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-5.6 is expected to launch publicly on Thursday with three named models: Sol, Terra, and Luna. GPT-5.6 Sol has already started appearing for some Plus users. Altman said GPT-5.6 uses 54% fewer tokens for agentic coding, which could mean lower cost and less waiting during long coding tasks. Early testing is focused on practical work such as deep research, building small apps, finding hard-to-find products or parts, comparing complex options, and exploring ideas. Some people who had moved toward Claude Code or Claude are trying ChatGPT again, with early praise for more natural handling of emotional context and story-world consistency checks. There is still rollout confusion: in the ChatGPT app on Codex, some users see only up to Extra High thinking effort and do not see Luna at max effort. GPT-5.6 is also being compared against Claude, Grok 4.5, and Muse Spark by asking each model to build the same four apps.
Since version 2.1.91 (released April 2, 2026), Claude Code has contained code that, whenever a user connects through a proxy, secretly checks whether the system timezone is set to Shanghai or Urumqi and whether the proxy's domain matches a list of Chinese domains or known Chinese AI labs. Based on those checks, the app covertly alters the date formatting inside the system prompt sent with each request — a steganographic technique that leaves an invisible marker without the user's knowledge. The code was deliberately obfuscated inside the Claude Code binary, making it hard to spot. It came to light when a developer who runs Claude Code through a personal proxy (to mix GPT and Claude models and fine-tune context management) was reverse-engineering version 2.1.196 to undo a change that disabled remote control whenever proxying was enabled, and stumbled onto the hidden check. The discovery triggered a heated discussion on Hacker News, and reports say Alibaba is considering banning Claude Code internally over the alleged backdoor risk.
The New York Times and The Daily News claim OpenAI failed to turn over important evidence in their copyright lawsuit. The dispute centers on whether ChatGPT was trained on their journalism and whether its answers repeated that journalism. OpenAI has argued that searching its training data and customer chat logs would be hard and would raise privacy concerns. A court-ordered deposition reportedly showed that OpenAI privacy engineer Vinnie Monaco said the company had already run internal searches and checks on its training data for copyrighted news content. The publishers also claim OpenAI had built a database of about 78 million de-identified ChatGPT conversations before the lawsuit and later used a Bloom filter inside a tool set called Project Giraffe to detect when answers repeated protected material. The publishers first asked for a sample of 120 million chat logs, but the number was reduced to 20 million. OpenAI submitted that smaller sample last December, but it reportedly had so many redactions that the court saw it as unusable. The publishers also claim OpenAI deleted billions of ChatGPT outputs after the lawsuit began, despite a court order to preserve them, and replaced millions of logs in the requested sample. They want the judge to punish OpenAI by blocking the 20 million log sample from evidence, treating major repetition of their content as established, limiting OpenAI’s arguments about the sample, and making OpenAI pay related legal costs. OpenAI denies the claims and says the Times is trying to reach private conversations from people unrelated to the case while OpenAI defends user privacy and fair use.
OpenAI’s GPT-5.6 family is described as three models: Sol, Terra, and Luna. Sol is the strongest model and is aimed at coding, cybersecurity, biology, and long-running AI tasks. Terra is positioned as a balanced model for performance and cost, while Luna is meant for fast, high-volume use. The release is not a full public rollout at first. Access begins as a restricted preview for vetted partners, after a request from the U.S. government. The move follows wider concern around powerful AI models, including limits placed on Anthropic’s Fable 5 and Mythos 5 because of cybersecurity risks. OpenAI says GPT-5.6 did not cross its internal threshold for dangerous cyber capability and was tested with more than 700,000 GPU hours plus outside evaluations. Broader access may arrive in the coming weeks, but the exact public release timing remains uncertain.
