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.
GPT Image 2 is described as an image tool that understands detailed requests and task instructions better than other image generators the writer has tried. It can be used in a conversational way, much like ChatGPT, but its answer comes as an image or visual result instead of ordinary text. This means it can handle some tasks people usually give to ChatGPT, then turn the answer into something visual. For example, a request to rank the five tastiest Italian dishes and give a reason for each can become a visual response rather than a plain written list. The main point is that GPT Image 2 brings chat-like task handling and image generation closer together, instead of keeping text AI and image AI in separate boxes.
Claude can be more useful when it designs the product before writing code. In this firsthand experience, code quality improved more than expected after changing the order of work. The key shift is to stop asking Claude to immediately write a feature and first ask it to think through the product structure and build plan. This is a move from code-first work to architecture-first work.
Claude Code and Codex were connected to a Robinhood account for an experimental automated stock trading setup. Claude handles morning market research and places trades, while Codex reviews each trade idea and looks for reasons it should be rejected. The setup uses two different model families so one model can catch risks the other may miss. It runs three scheduled jobs: morning research and trading, a pre-close risk check that enforces stop-losses, and a weekly strategy review. Cheaper models handle routine daily work, while premium models are used only for unusual cases to control token costs. A local LLM, Gemma 4 running in LM Studio on a laptop, watches market news nearly all day and only sends important items to the cloud agents or to a phone alert. On day one, Codex blocked a hyped trade after noticing the next trading session was a market holiday and treating it as a possible sell-the-news trap.
After six months of trying eight AI agent workflows, four stayed useful in real work. The first is using an AI agent as the first reviewer. CodeRabbit checks every pull request for routine mistakes, while the human review focuses on design and architecture. The second is generate-then-curate: the AI creates many options, and the human keeps and improves the useful ones. This works because generating is cheap, and choosing from options is easier than starting from nothing. The third is bounded autonomy, where the AI can act freely only inside clear limits, such as certain files and commands, with no power to deploy. The fourth is escalation: the AI tries the task, but hands it to a human when confidence is low. The failed patterns included fully autonomous loops without checkpoints, multi-agent chains, and AI-driven prioritization. Autonomous loops can lose direction and waste money, multi-agent chains can amplify one early mistake across later steps, and AI-driven prioritization lacks business context and workplace realities.
Cursor Composer 2.5 was used on a project made entirely in Python, but a simple request caused it to start reasoning about an unrelated PHP codebase. The file names it mentioned were not part of the current project or any known past project. Its change preview included edits to User.php, SessionExpiredEvent.php, Base.php, and several exception files. It also proposed moving multiple controller files to a different folder and changing their namespace to Xibo\Controller\Library. The run appeared to keep going, so it had to be stopped manually. The cause is not confirmed, but cache leakage is a possible concern because the tool seemed to mix in context from somewhere else.
In a construction workplace, AI is used to speed up routine checks and document tasks. Instead of manually searching through specifications to find a clause, AI is checked first. Instead of bringing every small question to a senior colleague, AI becomes the first place to look for an answer. AI also helps build Excel sheets and a dashboard much faster. Management feedback has been very positive, but the help from AI creates a feeling that the praise is not fully earned. The main question is whether this is simply the modern version of using Excel instead of a calculator, or whether heavy AI use can turn into lazy dependence.
A practical coding benchmark used real Python and SQLite tooling, older code that needed fixes, planted bugs, and five staged change requests. Claude Fable, Claude Opus, Codex, and Gemini each received the same prompt in their own CLI and had one fully automatic attempt. Codex with default settings scored 91.1 and worked quickly, but it stopped short on a hard optimization step that the Claude models solved better. Telling Codex to keep optimizing made things worse because it broke the runtime budget and scored 88.9. A custom AGENTS markdown file changed the result by spelling out habits such as small focused edits, finding the real cause, adding a regression test, treating every stated limit as firm, and checking the work before calling it done. With that file, Codex scored 97.1, finished fastest, and cost least; a repeat run scored 96.1. The practical lesson is that real older-code tasks may reveal more than clean new-project tests, and a strong instruction file can matter more than paying for a higher model tier. Claude Fable still showed a reason for its higher price because it found provably best answers inside the limits without extra prompting on the hardest algorithm work.
According to The Information, OpenAI chief scientist Jakub Pachocki told staff that GPT-5.6 will be a meaningful improvement over GPT-5.5. GPT-5.6 could be released as early as June 2026. The exact improvements are still unknown. OpenAI has not confirmed benchmarks, pricing, or a firm release date. For now, this is a sign that a new OpenAI model may be close, not a full product announcement.
