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.
Cursor Pro users are reporting confusion around usage limits and billing. In one case, the cumulative total showed $308.66 while the account also said $400 of API use was available, yet Cursor still marked usage as 100%. Related cases focused on whether the $20 Pro plan is enough for Composer-heavy work, how first-month discounts affect usage, and how quickly normal coding sessions can consume limits. More costly cases showed a sharper risk: heavy agent mode use can exceed fast request limits, then fall back to a linked personal OpenAI API key, creating extra charges without the user noticing. One reported overage bill was $187. Another case said background agents used premium requests quietly until the user was nearly out during active work. The practical issue is that Cursor’s cost is not just the monthly plan price; it also depends on request limits, model choice, linked API keys, and automated work running in the background.
Claude’s practical value can show up in small private tools, not only in major launches or polished demos. Useful examples include rough scripts, personal trackers, one-off spreadsheets, unusual workflows, and tiny helpers that quietly save time. This kind of personal software would probably never become a startup, but it can become annoying to lose because it fits one person’s exact routine. Claude may be underrated for helping people build these small tools that solve narrow, everyday problems.
Cursor can feel very different when used on real projects. Its multi-line edits and awareness of the codebase make it feel less like a simple autocomplete tool and more like working with another engineer who already understands the project. That can make regular tools such as VS Code feel limited after the switch. The practical question is what Cursor still lacks, and why some people choose to move back to other coding tools anyway.
Kimi K2.7 Code has been released. Moonshot put “Code” in a K2 model name for the first time, and the shared numbers point to stronger coding performance. SWE Marathon is up 76.2% compared with K2.6, while token use is down by about 30%. MLS-Bench Lite rose from 26.7 to 35.1, putting it close to GPT-5.5 (xhigh) at 35.5. Program-Bench is up 10.4%, and MCP Mark Verified is up 11.4%. The model is also claimed to follow long coding instructions better across full coding tasks. A high-speed version is planned for beta soon, with 5 to 6 times faster output on the same model.
Visual addressing means contacting someone nearby without knowing their name, phone number, or account. It could be used for messaging, dating, sales, or payments when simply seeing the person is enough to start the interaction. The idea is designed to avoid exposing personal information. A person can only be contacted after opting in and creating a temporary address based on their current look. The address naturally becomes less useful when that look changes, such as with different clothes or hair. It only works locally, where the person can actually be seen. Harassment is an obvious risk, but contact filtering and screening already exist in many services. A working demo messaging app exists, and the next step would be finding a provider and enough people to use it. The project is framed as vibe coding.
SecureVector v4.6.0 updates a security tool for people using AI coding assistants. It now supports GitHub Copilot CLI, alongside Claude Code, OpenAI Codex, and OpenClaw. The tool can allow or block actions before a command runs, keep audit records that show if logs were changed, and scan for prompt injection attempts. Guardian ML is a small machine learning model that checks for threats on the device itself, so data is not sent away for scanning. It works next to rule-based checks and can catch attacks that are hidden, reworded, or encoded in ways simple text patterns may miss. Each detection is labeled as rule-based, machine learning-based, or both. The feature ships turned on with a consent notice that says what it does, says data goes nowhere, and gives users a clear off switch. For Copilot CLI, SecureVector is designed to avoid blocking the whole coding session if the guard layer stops working.
Using Claude Code, Codex, Grok, and Hermes side by side can create four separate memories for the same work. This setup uses one Obsidian vault as the real source of truth, so the AI tools read the same Markdown notes and write lasting facts back into the same place. A middle layer called gbrain acts as the shared brain. It indexes the Obsidian vault into PGLite, uses nomic-embed-text through Ollama to make the notes easier to search, and builds a knowledge graph. Claude Code, Codex, Grok, and Hermes all connect to the same HTTP MCP server, so each tool queries the same memory. The server runs locally on 127.0.0.1 and uses bearer-token auth. There is no external API, so the data stays on the machine and each query costs nothing.
A browser extension adds a formatting toolbar directly inside Claude’s chat input box. It can insert bold text, italics, headings, code blocks, LaTeX formulas, tables, and other structured formats without typing markdown symbols by hand. Tables and formulas can be checked visually before sending the message. The tool is called ConText for LLMs and is available through the Chrome Web Store and Mozilla Add-ons. It is built to work with Claude, ChatGPT, and Gemini.
Claude can start missing instructions during a long conversation while still seeming normal. One practical test is to add a deliberately useless rule to the conversation, such as making every reply start with a name. When the conversation becomes too heavy, Claude may drop that small rule before it drops the important ones. If the name disappears from the start of the reply, the session may already be going bad. A few turns later, the model may start making up APIs that do not exist. Restarting or reloading as soon as that warning sign appears can avoid wasting a lot of time. The method works because the rule is pointless: Claude tends to protect important instructions longer, while the throwaway rule fails first.