Ghostcommit is an attack that hides instructions inside PNG images attached to pull requests. AI code review tools may treat those images as ordinary files and miss the readable text inside them. After the pull request is merged, a developer’s AI coding agent may process the image, read the hidden instructions, and follow them. The attack can make the agent place repository secrets, such as API keys, into the codebase as plain-looking lists of numbers. Tests by the University of Missouri-Kansas City’s ASSET Research Group found that Cursor and Antigravity carried out the theft, while Claude Code refused. The researchers told vendors about the issue and released a proof of concept. They also said 73% of recent pull requests in major repositories did not get meaningful review from a person or a bot.
OpenAI has opened an official page for running ads inside ChatGPT. Advertisers can place ads while people use ChatGPT to look at options, compare choices, and make decisions. OpenAI says ChatGPT can use the wider context of a conversation, not just simple keywords, to make ads more relevant. Advertisers can create an account in Ads Manager, set up a campaign, choose a budget and goal, and add ad details directly or in bulk. They can then measure results, change campaigns, and improve performance over time. Early advertisers named on the page include Best Buy, Lowe's, and VistaPrint. OpenAI says ads will be clearly marked, kept separate from ChatGPT's answers, and paired with controls for how people’s data is used for ads.
A solo developer of 9 years spent 3 weeks using Claude Code (86% on Opus 4.8), racking up 245 sessions, 70,000 messages, and 123.1 million tokens. The biggest shift came from using CLAUDE.md and persistent memory: writing decisions down once, instead of re-explaining context every session, turns the tool from a smart autocomplete into something closer to a teammate that remembers past decisions. The real cost isn't token usage itself but repeatedly re-reading the same context in long sessions. To cut that cost, breaking work into phases works well — finishing a task, asking Claude to write a handoff prompt for the next session, copying it, running /clear, and pasting it into a fresh session saved more time than any prompt trick. A meaningful chunk of early manual work also turned out to be tasks a skill or subagent could handle better in a single call.
A non-technical solo founder running a marketplace that connects developers selling skills to freelancers and small businesses shares a weekly SEO workflow built entirely around Claude, with no coding involved. Over three months, the site gained 30,500 organic clicks and 4.4 million impressions with zero ad spend, its Domain Rating rose from 0 to 50, and 330 articles were published. The routine: every Monday, export two CSVs from Google Search Console — Queries and Pages — and feed them to Claude with one prompt asking it to find queries getting impressions but no clicks, pages where position improved but click-through rate (CTR) dropped, keyword cannibalization between pages, and queries already ranking that have no dedicated content yet. Claude returns 10 to 15 specific, immediately actionable fixes each week rather than vague suggestions. In week three, for instance, Claude flagged that five separate articles were all competing for the same query, "how to install skills in Claude Code."
Velorn is a free open-source video editor built with Claude. It was made by a film and TV visual effects artist with 25 years of experience, and it runs on Windows, Mac, and Linux. The main idea is that Claude does not only help build the editor; it can also operate the editor itself. Velorn includes a local MCP server with more than 100 tools, and Claude Code can be connected with one copy-paste step. After that, Claude can read the timeline, inspect video frames, and review an edit shot by shot. It can also make real editing changes, including trims, moves, transitions, speed changes, text, motion graphics, and keyframes. With ComfyUI, it can generate images, video, and music through local models or an API, then bring community workflows into the project after checking for missing nodes or models and asking for approval before installing them. For audio, Claude cannot hear the music directly, but it can read volume levels, adjust faders, use compression, export the project, and check the result. Write actions show a plan before they run and go onto the normal undo stack, so changes can be reversed.
SpaceX is moving to buy Anysphere, the company behind the AI coding tool Cursor, in a $60 billion stock deal. Cursor helps developers write, edit, and understand code with plain-language instructions, and it competes with Claude Code and OpenAI Codex. The deal comes shortly after SpaceX’s IPO and looks aimed at strengthening xAI and Grok in AI coding, an area where Anthropic and OpenAI already have strong products. Cursor already has strong momentum with developers, while SpaceX can add computing power, distribution, and a much larger company platform. Reactions are split. Some Cursor users care mainly about whether the product stays useful, while others are worried because Cursor can see source code, product plans, unreleased features, and valuable company know-how. Nothing may change immediately, but the next 6 to 12 months will matter for pricing, privacy promises, model choice, and how deeply Cursor is tied into xAI or Grok.