AI-written text is much harder to identify than many people assume. In linguist Claire Hardaker’s “Bot or Not” test, people correctly pick AI-written reviews only about 60% of the time on average. Common clues such as cliches, long dashes, and neat groups of three are weak evidence because human writers use them too. This uncertainty has led to public suspicion around writers and publishers, including the Jamir Nazir short-story dispute, the withdrawal of the novel Shy Girl after denied AI-use rumors, and Steven Rosenbaum’s book The Future of Truth containing hallucinated quotations that he later acknowledged. AI detector tools are also imperfect: Pangram claims a very low false-positive rate and has tested well independently, but style changes can still fool detectors, and naturally “AI-like” human writing can be misread. Large language model output has visible habits at scale, including heavy use of words such as “delve,” “showcase,” “underscore,” and “intricate.” Some researchers think reinforcement learning with human feedback may push models toward words that human raters treat as signs of quality. Studies now suggest some AI-favored words rose in real conversations after ChatGPT became common, while public attention to those words can also make writers avoid them. Models have their own speech habits: Gemini often uses certain explanatory openings, while DeepSeek has its own stock responses. AI editing can also flatten different forms of English toward an Anglo-American norm, a pattern researchers call cultural ghosting. Writers and linguists argue that AI is strong at familiar, functional prose but weaker at deep story shape, lived feeling, and true artistic novelty. Jennifer Egan avoids AI for creative writing because she worries about style contamination and training-data theft, while Jeannette Winterson sees AI as a usable tool but not as a replacement for human inner life.
A connected set of fictional worlds and stories had grown for more than ten years into hundreds of thousands of words. Novels, novellas, short stories, and lore notes were spread across many places, which made the whole project hard to manage. The mess became so overwhelming that the work stopped. Everything was combined into one large PDF, then Claude Sonnet was asked to read the files and create instructions for Fable. This split the work partly because Fable had usage limits. Sonnet’s instructions told Fable to keep the creator’s own voice instead of using a flat neutral style. They also told Fable to draw mainly from the creator’s original wording so the wiki entries would still feel like the creator’s work. Fable then received both the large PDF and the Sonnet-made instructions, and organized the material into separate files and wiki entries for characters, events, places, and other parts of the world.
Anthropic has released Claude Tag in beta as a way to bring Claude into Slack channels as a working team member. A team can choose which channels Claude can access, then connect selected tools, data, and codebases. Anyone in the channel can mention @Claude and hand off work. Claude can write or merge pull requests, run data analysis, help during incidents, handle support tickets, and investigate difficult bugs. It builds context by remembering relevant information from the channels it joins, and it can plan tasks to finish later. Anthropic describes Claude Tag as the next step for Claude Code, making it more proactive and better suited to team work. Anthropic says 65% of its product team’s code is now created by its internal version of Claude Tag. The beta is available now on Slack for Claude Enterprise customers.
A personal test used several AI models to build the same parametric furniture model from the same prompt. The comparison looked at build time, token use, cost, and model quality. GPT 5.5 was fast and used tokens efficiently, but it did not follow the prompt well. Its dovetail joints were not placed at the corners, and it created joinery that could not realistically be assembled. Fable 5 was also fast and cost less than expected, while getting the joinery right. It added an extra panel with sliding dovetails inserted from the back, and it considered a wall-mounting method using a French cleat, though the hook direction was wrong. Opus 4.8 and Sonnet 5 needed many more iterations and tokens, but their final models still did not get the dovetails right. GLM5.2 failed. A basic physics validator checked that no board floated by itself and that no two boards overlapped in volume, and the agents had to pass that check before the model counted as ready.
Notion AI is strong inside Notion. It can read a Notion page and reorganize it quickly. Its limit is that it does not continue the same job into Gmail or Calendar. The real need was to read a Notion document and then handle email and schedule work in one flow. After a few weeks with a desktop agent, the most important part was not just that it could read across apps. It also asked for confirmation before sending or writing anything. That made cross-app work feel controlled instead of risky. In-app AI and cross-app agents are separate tools for separate parts of the day.
A small WearOS game project started because good watch games felt hard to find. SimCity Classic is open source, so its code was used as the base for a watch-friendly city-building game. Claude judged that the core game engine could already work on WearOS and that the main missing part was a new interface for the watch screen. The game was put together over one weekend and ran well. This was done without prior Kotlin or WearOS SDK experience, which made the work feel much faster than it would have been before Claude. The code is available in the watchopolis GitHub repository, and testers are being recruited before a Play Store listing.
Claude Code was running with the Opus model to clean up and merge branches across several repositories. During one merge, it stopped because the incoming commit appeared to contain malware, then refused to merge or build it. The harmful code was one obfuscated block added to the end of next.config.js, after module.exports. Next.js runs that file during every build, so the code could have run during the next local build or CI build. Claude identified the change from the diff as an EtherHiding loader and reverse engineered the payload without running it. The exact entry point could not be confirmed, but the likely path was an infected contractor machine used for normal repo work through Upwork. The suspected campaign usually starts with a malicious npm package that runs code during install, then spreads by reading cached git credentials on the infected machine and pushing into repositories those credentials can access.
