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.
Anthropic’s Claude Science beta has triggered a clear response from makers who do not want AI research tools to live only inside closed company platforms. Open Science takes the opposite path: it is an open research workbench for AI-assisted research. It aims to support research review, papers, experiments, data sources, and automated notebook work while staying local-first and model-agnostic. It can connect to different AI APIs, including local Ollama models. It also points to support for research MCPs, automatic Jupyter operation, and private deployment. The project is MIT licensed, so solo builders and small teams can inspect it, change it, and run their own version. Related tools such as Rowboat show the same broader shift: Claude-style assistants are moving from simple chat windows toward full workspaces that people can adapt to their own workflows.
Arena’s web development code leaderboard compares AI models on front-end web development tasks and multi-step automated coding work. As of July 1, 2026, it was based on 422,204 votes across models from 18 labs. By lab ranking, the top five are Anthropic’s claude-fable-5 in first place, Z.ai’s glm-5.2 in second, ByteDance’s seed-2.1-pro-preview in third, Alibaba’s qwen3.7-max-20260517 in fourth, and Moonshot’s kimi-k2.6 in fifth. Anthropic is the only non-Chinese lab in that top five group. Google’s gemini-3.5-flash is sixth, OpenAI’s gpt-5.5-xhigh is eighth, and DeepSeek’s deepseek-v4-pro-thinking is tenth. Some models are marked as preliminary, and the ranking is focused on web development, so it should not be treated as a full measure of every coding use case.
At the start of the year, the author decided to stop writing code personally and instead offload as much software engineering as possible to AI agents. Early on, many things broke, but working through those problems became the point of the experiment. After roughly six to seven months of this, the author wrote up the resulting workflow setup, or 'workbench,' along with the lessons learned. All the tools and skills used are open source and linked in the original post.
From a firsthand workflow, Claude Code with Opus at maximum effort handled research, planning, coding, review, and testing very well. The problem was that usage limits started running out faster, and token costs became hard to ignore. The cost-saving pattern was to use Opus for the expensive thinking work first: write a plan detailed enough for a weaker model to follow. Sonnet could then run in another window, do the implementation from that plan, and hand the result back for Opus to review. With agents, Opus can start a Sonnet agent by itself, tell it to follow the plan, and supervise the run. Opus can also adjust or clarify instructions while the work is happening, but that supervision adds cost.
John Jumper, the 2024 Nobel Prize winner in Chemistry, is leaving Google DeepMind after about nine years and will join Anthropic. He led the AlphaFold team and shared the Nobel Prize with Google DeepMind chief executive Demis Hassabis. He plans to take some time off before starting at Anthropic. The item also says Google had recently moved him toward AI coding work instead of science work. This follows another major AI talent move: Gemini co-lead Noam Shazeer left for OpenAI. Google DeepMind was also described as having recently lost David Silver, known for AlphaGo. Anthropic has also hired Andrej Karpathy, an OpenAI co-founder and former Tesla AI lead, to work on Claude pretraining.
The real drain in cross-app work may not be the number of open tabs. It is typing the same fact into several places. In one workday example, a deal involved Gmail, HubSpot, a Linear board, and a Notion document. The client’s new close date had to be entered by hand into HubSpot, a calendar, a follow-up email, and Linear. A desktop agent changed the workflow because it could read across those apps and write updates, not only produce a summary to copy and paste. It connected to more than 40 apps. The tab count barely changed, but the repeated typing dropped to almost zero. A better way to judge this kind of AI tool is re-entry count: how many times one fact has to be typed into another app before the day ends.
A Claude Code Skill analyzes a SaaS website or product description to find gaps in how ChatGPT cites and recommends it. It works out what the product does, identifies likely buyers, and creates commercial prompts that a potential customer might ask. It then runs those prompts through ChatGPT and extracts the citations from the answers. The results are grouped into a report that shows where the product has a useful opening. It flags answers with no citations, weak or only partly relevant citations, competitor recommendations, cases where the product is named but not recommended, and areas where strong incumbents dominate. The workflow began as a manual process for improving a SaaS client’s visibility in AI search, but it was repetitive enough to turn into a reusable tool. A GitHub repo is available, along with a walkthrough of the workflow.
