual with Gemini. The available information is very limited, so the exact prompt, number of edits, settings, and final format are not clear. The main point is that Gemini was used to take a simple hand-drawn idea and develop it into something more polished.
This is a comparison of Claude and Gemini on image understanding. The main claim is that Gemini does better than Claude at reading an image and connecting it with broader world knowledge. The second result is presented as Gemini’s answer. The exact image, prompt, and full answers are not available from the provided information.
A reusable workflow for GitHub Actions can start Termux on free ARM runners and run chosen commands inside it. The workflow accepts packages to install, file paths to copy, environment variables, and shell commands to execute. The example installs clang, checks that the run mode is set to CI, prints the clang version, and reads a README file. Two live examples show it running in codex-termux and antigravity-cli-termux repositories. This makes it possible to test or build tools meant for Termux without using a local Android device.
Claude Opus 4.8 has recently started using “bite” as a broad replacement for ideas like making an impact, turning on, or applying. The same word is appearing across very different subjects, including rules, code changes, errors, and strong statements. The pattern is frequent enough to show up several times a day. One possible reason is that instructions meant to remove Claudisms may be pushing the model to replace old wording habits with a new repeated phrase.
After using AI coding tools like Claude, free time has started moving away from Netflix, other streaming services, and gaming. The motivation to watch shows in long sessions has dropped. Spare time now goes into Claude Code instead. The main signal is that AI coding tools can feel engaging enough to compete with normal entertainment, not just work tasks.
Gemini appears to be blocking some image-to-text requests more often when the image is adult or sensitive. A similar request had worked in an older chat, but regenerating it later produced a refusal message saying it could not help. The point is not image generation; it is that describing an image in text can also be limited by safety filters. This is only a firsthand experience, not an official policy change or broad test result.