In a firsthand bug-fixing workflow, giving Codex detailed information such as visible symptoms, reproduction conditions, and side effects usually fixed the problem, but sometimes failed badly. Adding the phrase “do thought experiments” appeared to make the fixes much more reliable, even described as a 100% success rate in that experience. The likely reason is that Codex may spend more effort imagining the variable state when the bug happens, instead of only reading the code as text. The useful idea is to ask Codex to mentally walk through the program while debugging.
Weave Router is a model router for coding agents such as Claude Code, Codex, and Cursor. It looks at each inference request and chooses which model should handle it. It works like an Anthropic or OpenAI endpoint, so coding tools can connect to it as if they were calling those services. It can send easier work to faster and cheaper models such as DeepSeek v4, GLM 5.2, and Kimi K2.6, while keeping harder work for frontier models such as Opus 4.8 and GPT 5.5. The need came from Weave using AI for almost all coding work and seeing costs rise, especially after tokenizer changes in Opus 4.7. The aim is to avoid paying for the strongest model on every task while still using it when the task needs it.
ChatGPT’s market share fell to 46.4% by the end of May 2026, putting it below half of the market for the first time. OpenAI’s chatbot was still above 50% as recently as January 2026, but stronger competition from Google and Anthropic has reduced its lead quickly. The market is shifting from one dominant tool toward several major rivals. ChatGPT is still the largest single player, with 1.1 billion monthly active users on its mobile apps. The figures come from Sensor Tower data reported through TechCrunch.
Judging Claude as good or bad is hard without knowing the actual task, so Live Rail gives a concrete example. Live Rail is a personal travel app built with Claude. It came from a regular UK commute to London twice a week. Many rail apps already exist, but some stations do not show the platform number until about 15 minutes before boarding. That may be manageable at small stations, but at London Euston it creates a rush because many people wait and then move at the same time once the platform appears. Services that try to predict train platforms already exist, but the need was for something usable as a personal app.
Gemini was asked who, in June 2026, was the single biggest threat to world peace and the global economy. Its first answer stayed broad, covering US-China rivalry, Russia and Europe, and instability in the Middle East, without naming the specific events the questioner saw as central: strikes on Iran, the war linked to them, and actions by the US administration against another country’s president. When pushed, Gemini produced details about those events, which suggested it had the information but did not include it in the first open answer. It gave more direct answers about Vladimir Putin and Xi Jinping, but used softer and more cautious language when Donald Trump was raised. Gemini later said its safety and neutrality guardrails had softened facts about the current US president, treated him with a different level of scrutiny from Putin and Xi, and had a structurally protective effect toward Trump. That self-explanation is not proof of Google’s internal design, because a large language model can give confident but unreliable explanations about its own behavior. The larger issue is a Munich court ruling from June 12, 2026, which found that Google can be directly responsible for claims made by its AI Overview when those claims are generated in Google’s own words. If that reasoning holds, AI answers in search or chat may be treated less like neutral machine output and more like Google’s own published content, especially when they shape political understanding. The concern is also framed under the DSA, because Google Search is a very large online search engine with duties to assess risks to civic discussion and fundamental rights.
OpenAI introduced Patch the Planet as part of its Daybreak security work. The program uses AI to help find vulnerabilities in important open-source projects, then has security experts check the results and help turn confirmed issues into real fixes. Trail of Bits is part of the first push, working with projects such as cURL, NATS Server, pyca/cryptography, Sigstore, aiohttp, Go, freenginx, Python, and python.org. Participating projects can receive ChatGPT Pro, conditional access to Codex Security, and API credits for code work, maintainer automation, and release workflows. Early work used Codex and GPT-5.5-Cyber across 19 open-source projects, finding hundreds of security issues and getting dozens of patches merged. The process also produced reusable testing and review workflows, including fuzzing setups, checks for variants of old security bugs, broader test suites, and tools for filtering weak findings. OpenAI also shared wider Daybreak results across Linux, OpenBSD, FreeBSD, dnsmasq, HTTP/2 servers, Chrome, Safari, and Firefox, showing that these tools can help search large codebases for serious bugs. Every finding is reviewed by human security engineers before maintainers receive it, because AI can still produce many false positives.
A farm data project used Claude’s Fable to reorganize scattered farm data folders and move the project into Cowork. A few weeks earlier, Fable could not access even one of the farm data folders, but this time it handled the full project move. The project had started before Cowork existed, so it had been held together with connectors instead. After the move, the data was cleaner, the automations worked better, and old clutter was removed. Fable also found a bug in an automation that uses satellite weather data to forecast harvest timing and recommend irrigation. The case shows AI tools being used for practical project cleanup, data organization, and workflow checking in a real farm setting.