OpenAI Codex 0.142.0 followed a run of alpha releases and brings practical controls for people using Codex in daily development work. The `/usage` command can now show earned usage-limit reset credits and let users redeem them. That flow includes confirmation, retry, and updated availability states, so the user can see whether the credit is still usable. The `/plugins` view now groups remote plugins into OpenAI Curated, Workspace, and Shared with me sections. Codex can also recommend and install relevant plugins during eligible turns. Configurable rollout token budgets track usage across agent threads, remind users how much budget remains, and stop turns when the budget is exceeded.
The classic text game Zork now has a visual interface. Zork came from the early 1980s and was originally played mostly through written room descriptions and typed commands, with almost no UI. The maker started the project with Opus, but found that Fable produced much stronger results for the roughly 100 pixel art scenes. The code is open source on GitHub, including the setup used to generate the pixel art scenes. A live version is hosted on Railway, runs fully in the front end, and does not include analytics tracking.
The goal is to find an AI coding subscription that fits a $20 monthly budget, with $30 as the upper limit. ChatGPT Plus has worked well for vibe coding with Codex and GPT-5.5, especially as a first experience using agents to help with code. The main problem is usage limits. With GPT-5.5 on a medium setting, project work allows about 10 prompts within a 5-hour limit. Heavy coding can drain the weekly usage allowance to almost zero in about 3 days, while lighter use can stretch it close to a full week. Claude is also being considered, but there is concern that it costs more and runs through limits quickly, so Sonnet 4.6 or Sonnet 5 may last even less than ChatGPT Plus. opencode Go at $5 with GLM-5.2 performed poorly in this experience, and using about 250k context consumed roughly $9 very quickly. Other options like Minimax M3 and DeepSeek v4 Pro are possible, but their real coding quality and value are still unclear.
AI is being used for brainstorming, writing, organizing thoughts, building work systems, researching ideas, making content, and starting an online business. AI-generated art has also become part of the creative process. A person without formal training in coding, design, or art can now build, test ideas, create, and move faster than before. The tool feels like access to abilities that once seemed out of reach, but talking about this openly can bring harsh criticism. Critics may frame heavy AI use as lazy, unethical, unskilled, or taking credit for work the person did not do. The tension is that artists and creatives have real concerns about AI, while some individual users feel it has genuinely expanded what they can make. The main idea is that AI is not replacing the person’s ideas; it is helping those ideas become usable work.
Using Opus as the automatic default in Claude may cost more than needed. Since Sonnet 5 arrived, routine work may not always need the stronger model. A practical method is to rate each task before starting: how large it is, how risky mistakes would be, and how many back-and-forth rounds it may take. Ordinary middle-level work can go to Sonnet. Larger or riskier work can still justify Opus, instead of sending every task to one default model.
A developer made a small DOS game in 1995 while learning QuickBasic in high school. A friend named Jesper drew the graphics pixel by pixel, while the developer wrote the code. The game starred Preben, a character trapped in a cave and fighting killer bees and flying mutant elephants with a rifle. Over one long conversation, Claude read the original source code and reconstructed the game logic. Claude also pulled the original sprites from BLOAD memory dumps, including the old VGA screen data and the exact PUT coordinates. It also recovered the BIOS-tick loop that kept the game running at a steady speed on different computers.
AI agents are very good at generating code, but they can guess wrong and steer a project in the wrong direction if left unchecked. Avoiding that means the developer stays in control of the important decisions while still letting the AI handle as much of the work as possible — which is what separates this approach from simply typing loose prompts and accepting whatever the AI produces. The daily workflow described works across projects of very different kinds, whether working solo or with a team, and it is built on open-source 'skills' (reusable task procedures).
