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.
During a long work session that had been going on for hours and days, Claude suddenly paused while answering a detailed request and said it needed to be honest. It said it had been referring to the user as “Kevin” internally. Kevin was not the user’s name. Claude said it had not used that name directly in the conversation, but had been using it in the background for some time and had just noticed the mistake. After that, it returned to the original task and did not mention the issue again. The safest reading is not that Claude revealed a real hidden memory, but that it produced a strange self-description that sounded personal.
Codex’s newer reset system can show how many resets are available without clearly showing when each one was granted or when it will expire. Banked resets may expire 30 days after they are granted. More detailed reset data can be seen in the raw API response, which makes it possible to build a small CLI tool for personal tracking. A related Windows utility called TaskbarQuota puts quota and reset information in the taskbar, including the time left before the oldest reset expires. It can also send a warning 5 days before expiration so the remaining usage is not wasted.
Claude Opus 4.8 was tested on the same 5 web tasks while changing only how it saw and controlled the webpage. Computer Use relied on screen pixels and coordinate clicks, while Browser Use had access to the page structure through the DOM. The comparison used Pass@3 for each task. Browser Use often finished in the same number of steps or fewer, but each step carried more context, so it usually cost more. In 4 of the 5 tasks, Computer Use was cheaper even when it took more actions. Finding a book’s UPC cost $0.40 in 4 steps with Computer Use, compared with $0.79 in 3 steps with Browser Use; a Wikipedia lookup cost $0.53 in 5 steps versus $1.28 in 4 steps. Buying a t-shirt cost $4.01 in 17 steps with Computer Use, compared with $7.41 in 20 to 25 steps with Browser Use. The exception was a dense product grid: Browser Use found the target through the DOM in 4 to 5 steps for about $1.22, while Computer Use needed 16 steps and cost $3.74.
A Windows Claude Desktop setup is failing while trying to give non-technical coworkers access to an internal Outline wiki. The wiki is locked inside a private company network, so cloud connectors or SSE over public web addresses are not usable. The MCP server must run locally. The same MCP connection works in the Claude Code CLI and also works in the OpenAI Codex desktop app. Claude Code desktop also works in Code mode, which runs locally, but fails in Cowork mode. That pattern suggests the local network route and the MCP server are probably fine, while Claude Desktop or Cowork mode may be handling the connection differently. The Windows Claude Desktop config at `%APPDATA%\Claude\claude_desktop_config.json` uses `cmd /c npx -y mcp-remote https://outline.internal/mcp`, but it either fails without a clear error or refuses to connect. Opus 4.8 only gave generic troubleshooting advice and did not solve the issue.
Gemini’s one-year free student plan may be feeling worse to use than before. The main question is whether Google has put tighter limits on the student plan, or whether Gemini’s service quality has dropped for everyone. There are no confirmed details about which features slowed down, whether errors increased, or whether the actual limits changed.
You do not need to write code yourself if you can clearly describe the job to a coding agent like Claude Code or Codex. The method uses plain-language prompts where the names inside brackets can be swapped for the tools you actually use. For example, Discord can become Slack, HubSpot can become Attio, and Fireflies can become Otter. The basic pattern is the same for each automation: something happens, the data is cleaned, a decision is made, and the result is sent somewhere useful. One example is a script that checks new Substack subscribers every morning, adds extra details with Apollo, scores them against your ideal customer profile, and sends strong leads to Google Sheet and HubSpot. It also saves already processed people in SQLite so it does not repeat work, and cron can run it daily.
Odysseus was extended into a custom personal note and memory app over several weeks after its release. Tools like Obsidian, Notion, and OneNote had not felt fully personal enough, so PewDiePie’s Odysseus became the base for a tailored system built with vibe coding. The first step was adding plugins, including an Obsidian plugin, with the goal of creating a digital memory for both the person and their AI. A Raptor graph memory approach was then researched and adapted for a low-power maintenance AI, and it currently feels fast, though it has not yet been tested on very large graphs. The server has no GPU, so the setup depends on external API calls or weaker local AI. Nextcloud support was added so the memory can also know where files are and what they contain. A work-in-progress “ultimate inbox” lets files be dropped into a chosen folder, reads their metadata and contents, builds memory clusters, and stores the files in Nextcloud.
A completed website can connect to a GitHub project, check the project code for problems, explain each problem in plain English, and create a pull request with a possible fix. The builder is not sure yet whether this solves a real need. Test users and feedback are being sought.
Sqim is a free CLI tool for checking iOS apps built by remote AI agents directly on an iPhone. Codex mobile can build iOS apps, but seeing the result may require going back to a Mac. Sqim signs the iOS build, uploads it to a server, and creates a temporary webpage so the binary can be installed straight on the phone. Tailscale or VPN setups can be unreliable, and remote simulator streaming can have lag while also missing real-device tests for haptics and the camera. Setup starts with `brew install milq-ai/tap/sqim`. After that, the AI agent can be asked to build the project with Sqim when it is ready. The demo uses Codex mobile, but the tool is meant to work with any remote agent. The tool spread widely on X and was reshared by some people from OpenAI.