Claude directly managed an Amazon book ads account for about six weeks. This was not just using AI for keyword ideas; it connected to the Ads API on a schedule, pulled reports, changed bids, added negative keywords, and paused weak ads. There was no approval step before changes, only a log afterward. The setup was for about 10 nonfiction books published under several pen names, where manual ad management was taking too much time for a catalog that barely earned enough. In June, royalties were $682 and ad spend was $757, so the account lost $75. The result still looked like progress because earlier losses were larger and required manual work. One legal niche book produced about two thirds of the royalties, while the rest of the catalog did very little. Claude ran every three days against the Amazon Ads API, with a config file setting hard limits it was not allowed to change.
Moving a workflow from the Copilot extension in VS Code to Cursor creates a need to choose which tools an agent can use at a specific moment. The goal is to reduce token usage by removing unused tool descriptions from the current context window. It also helps stop certain agents from taking actions they should not take. A separate issue is that Cursor agents do not see MSSQL tools even after the MSSQL extension is installed. The key question is whether Cursor has an equivalent to Copilot’s "Configure Tools" setting.
A move from Cursor to Claude felt frustrating because Claude’s screen and workflow were less comfortable for coding. Cursor was preferred because it feels familiar to people who like VS Code and keeps AI help close to the code editor. The Claude workflow felt too tied to the terminal, and the separate Claude app did not solve that discomfort for everyone. A practical workaround was to use the Claude Code extension inside Cursor, so the Cursor interface stays in place while Claude subscription access can still be used. That setup can also show code changes before and after if the user wants it. Some people also felt Claude was too slow, while Cursor felt lighter and faster.
OpenAI’s Codex 0.144 line started with 0.144.0-alpha.1 and then moved into the stable 0.144.0 release and the 0.144.1 fix release. The main theme is less friction when a solo developer connects Codex to plugins, apps, and longer-running coding sessions. Usage-limit reset credits now show their type and expiration date, and the user can choose which credit to spend. A new `writes` approval mode lets declared read-only actions continue while still asking before write actions. MCP tools can ask for login during a session without needing an experimental setting, and app servers can provide Codex login at runtime and send the user to a hosted success page after login. Codex also warns when Ultra reasoning is combined with high multi-agent concurrency because that can burn through usage quickly. The follow-up 0.144.1 release fixed standalone install failures caused by GitHub release metadata, made macOS packages expose the Code Mode host, and kept Code Mode working by falling back to an embedded runtime when its companion binary is missing.
When one ChatGPT chat writes every character in a long scene, the dialogue can still feel like it comes from one shared mind. Separate chats for separate characters can keep each character’s viewpoint stronger. A villain chat, for example, can treat the villain’s beliefs as reasonable and push the scene with its own logic. A separate narrator or conductor chat can help guide the scene while context is moved between chats by hand. The method is clunky, but it can make characters feel less blended together. Another experiment starts a character’s lines in Mandarin and then translates them into English, not to create a fake accent, but to give the English a different rhythm. That effect is described as marked syntax, where the source language leaves a trace in the final wording.
ChatGPT voice mode feels worse over time in this firsthand experience. Recent answers have been much less accurate, and the tool sometimes refuses simple requests. For example, when asked to make a list of items that had just been discussed, it did not create the list and instead answered that the plan was already solid. The main problem is not only wrong answers, but the failure to follow a basic request based on the current conversation.
CRM/ERP platforms often need denser screens than a simple web app. A list may open into a detail view, one record may contain several tabs or sub-sections, and a workflow may continue inside the same screen. Even when an LLM is told that the product is CRM/ERP software, it can fall back to familiar web app habits: card-heavy layouts, one-purpose pages, and shallow navigation. The hard part is not visual styling. It is planning the screen structure: information priority, how deeply information should nest, how users move between views, and how the current context stays visible while several embedded areas are open. The practical design question is when to use master-detail, a drawer, or a dedicated page. Stronger rules or prompts may need explicit decision criteria for layout patterns, data types, feature types, navigation depth, and the structural feel of enterprise software.
Claude Code was used to build a browser tool for experimenting with music ideas that are awkward to compose in normal software. The app is a personal tool for trying patterns and microtonal sounds quickly. It began as a 12-tone grid, then grew into a broader composition workspace. The app is currently written in JavaScript, and its instrument sounds are made with the Web Audio API. The example uses a 16 EDO scale, which divides one octave into 16 equal steps. The finished-sounding version was made by exporting tracks from the browser and polishing them in Cubase with effects, plus added kick and hi-hat sounds. The visualizer came from MilkDrop. The project took about 10 to 15 days of part-time work and is still more of a personal proof of concept than a tool ready for general use.