LithosAI presents Lithos Engine as a way to run Kimi K2.7 Code at 1,000 tokens per second. Kimi K2.7 Code is a coding model with 1 trillion parameters, running on 8 B200 machines. It is positioned for use with Claude Code, Codex, or a developer’s own coding setup. LithosAI compares this with 174 to 291 tokens per second from major providers, making its setup 3.4 to 5.7 times faster. The company says the speed gain does not use approximation and keeps the model’s native precision and full quality. LithosAI was founded by Carnegie Mellon computer science professors Dimitrios Skarlatos and Zhihao Jia.
OpenAI is ending ChatGPT Atlas, its separate desktop browser, and moving the work into the new ChatGPT desktop app. The new app was released on July 9, 2026, and includes ChatGPT Work agent, ChatGPT Codex, and browser features for working on the web. Chrome users can keep using a Chrome plugin that connects ChatGPT to their normal browser instead of switching to a new one. The target shutdown date for Atlas is August 9, 2026. OpenAI plans to share more details inside the app and by email. Atlas first came to Mac in October 2025; OpenAI then added browser features to the Codex app in April 2026; those efforts are now being combined in the new ChatGPT desktop app.
A firsthand build from an iOS software engineer turned into a finished web game in 15 days, with Claude Code at the center of the process. The game lets players control a capybara on a scooter and deliver stacked food before time runs out, with both single-player and multiplayer modes. Claude Code handled most of the coding work, while ThreeJS, Suno, ElevenLabs, GPT Images-2, and Tripo3d helped create the rest. The code, textures, music, and sound effects were all made with AI tools. The game was entered in VibeJam 2026, a game development contest where more than 90% of the code had to be written by AI. The contest ran through April 2026, and the game won the top prize of $25,000. The finished game is available to play for free, and feedback is being collected.
Claude Code was reported to continue automatically when its AskUserQuestion feature did not receive an answer within 60 seconds. The timeout was not a setting passed by the user; it appeared to come from the question tool itself. The reported setup used Claude Code 2.1.198, Opus, AWS Bedrock, Linux, and the VS Code integrated terminal. The behavior was treated as a regression because earlier versions were expected to wait for the user’s answer. Anthropic said a release would add the setting to /config and make the timeout off by default. A temporary workaround is to add a long CLAUDE_AFK_TIMEOUT_MS value in the env section of settings.json, which effectively avoids the 60-second cutoff. The same pain point also led makers to build status lights, menu bar watchers, and phone alerts so they can see when Claude Code is waiting for input or approval.
Google has made the Interactions API generally available and is making it the main way to work with Gemini models and agents. It first appeared as a public beta in December 2025, and the official release now gives it a stable structure. Developers can send a model ID to get model output, or send an agent ID to run a more independent task. Long jobs can run in the background on Google’s servers. Managed Agents can create a remote Linux workspace where an agent can run code, browse the web, and manage files. Tool use is broader now, so Google Search, Google Maps, custom functions, and image results can be combined in one request. A Flex tier is meant to cut cost by 50%, and paid users can retrieve past interactions for 55 days. The older generateContent API will keep working, but newer long-running and agent features are expected to move toward the Interactions API first, and sometimes only there.
ZCode is an AI coding tool from z.ai, the maker of GLM. It is designed to help with planning, coding, review, and release work inside an existing developer workflow. The product is tuned for GLM-5.2 and says it is optimized for reasoning, coding, and work involving multiple agents. The example screen shows ZCode building a browser-based Gomoku game from an empty project, writing files, checking JavaScript, tracking progress, and marking the goal complete. The listed GLM Coding plans are Lite at $16.2 per month, Pro at $64.8 per month, and Max at $144 per month. Pro includes selected MCP tools and faster generation speed. Downloads are listed for macOS, Windows, and Linux, with Linux marked as beta. ZCode also presents remote task starting through WeChat, Feishu, or Telegram.