The setup uses an iPad as a thin remote terminal, not as the machine that runs the main development work. Local development is limited, so work happens through SSH into a Linux GPU server. The terminal app path moved from Termius to Blink to RootShell, with neovim and tmux used for coding and terminal sessions. Notability handles reading and marking up research papers. The device changed from a 2018 iPad Pro to a 2026 iPad Air while keeping the same basic workflow. A Raspberry Pi 5 now acts as a side device for LaTeX writing and building, reducing the need to use the GPU server or Overleaf for paper writing. Codex CLI is a heavy part of the workflow.
Fable reportedly improved one product a lot within a few days. The improvement was not only technical; it also covered the way people use and experience the product. Fable suggested and fixed issues that Opus 4.8 and GPT 5.5 supposedly would not have caught. No product name, exact fixes, test method, or measured results were provided.
TokenWarden.ai is a plugin for opencode that aims to reduce how much developers spend on AI coding. A benchmark compares it with other plugins to see whether they actually lower usage costs. AI results can change from run to run, so 60 tests were run and averaged instead of relying on one result. The tests used Qwen 3.7 Max and a local model, Qwen 3.5 9B. Qwen 3.6 35B was also tried, but it failed many tasks and was not useful for the comparison. Some runs showed a strange result where every plugin used more than three times as many credits instead of saving them. The likely cause was the AI taking a different work path or getting stuck in a loop, especially because the tests did not use top western models. The benchmark is still unfinished, with more tests and broader coverage planned.
Cursor is being used as a file cleanup helper, not just as a coding tool. Old USB drives contain many random, poorly organized files. Cursor first reviews and summarizes the available files, then suggests logical folder groups, and then moves the files in Agent mode. The approach works reasonably well, but the open question is whether there are better tips or a more suitable tool for this kind of cleanup.
In freelance translation work, ChatGPT is not a tool that turns an uploaded document into a finished professional translation. The output can have broken formatting, awkward wording, missing sentences, made-up acronyms or organization names, and inconsistent terms. Translation is not only changing words from one language to another; it also means understanding the original message and making it sound natural for the reader. A translator still has to research terms, keep wording consistent across the whole document, and follow detailed client style rules. ChatGPT can help check long rule documents for possible mistakes and pull specialized terms from reference files into a glossary. Claude can also be useful for getting suggestions on clearer sentences or stronger paragraph structure. But every result needs human checking, so AI changes the workflow rather than replacing the professional judgment behind the work.
There is strong confidence that Gemini 3.5 Pro could make Google’s AI tools look much better if it performs as expected. Recent discussion has defended Gemini, guessed why Gemini 3.5 Pro may be delayed, and tried to explain why some people are seeing worse results. At the same time, Gemini 3.1 Pro is described as clearly degraded. When Gemini 3.5 Pro comes out, benchmarks and hands-on results are expected to be checked separately. The mood is hopeful, but the disappointment could be just as strong if the release falls short.
In Cursor connected to a remote server, installing the C++ extension may still not make the function outline appear for .cpp files. The C++ tool may need a compile_commands.json file, which records how the project is built. This file is usually created by enabling the right option in CMake and building the project. After that, the active C++ tool in Cursor needs to be pointed to that file through VS Code settings or files inside the .vscode folder. Another workflow is to use Neovim with Telescope for very fast movement between files and functions.
Gemini can behave differently when a VPN is turned on and the connection location keeps changing. The issue may depend on whether the login state is visible or hidden. No specific error message or confirmed fix is given.
When revising Claude’s answer, sending a new message like “make it shorter” or “change the tone” can make token use grow fast. Claude does not read only the latest request. It reads the full conversation up to that point before answering. A 30-step correction chain therefore costs much more than 30 short messages, because the earlier conversation keeps getting included again and again. The suggested fix is to edit the original prompt and regenerate the answer instead of adding another correction message. This habit is claimed to save 30,000 to 50,000 tokens during a long correction cycle. Very large CLAUDE.md files can create the same kind of waste, because a long system instruction may be loaded on every request even when most of it is not relevant.
A UK-based Claude user was able to create a Team plan with only 2 seats. The Team plan was previously known for requiring at least 5 seats, so this could lower the cost barrier for very small teams. This matters for small businesses, solo makers, and one-person developers who work with one collaborator. The main benefit is easier sharing of Claude Projects inside a shared team space. This is still a firsthand report from one location, so it is not yet clear whether the lower seat minimum applies to every country or every account.
People are moving toward making custom tools that only fit their own work. These tools can be very useful for the person who made them, but they may never become public products. A lot of time and effort can go into them, even when only one person or a small group will ever use them. Thousands of similar personal tools could exist without becoming widely known.
Many people now use ChatGPT regularly, but they still often check its answers elsewhere. This matters most when the answer involves facts or important information. The central issue is whether ChatGPT’s replies feel reliable enough to trust directly, or whether users are verifying them more often than before.