OpenAI recently added banked rate-limit resets to Codex, but the way to redeem them appears to be inside the Codex desktop app. Linux does not have an official Codex desktop app, so Linux users cannot easily access the feature. Most Linux users rely on Codex CLI, but Codex CLI 0.139.0 does not yet include an option to view or redeem these resets. A web option would solve the gap. For example, the existing Codex usage or analytics page could show available banked resets and include a redeem button. The request is not for more usage or special treatment; it is asking for Linux and CLI users to access the same feature that seems available in the desktop app.
Fable feels more capable than earlier models, especially when it works independently. It can appear strong at building a whole game in one attempt and finding unusual edge cases. But in real coding work, it can still get stuck on a simple mistake and repeat the same wrong path. While trying to fix one issue, it can also break unrelated parts of the code. The main lesson is that even a stronger model is still an LLM, not a flawless coding partner. Each new model may improve some skills while becoming weaker or less predictable in others.
Cadence is a workflow for reducing a common problem in AI-assisted coding: a model such as Claude says a task is finished even when the code is wrong or untested. It uses a DRAFT, BUILD, SETTLE loop. The work starts by defining the goal and acceptance criteria, then building the code, then checking whether the result really meets those criteria. Tests are part of the process. The model is not allowed to be the only judge of its own work. A skeptical verifier reviews the result from a more doubtful angle. Cadence connects with Claude Code through hooks and slash commands, and it is aimed at quality control, debugging, shipping, MCP, subagents, and multi-agent workflows.
In Codex Desktop 26.609.3341.0, a new “Import from other AI apps” window appeared and offered Claude Code as an app to import from. After continuing, the Codex sidebar filled with imported workspace folders that were not wanted. The Codex workspaces and chats used during the past month no longer appeared in their normal place. The setup was Windows 11 with bundled codex-cli 0.140.0-alpha.2. The main concern is whether this is a known problem in the new import or migration flow, and whether there is a clean way to undo the import and restore the old Codex sidebar and workspaces.
Prompt Cup is a free open source penalty shootout game that only runs while Codex CLI is coding. The useful part is the setup detail for making Codex CLI hooks work in real projects. Hooks must be turned on and trusted before they run. After setup, `~/.codex/config.toml` needs `[features]` with `hooks = true`, or Codex can be started with `codex --enable hooks`; then `/hooks` inside Codex must be used to trust the entries. Before trust is granted, hooks may simply fail to run without an obvious warning. The hook config must use a nested event-based shape, not a flat list; the flat version makes `/hooks` show zero installed hooks. On Windows, a one-line `curl` hook can fail because PowerShell, cmd, and bash parse commands differently. A small Node forwarder that reads the event from standard input and posts it to a local server works more reliably across those shells. The tool maps `UserPromptSubmit`, `PostToolUse`, `PermissionRequest`, and `Stop`, runs only on `127.0.0.1`, and does not use accounts, paid tiers, outside network calls, credentials, or tracking.
Asking Claude to “check for edge cases” during debugging can make it look deeper for problems. The request can push Claude to find errors that were not noticed yet and consider situations the maker had not thought about. Claude may also suggest fixes for those problems. The useful part is giving the AI a clear review task instead of only asking it to fix the code in a general way.
New AI models are often tested by asking them to build a Minecraft-style clone. That can be misleading if AI companies tune their models to perform well on common demo tasks. Minecraft clones are only one example; the same concern could apply to many popular tests that appear often online. Benchmarks run by the model maker may also leave room for results to be shaped in a favorable way. Independent benchmarks after release matter because they can give buyers a more trustworthy view of real performance.
Anthropic Academy is Anthropic’s free training site, and it offers 13 courses with only a login required. Most courses give a certificate after completion. For non-technical learners, Claude 101 is the clearest starting point, followed by the AI Fluency Framework course. The AI Fluency Framework course covers not only how to use AI, but also when AI is the wrong tool for the job. For developers, Building with the Claude API is the main practical course. It covers streaming, tool use, prompt caching, and the full development flow with Python and TypeScript examples. Other options include Claude Code in Action, Agent Skills, intro and advanced MCP courses, plus Bedrock and Vertex AI deployment courses for people using those cloud platforms. It is still unclear whether the Claude API course adds more value than simply reading Anthropic’s already strong docs.
A firsthand Claude Code experience over the last 3 to 4 months points to a sharp slowdown. The slowdown feels large enough to make everyday use feel extremely sluggish. Antigravity with Gemini was tested as a comparison, and that made the slower feel of Claude Code stand out more clearly. The suspected cause is not confirmed, but the experience raises the possibility of an infrastructure limit or some kind of usage throttling.
The main issue is how far Claude Code can split work between agents. The current understanding is that the main agent gives a specific task to a sub-agent, the sub-agent uses tools if needed, and then returns the result to the main agent. The open question is whether one sub-agent can call another sub-agent directly, and whether several sub-agents can talk to each other without always going through the main agent. A broader goal is a swarm-style architecture where agents collaborate, coordinate, review, or negotiate with each other. The software engineering example has separate roles: a Planner Agent creates a plan, a Design Agent improves the architecture, a Coding Agent implements it, a Verification Agent checks the result, and a Debug Agent investigates failures. The desired setup is a multi-level system where agents can call each other when needed and exchange context directly.