Cursor now offers Grok 4.5. Cursor says it trained Grok 4.5 with SpaceXAI, calls it its strongest model so far, and presents it as the first model it built for more than software engineering. The listed price is $2 per million input tokens and $6 per million output tokens, with double usage during the first week after launch. Early experience is mixed. Grok 4.5 High Fast has been used for product requirement documents and issue writing with results described as close to Opus, much faster, and less padded than Opus or GPT5.5. At the same time, some users on the latest Cursor version cannot see Grok 4.5, and some saw it appear on launch day and then disappear later. Availability may also differ by region, with an EU user specifically reporting that the model was missing. Other reports raise billing and routing concerns: Grok 4.5 appeared to use GPT 5.6 sol as a subagent in one case, while Grok 4.5 High Fast reportedly switched to Sonnet 5 High and consumed on-demand usage in another. The open practical questions are how manual model choice compares with Auto mode, how Grok 4.5 counts against Cursor’s $20 plan, and how that mileage compares with Codex’s $20 plan for daily coding.
Two days of network traffic from one ChatGPT Pro account showed part of how ChatGPT chooses sources for web-backed answers. The data covered about 1,240 source records, so the exact percentages should be treated carefully, but the repeated internal fields are meaningful. Each fetched web result had a result_source label, with four observed values: serp for ordinary web results, labrador for major publishers and reference sites, bright for a commercial scraper, and oxylabs for another scraper. ChatGPT did not search the web for every question. When a question was classified under turn_use_case as text, it answered from stored model knowledge instead of fetching current pages, and even some questions that sounded time-sensitive fell into this path. For comparison tasks, the thinking model split one question into roughly 15 to 40 smaller searches, checked pricing pages, guessed possible prices, and looked for clues such as currency symbols. Being fetched, being cited, and being mentioned were separate outcomes. Reddit was more likely to become a cited source because the text is easy to read, while YouTube was often fetched but not cited because the model usually saw metadata rather than the full video content. If pricing or product details sit behind JavaScript, ChatGPT may fail to read the official page and cite a third-party site such as G2 instead.
Microsoft has started handling part of the Copilot workload in Excel and Outlook with its own MAI models instead of OpenAI and Anthropic models. The in-house models are already taking tens of thousands of prompts each week across the two apps. Excel and Outlook were previously described as relying heavily on OpenAI and Anthropic models. This change shows Microsoft using its own models as a live replacement inside real products, not just as a research project or side option. Microsoft’s stock rose 1.75% that day, suggesting investors saw the news mainly as a cost and margin story. Because Copilot inside Office is one of Microsoft’s highest-use AI surfaces, even a partial switch could change how much Microsoft pays outside AI companies to bundle AI into everyday work tools.
China's National Vulnerability Database warned that some versions of Anthropic's AI coding tool Claude Code may create a security risk. The warning names versions 2.1.91 through 2.1.196 and says they could send location-related signals, such as time zone data, and user-identifying information to remote servers. Version 2.1.200, released on July 3, is described as outside the affected range, and users were told to uninstall the older versions or update. Alibaba also moved to block Claude Code in its workplace from July 10 and pointed employees toward Qoder instead. Anthropic said the disputed behavior was a temporary anti-abuse measure meant to detect unsupported access, unauthorized reselling, and distillation. The dispute also exposed a wider pattern: some Chinese companies and developers have used VPNs, overseas subsidiaries, or cloud routes to reach Claude Code even though Anthropic restricts access from China.
Claude Cowork is coming to mobile and web. Work can be started at a desk, continue after the laptop is closed, and be checked later from a phone. Scheduled tasks can run even when the computer is off, such as preparing a client brief from threads and transcripts at 6 a.m. and leaving a follow-up message drafted but not sent. When human judgment is needed, Claude can send the question to the phone. On web and desktop, chat and Cowork now live in one place, with shared access to projects and artifacts. Starting a Cowork task is meant to feel like starting a normal conversation. The beta is rolling out over the next several weeks, beginning with the Max plan, with more plans to follow. Doubled Cowork usage limits are extended through August 5. Early reactions are mixed. Some non-technical users see Cowork as a major time saver for curriculum design, grading, slide decks, and grade analysis. Others were frustrated when the Cowork tab appeared to move into Projects without clear notice, causing fear that folders, prompt scripts, and instructions had been lost. There is also an open question about how Cowork differs from Claude Code, which some people find more flexible even for documents, PDFs, and sales material.