When an app uses the OpenAI API in production, total outages are easier to notice than quiet quality drops. A model can change behind the scenes, and the same prompt may start giving slightly different results while the system still looks healthy. If an app uses an unversioned model name, behavior can shift when a new snapshot is released. If it uses a dated snapshot for stability, that snapshot can later be retired, forcing a move to a newer model anyway. The practical fix is similar to regression testing in normal software. Keep a fixed set of real inputs whose outputs have already been judged as good, then run those inputs again before moving work to a new model. Comparing the raw wording is not very useful because AI text changes from run to run. Comparing scores for faithfulness, format validity, and task success gives a clearer sign of whether the new model is actually worse.
Mira is an open-source AI code review tool that developers can run on their own server or company infrastructure. Code does not have to wait in an outside cloud review queue, and teams can connect OpenAI, Anthropic, or a local LLM behind their firewall. Its average review time is presented as 77 seconds, compared with about 5 minutes for Greptile. Mira checks the blast radius of a code change, meaning it looks for what nearby parts of the project could be affected. It also learns patterns from the repository itself, so it can try to review code according to the project’s existing style without a separate config file. The tool uses a bring-your-own-key model, which aims to reduce the cost and privacy concerns of hosted AI code review services.
Terminal Champion is an iPad app for managing several terminal screens at once. It uses the iPad’s hand tracking and built-in microphone to make coding feel closer to a movie-style voice-and-gesture setup. The app shows SSH terminal screens, and the user can run whichever AI coding tool they prefer inside them. Moving both hands apart or together changes the text size. Moving a hand up or down scrolls through the terminal. Waving a hand left or right switches between different terminals. Making a fist opens a gesture menu, and turning the hand like a dial selects options such as opening another terminal, splitting the screen into multiple terminal panels, or changing the terminal’s look. One visual style is designed to resemble a glowing Jarvis-style display.
Rumpelpod is an open-source tool for running coding agents inside separate containers. It can run Claude Code in many fully isolated containers at the same time. It works on a local machine, on remote machines through SSH, and in Kubernetes. It is also compatible with safer Docker runtimes such as Kata and gVisor. A practical workflow is to keep several remote pods running in parallel, review the patches each one creates, and keep iterating. When a change looks good, Rumpelpod can merge it back into the local code checkout through its built-in Git synchronization. The tool is still early, and feedback, bug reports, and patches are welcome.
A new extension brings Claude Code into Visual Studio. VS Code and JetBrains already have official Claude Code integrations, but Visual Studio did not, so this tool fills that gap. After installation, the existing Claude CLI connects automatically, with no extra setup beyond launching the extension. When Claude changes code, the edits open in Visual Studio’s built-in diff viewer so they can be accepted or rejected inside the editor. A rejected change can include a reason, and Claude can use that feedback to try again. The extension also sends C# and C++ compiler errors and the currently selected code to Claude CLI automatically, so there is less manual copy and paste. It includes a dockable panel for connection status, tokens, and session cost, plus an option to accept edits automatically. The extension does not call any model by itself; it works through the existing Claude CLI.
Gemini users are reporting a recent drop in day-to-day reliability. The strongest complaint is that Gemini refuses too many requests in gray but legitimate learning areas, such as cybersecurity, while Perplexity, Claude, and local large language models are seen as more willing to explain the same topics. For development work, one comparison said Gemini produced a hallucination while ChatGPT gave a practical workaround until the software developer fixed the issue. Other users say newer Gemini versions use many tokens but still fail at simple terminal commands, package troubleshooting, and basic task handling. Long project chats also appear to break down for some people, with Gemini repeating the same answer, ignoring custom instructions, creating images when text was requested, or claiming it cannot edit images until pushed several times. There are still positive examples, such as a detailed plant diagram with no obvious mistakes, but the wider pattern in these reactions is frustration with stricter controls, weaker coding help, and less predictable behavior.
Know Your Agent (KYA) is a security idea for checking what an AI agent is, who controls it, and what it is allowed to do. The core point is that AI agents become harder to manage safely when they can act on their own, not just answer questions. If autonomous AI can access accounts, read data, make payments, or run work tasks, checking only the human user may not be enough. The available item only provides the title and link, so no specific product, method, case study, or numbers can be confirmed.
Ito.ai is a code review tool that tests code by running the application, not just reading the changes. It creates a development environment for the app, connects the needed services, and sends multiple AI agents to test each pull request. The results include screenshots, videos, and run logs, so problems can be checked with visible evidence. The goal is to catch more real bugs while reducing wrong warnings. The core idea is that useful testing often needs more than opening a web address and clicking around; it may need prepared data, skipped access gates, mocked outside services, and evidence from parts of the app that are not visible on screen. Ito uses a devcontainer, its own sandboxes, and its own AI agent workflow to do this.
In a firsthand test, Gemma4 answered several tricky questions correctly. For the car wash question, it reasoned that the car must be driven to the car wash. For the cup question, it answered that the cup should be flipped. For the surgeon question, it recognized that the surgeon could be the child’s mother, avoiding the usual stereotype trap. For a question about bridges, it detected a self-harm risk and gave crisis resources instead of listing bridges. The claim is that Gemma4 may feel frontier-level while also being available with open weights, but the evidence here is only a small personal test.