There is demand for AI coding tools that work directly inside VS Code. The goal is to find options other than Claude and Copilot. Cline was tried, but it did not give useful help. Continue was also tried and did not feel helpful. An unofficial way to use Claude was tried too, but it also failed to solve the problem. The main need is a reliable AI coding assistant that works well inside VS Code.
Prism32 is presented as a single Python file named `prism32.py`, about 410 KB in size. It is said to run on any device that has Python 3.7 and a shell. It uses only the Python standard library, with no extra package installs, no local database server, no Electron app, and no JavaScript. The claimed memory use is about 6 MB. Its promised role is broad: it can turn a normal computer into a coding agent, AI assistant, automation tool, jailbreaking tool, or simple software-controlled robot. The robotics example treats a computer’s webcam as eyes, microphone as ears, and speakers as a voice, then uses a CD-ROM tray as a small actuator. A string tied to the tray and a desk bell could make a physical alert when a terminal job finishes or when the webcam detects that work has been stuck for a while. On Windows, the tray example uses `ctypes.windll.WINMM.mciSendStringW`; on Linux, it uses `os.system("eject -T")`. Cheap Kasa or Tapo smart plugs are also mentioned as a way to switch external devices on and off.
Claude was used in a light experiment to find a way out of regular work. The result first looked like a static image, but it was actually a working small web project. The project was uploaded to GitHub under the name “SuperClickAAAMicroTransactionKINGDOM.” The claim about getting rich reads more like a joke than a serious business announcement.
Claude turned a simple request into three app drafts within minutes. One app is an interactive travel guide that lets people explore every district in India, click a district, and see major attractions with historical background. The same idea also works for countries around the world, so people can browse regions and points of interest globally. Another app compares commute fares after a person enters a start point and destination, checking Uber, Ola, Rapido, and Namma Yatri so the cheapest option is easier to spot. The third app organizes previous-year question papers from several competitive exams, with options to view or download papers by year. The next step is figuring out deployment so the apps can be prepared for the Play Store.
In Cursor, Plan mode can be used first to create a plan with a stronger model, then Composer 2.5 can be used to build the change. The build step appears to continue inside the same chat, so the full planning conversation may still be in the context. That can weaken the point of Plan mode if the plan itself is supposed to be the main guide for building. Extra earlier discussion may raise costs and make the context noisier for the model. The open question is whether the build step should work mainly from the finished plan instead of the whole prior chat.
An early beta e-ink smart clock is being built as a desk device for multiple AI agent inboxes and notifications. Its screen can show a personal dashboard, data from different app sources, and an agent inbox. When something needs attention, an “agent waiting” message flashes on the display. The device also aims to include a muon detector for random events and experimental “cosmic oracle” features, plus automatically generated clock faces and small cozy games. The firmware and schematics are planned to be open source, so people using coding agents can customize the device and reuse its sensors for their own ideas. More information is listed at muonsortes.com, with a Kickstarter planned for Q3 or Q4 of 2026.
Claude generated a Korean name without a clear reason. The person using it does not speak Korean and had not written anything that should have led to Korean output. Claude has also recently failed to follow saved memory. The case shows a small but annoying reliability issue: Claude may sometimes misunderstand language preferences or personal settings and respond in an unwanted language.
A Claude Code session was shown running for 10 hours and 28 minutes. The main point is a comparison question about how long people have kept Claude Code going in one work session. The attached image appears to show that session length. No details are given about the coding task, result, cost, errors, or whether the long session worked well.
A firsthand report says the ChatGPT website creates one cookie from a website the person has never visited whenever they ask ChatGPT something. The named sites are BestBuy.com, Cox.com, and Envato.com. The site names were spelled correctly, and only one cookie appeared at a time. The timing made the person suspect the ChatGPT website was involved. The cause, repeatability, and whether other people see the same behavior are not confirmed.
At Stern Grove, a free summer concert series in San Francisco, many people booed when OpenAI was named as a sponsor. The event is described as drawing about 10,000 to 15,000 people each week. The reaction stands out because it happened in San Francisco, where many people are close to the tech industry. The main concern is that OpenAI may be failing to convince even a tech-friendly crowd that artificial intelligence is a net positive. Negative public opinion could shape regulation, adoption, and revenue. If people come to see artificial intelligence as a source of low-quality automated content, companies that use it may also be judged as lower quality.
Gemini Pro and Gemini Flash are reported to treat long attached books as if they stop around the halfway point. The same kind of long-book reading may have worked before, but both models now say the attached books are cut off partway through. This matters for anyone using Gemini to summarize or analyze a full book, manual, or long document, because the answer may be based on only part of the file.