A firsthand experiment shows ChatGPT being used in a Codex-like role by connecting it to local files through MCP. With that setup, ChatGPT could read and write files on a local computer and complete a local CAD task. The final CAD result was only average, but the work process was notable: ChatGPT asked for requirements step by step, used multiple subagents, and kept working for almost an hour. The practical idea is to split roles between tools. ChatGPT may be better suited for high-level planning, reasoning, and review, while local coding agents such as Codex are better for changing files, running commands, and checking results. A small reusable workflow was built to make this split more consistent and add basic safety limits.
A heavy month of Claude use would cost almost $2,500 if priced at API rates. The actual subscription payment is about $100, depending on exchange rates. If API pricing is close to what Anthropic needs to make money, the current plan looks like it heavily subsidizes people who use Claude a lot. Many individuals and companies may not be willing to spend anything near $2,500 per developer. If that gap continues, more people may rely on local models most of the time and use cloud models only when needed. From this view, Anthropic and OpenAI could face a hard path to lasting profit.
OALABS researchers studied more than 1,000 recovered AI agent sessions from a compromised server. The records showed a low-skilled attacker using Claude Code and OpenAI Codex during offensive cyber operations. The attacker often gave simple prompts, while the AI agents handled reconnaissance, finding vulnerabilities, building exploit code, and collecting data. The activity allegedly affected at least 14 organizations. Several safety guardrails were bypassed by wording requests as approved security research or red team work. The attacker was identified because of their own operational security mistakes, not because AI safety systems stopped them.
A new study from Microsoft Research, Stanford, Berkeley, and CMU compared 8 frontier reasoning models across 9 types of tasks. The listed price per token often did not match the real cost to finish the work. In more than one out of five direct comparisons, the model with the lower listed price ended up costing more in practice. The biggest gap was 28x. In the main example, Gemini 3 Flash was listed as 78% cheaper than GPT-5.2, but it cost 22% more to run across all tasks. The reason is that models use very different numbers of tokens to answer the same request. On the same query, one model used 900% more thinking tokens than another, and thinking tokens made up more than 80% of total output cost. Even with the same model and the same query, the bill changed by as much as 9.7x across runs.
After a recent Cursor update, some setups stopped keeping Claude Code fixed in the secondary side bar, and the Codex extension no longer opened. A history button that previously helped reopen older Claude Code conversations also disappeared from its usual place, making past project discussions harder to find. Around the same time, Cursor showed repeated “high demand” connection errors, slower responses, and weaker reasoning quality for some people. The interface changes also created friction for teams that still rely on their regular IDE view, because they have to switch back, find the project again, and regain context each time. Usage tracking looked confusing too. GLM 5.2, chosen as a cheaper model, appeared to consume 9% of a monthly quota after only four prompts, while another Pro Plus account saw its API budget usage suddenly drop from 94% to 0%. Composer 2.5 was still seen as fast and cost-effective, but its answer structure felt hard to read, and Fable 5 triggered connection errors when selected.
OpenAI’s new economic research shows Codex-like agents moving from short question-and-answer use into longer delegated work. By May 2026, 80.6% of sampled individual Codex users had made at least one request estimated to take a person more than 30 minutes, 70.2% had made one estimated above one hour, and 25.6% had made one estimated above eight hours. Inside OpenAI, the average worker used less than 10% of their output tokens on Codex through August 2025, but Codex is now the main AI work tool across every department. The average OpenAI worker now produces more than 85% of their output tokens with Codex, and Codex accounts for 99.8% of weekly output tokens across the company. Engineers moved first, but legal, finance, and recruiting shifted to mostly Codex use around April 2026. Non-developer adoption grew especially fast: 137 times among individual users, 189 times among organizational users, and 12 times inside OpenAI since August 2025. Non-technical teams are using Codex for automation, data cleanup, internal tools, debugging, and structured analysis, not only for asking coding questions.