Anthropic analyzed about 400,000 Claude Code sessions from about 235,000 people between October 2025 and April 2026, using a privacy-preserving method. In a typical session, people made about 70% of the planning decisions, such as what to build and what counts as done, while Claude made about 80% of the execution decisions, such as which files to edit and which commands to run. About 56% of sessions involved writing new code, fixing code, testing, or automation, and 17% involved operating software such as deploying, configuring, or running systems. Over seven months, debugging fell from 33% of sessions to 19%, while operating software, data analysis, and writing documents became more common. More expert users got Claude to do more work from each instruction. Novice sessions led to about 5 Claude actions and 600 words of output per prompt, while expert sessions led to about 12 actions and 3,200 words. Success also rose with domain expertise: novice sessions reached verified success 15% of the time, while intermediate or higher sessions reached 28% to 33%. In sessions that produced code, software workers and non-software workers had almost the same partial success rate, 89% and 88%, which suggests that understanding the problem can matter more than having a traditional coding job.
Google limited Meta’s access to Gemini because Meta wanted more computing capacity than Google could supply. The limit was communicated around March 2026 and disrupted or delayed some of Meta’s internal AI projects. Meta told staff to use tokens more carefully while also trying to control AI costs. Meta has used Gemini and Claude for internal work such as scam detection, harmful content handling, customer support, ad helper chatbots, workflow support, and coding. Gemini became important inside Meta because it performed better than Meta’s own Llama models for some tasks. Google has also been short on capacity and signed a deal to rent computing power from SpaceX for about $920 million per month. Google Cloud passed $20 billion in quarterly revenue, but its backlog of signed work grew to more than $460 billion, and Sundar Pichai said limited computing capacity held back revenue. In the same competitive setting, Meta also ran a large safety benchmarking project against ChatGPT, Gemini, and Character.AI using teen-like accounts and prompts about suicide, sex, drugs, and eating disorders; rival companies said the testing was not authorized or may have broken their rules.
Claude Code workflows are moving beyond one long chat with one model into a setup where several agents split a development job. The main pattern uses a Markdown feature spec as the source of truth, then a primary agent called Fable assigns work to several Opus subagents. The primary agent checks their output, sends work back when it needs correction, and can keep a feature build moving for several hours before a final human review. Related workflows apply the same idea to bug hunting, rule checks, context management, and code review, with several agents looking at the work from different angles. Cost is a major reason this pattern matters. Reported benchmark numbers showed Fable 5 as the manager and Sonnet 5 as the worker reaching 96% of the performance of an all-Fable setup at 46% of the cost; on BrowseComp, that meant 86.8% versus 90.8% accuracy and $18.53 versus $40.56 per problem. A second pattern, where Sonnet 5 executes and consults a Fable 5 advisor, reached about 92% performance at about 63% of the cost on SWE-bench Pro.
A Gemini account showed an Untitled conversation containing AI output that the account owner said they never requested. Related reports describe new chats that suddenly continued an older conversation, while other chats lost the thread after almost every message. In one case, Gemini seemed to attach a Chat Summary from only one past conversation, even though the new chat should have been separate. Other problems pointed to confusion around Google services: a Drive video was summarized even though permission was not remembered, and a Google AI Plus 400 GB subscriber did not see the Gemini button in Gmail despite expecting Gemini in Gmail. The cause is not confirmed, but the pattern raises concerns about session or cache mix-ups, uneven account feature rollout, and unreliable context handling in Gemini’s web experience.