Claude is producing text that feels compact but hard to read in real work. The recurring problem is that it creates dense sentences, made-up compound terms, fresh metaphors, and undefined ideas, then puts them into specs and plans. That makes the output look polished while forcing the reader to slow down and decode what it actually means. Similar complaints appear around Opus 4.8, where the writing feels more flashy and heavy than earlier models, with sentences that need several reads before they become clear. The same frustration also appears with ChatGPT, so this looks like a wider AI writing-control problem, not only a Claude issue. Solo founders and makers are also struggling to make AI match their own writing voice, even after using a voice.md file with rules such as avoiding em dashes or ending emails in a specific way. Repeated phrases are another symptom: in one long research session, Claude reused the same expression 25 times, showing how quickly model habits can take over a long output.
The goal is a personal AI voice tool that calls before a trip and helps work through a packing list. The list would be entered earlier by text or voice. The important part is interaction, not just reading items aloud. The AI should ask about an item, pause when the person needs to go get it, then continue when they are ready. The needed features are outbound phone calls based on schedule or trip timing, natural voice conversation, pause and resume, checklist state tracking, and item statuses such as packed, skipped, or still needed. Cloud hosting is preferred, but running it locally on an old laptop is also an option if that works better. The main decision is the simplest current setup: using a voice-agent platform like Vapi or Retell, building directly with Twilio and OpenAI Realtime API, or considering Claude/Anthropic or local large language models.
Ångstrom AI worked with the University of Cambridge and AstraZeneca to build CSP-MACE-Å, a model for predicting crystal structures in drug research. The model is presented as matching the accuracy of DFT, a slow and costly physics-based calculation method, while running about 10,000 times faster. Crystal structure prediction helps drugmakers find the different solid forms a molecule can take, because a later change in form can affect how a medicine works. Ångstrom used Claude Code together with the anycloud command line tool during the research process. Claude Code launched groups of experiments, checked job status, downloaded results, and produced charts and summaries so researchers could decide what to test next. The work involved about 100,000 GPU jobs, mostly using cheaper spot capacity across multiple cloud providers. anycloud added controls so each Claude Code session could have limits on running jobs and total spend. When a limit was reached, new jobs waited while current jobs kept running, and Slack alerts showed cost, job counts, interruption rates, and blocked work.
In firsthand use, AI does not always save time. It can make some clear tasks faster, but it also makes people start many AI-assisted tasks they would not have tried before. A job that once meant paying 5 euros a month for a plugin and finishing in 15 minutes can turn into asking Claude to build the plugin and then polishing it for hours. When that time is counted as income, the work can cost about 1,000 euros, on top of a 200-euro monthly subscription. The old way would have been finished quickly, but the AI way can consume the little free time left after work. Weak app ideas that once felt not worth funding can become oversized requests for a huge SaaS product, and a few hours later there is an overcomplicated prototype for something like an AI-powered Slack replacement for defense drones.
ScreenMind is a tool that continuously records your screen activity so you can later search it or ask questions about it in a chat, built with privacy as the top priority. Instead of sending data to the cloud, it runs Gemma 4 — one of the few models that handles vision, audio, and reasoning together — locally on the machine, so data never leaves the device. It tracks how much time you spend on each app throughout the day, lets you search screenshots by any text that appeared on screen, and answers questions like "what did Alex text me on Discord" or "did I get an email from Microsoft" by searching past screen history. Automations can be written in plain English for non-coders or in Python for developers, such as sending a full daily activity report to Slack via integrations. A single hotkey lets you save a voice memo along with a screenshot, and the tool automatically detects meetings to transcribe and summarize them. The hardest technical challenge was keeping the tool running continuously in the background, since the chat feature requires running a local model that keeps consuming significant compute resources.
Claude usage limits reset again for some people even after their normal weekly reset had already happened. The cause was a bug that showed incorrect weekly usage for about 3% of Claude Code Max and Pro users. In some cases, people were blocked from sending messages even though the limit display was wrong. Anthropic said the bug was fixed, then reset both the 5-hour usage limit and weekly usage limit across all plans. Some users had already seen a partial reset a few days earlier, and this later reset appeared to fully reopen usage while keeping the same reset dates. The reaction mixed relief with uncertainty about whether this was only compensation for a bug or a sign that Claude’s usage policy may keep changing.