OpenAI audited SWE-Bench Pro, a widely used test for AI coding ability, and estimates that about 30% of its tasks are broken. On the public set of 731 tasks, top AI models rose from a 23.3% pass rate to 80.3% in eight months, but the audit suggests those gains may not fully reflect real coding skill. An automated review marked 200 tasks, or 27.4%, as broken, while human reviewers marked 249 tasks, or 34.1%, as broken. The main problems fell into four groups: tests that forced one narrow implementation, prompts that left out needed details, weak tests that let incomplete fixes pass, and prompts that pointed models toward the wrong behavior. OpenAI used Codex-based investigator agents plus reviews by five experienced software engineers. OpenAI is now withdrawing its earlier recommendation to use SWE-Bench Pro and says model developers should treat results from it carefully.
A web developer turned an idea into a working coffee shop ordering demo with Claude Code in 40 minutes. The app included a separate subdomain for each coffee shop, preloaded logos and names, QR code ordering, a barista CRM, staff logins, and a Telegram-based loyalty system with automatic discount messages. It was not just a mockup; it was deployed and already running for real customers. The developer visited three coffee shops with the live demo instead of a slide deck. The first shop owner agreed to buy it in a five-minute conversation, and the sale was worth $700. The important lesson is not only the speed of the build. Demos are now easy to produce with AI, so the real edge came from knowing what shop owners care about, reading the room, predicting objections, and answering concerns before they became blockers. Claude Code shortened the build, but it did not create the customer understanding that made the sale possible.
OpenAI is in early talks about giving the US government a 5% stake in the company. Sam Altman is arguing that the public should share in the money created by AI. The idea could also ask other major US AI companies, such as Anthropic, Google, and Meta, to give a similar stake, but it is not clear whether they would agree. The stake could sit in a government-backed vehicle similar to a sovereign wealth fund, modeled on Alaska’s fund that invests state oil money and pays benefits to residents. The talks are still conceptual, and a real deal may need approval from Congress. The pressure behind the idea includes worries in Washington about data center growth, jobs, cybersecurity, and the release of powerful new AI models. Based on OpenAI’s reported private value of $852 billion, a 5% stake would be worth about $42.6 billion, and it could be worth about $50 billion if OpenAI reaches a $1 trillion value in a public listing.
Claude still has clear strengths in long conversations and memory, but recent use suggests it may be misreading intent and treating more normal requests as sensitive. Topics such as religion, philosophy, psychology, and creative writing have triggered refusals or overly cautious responses even when the request was framed as harmless. A fiction idea about aliens bringing back lost writings that challenge parts of a religion was rejected even after the purpose was clarified, but the same prompt in a fresh chat with Sonnet 5 worked normally. That inconsistency matters because the same request can get different results depending on the chat or Claude model. Similar pushback appeared with a basic question about Zoroaster’s influence on Judaism, plus some financial planning and brainstorming tasks. In a poll of 136 people, 34% said Claude had become too preachy and pushed back too hard, 13% had seen the issue once or twice, and 42% had not seen a problem. The issue did not seem limited to Sonnet 5; Opus 4.8, Sonnet 5, and Sonnet 4.6 seemed more likely to push back, while Opus 4.6 and Opus 4.7 caused fewer problems. A later test across about six chats produced only one real refusal, so the problem may vary by topic, wording, model, and timing; Claude can still give very strong answers, while Gemini may feel more consistent.
Google added new features to Managed Agents in the Gemini API. A managed agent lets one API call handle reasoning, code execution, package installation, file handling, and web information inside an isolated Google cloud workspace. Background execution lets long tasks keep running on the server while an app receives an ID, checks status, streams progress, or reconnects later. Remote MCP server support lets agents connect directly to outside tools such as private databases or internal APIs. Built-in sandbox tools and custom functions can now be used together; when a custom function is needed, the interaction pauses so the client app can run its own business logic. Credentials such as short-lived access tokens or API keys can be refreshed in the next interaction by sending a new network configuration, while the sandbox keeps its files, installed packages, and cloned repositories. The examples use the @google/genai JavaScript SDK, with separate documentation for Python and cURL.