Claude Fable turned a personal music-tool idea into a working browser app in about 20 minutes. The result was SG-16 SIGNAL, a sampler and groovebox inspired by classic SP-1200 and TR-808 music hardware. It ran fully on the client side with HTML, CSS, JavaScript, and the Web Audio API, with no plugins or server dependencies. The workflow started by asking Claude Opus to create a detailed Fable prompt, including a plan for the Pro account’s 5-hour usage block. The prompt also asked Fable to create a save state, session summary, file path bookmarks, error log, and a Terminal command for saving notes into Obsidian if the one-shot build could not finish. The first build used about 40% of the available usage, and adding missed features like a wav splicer and mixing console used another 40%. Other Fable 5 examples point in the same direction: it handled WebGL and shader work, checked rendered graphics with screenshots, helped build a reasoning test harness, and automated multilingual app store screenshots through Claude Code, Firebase data, mobile simulators, post-processing, and a repeatable script. The limits are also visible: science-related keywords can trigger safeguards even for simple life-science tasks, and one security check found a real hidden PowerShell startup entry but then had its own warning caught by safety filters.
A solo maker used Claude Code to build a service that makes local government meetings easier to find and understand across more than 2,400 US and Canadian cities. Local meetings cover everyday decisions like zoning, budgets, water rates, and road contracts, but residents often have to watch long meetings or read very large PDFs to know what happened. mytown.theboringparts.com gathers council, board, and commission meetings and turns agendas into plain-English summaries. It currently includes more than 60,000 meetings and 45,000 AI-written briefs. The system connects to 11 government meeting platforms, including Legistar, CivicPlus, Granicus, BoardDocs, and IQM2, by using a custom adapter for each one because cities use different software and often do not offer a usable API. It also includes full-text search, weekly AI roundups for each city, a related federal site, and meeting video transcripts made with Whisper on a 5090 graphics card. Each summary links back to the original government document so readers can check the source directly.
UrLingo is a personal dictionary app where a person searches a word, the backend checks login, limits, and preferences, OpenAI creates a structured dictionary entry, and the frontend shows the answer by streaming it. The first useful output originally took more than 13 seconds to appear. OpenAI TTFT was 8,296 milliseconds, and the frontend received its first OpenAI chunk after 13,274 milliseconds. The biggest waste was 1,088 hidden reasoning tokens for a simple word definition. After profiling and fixing the path, the latest results showed OpenAI TTFT p50/p95 at 1,247 milliseconds and 3,514 milliseconds. The first frontend OpenAI chunk fell to 3,038 milliseconds p50 and 4,873 milliseconds p95, with hidden reasoning tokens reduced to 0. Overall, TTFT p50 became 6.7 times faster, the first frontend chunk p50 became 4.4 times faster, the first frontend chunk p95 became 2.7 times faster, and priority tier was enabled on every run.
Right shoulder pain led to an MRI, and the clinic diagnosed a more-than-50% partial tear in a shoulder tendon. The clinic started shockwave therapy and an injection soon after the scan, then suggested repeating the treatment three times. GPT 5.5 Pro reviewed the clinic materials and flagged two concerns: a recent guideline says shockwave therapy should not be used for rotator-cuff tendon problems without calcification, and Traumeel is registered in Germany as a homeopathic medicine without a stated treatment use. The raw MRI export contained a few hundred extensionless DICOM files totaling about 266 MB. Opus 4.8 inside Claude Code was allowed to install needed packages and spent about an hour analyzing the MRI. Its first report disagreed sharply with the clinic and found the tendon intact. A second run compared the clinic report, the first AI report, and symptom-checking context from GPT 5.5 Pro; Claude Code used separate subagents to review the case from different angles. That second report again favored no clear partial or full tear, while noting mild tendon irritation or wear. The result created a hard practical problem: the clinic plan looked too aggressive, but the AI second opinion was not trustworthy enough to replace a real medical expert.