Gemini 4 Argon’s introductory usage prices double after November 30. GitHub retires three Copilot models on November 16, with an exception for annual individual subscribers. Cursor’s new subscription bundle separates its included Grok allowance from charges for other providers’ models.

Coding Agents: prices, allowances, and retiring models

Google’s Gemini 4 Argon targets complex, multi-step coding. Its API is the service an app or coding agent uses to send requests to the model. Introductory rates are $2 per million input tokens and $10 per million output tokens. Tokens are the small chunks of text that providers count when billing requests.

Those rates rise to $4 and $20 after November 30. A coding agent can make repeated requests during one task, so the increase affects the cost of continuing work as well as starting it. Argon is also available through GitHub Copilot and Vercel AI Gateway, with access and billing depending on the route.

Copilot’s retirement date and review costs

On November 16, Copilot retires Claude Opus 4.8, Claude Opus 5, and GPT-5.6 Terra, but annual individual subscribers retain Opus 4.8. GitHub suggests Claude Opus 5.5 for the retiring Opus models, and GPT-6.1 Sol or GPT-6 Luna for Terra. Business and Enterprise plans enable the replacements automatically unless an administrator has disabled the model or automatic enablement. Change any saved workflow that names a retiring model before that date.

Copilot also updated HydraFusion, its research-preview option that can assign drafting and review to different models. It now uses medium reasoning effort to draft and high effort to review, and a new Max option raises effort further. GitHub bills each model it uses at that model’s standard rate, so additional passes can consume more AI credits, Copilot’s usage balance. A discarded draft can leave file edits behind, so inspect the changed files before saving them in a Git commit.

Cursor separates Grok from third-party usage

Cursor joined xAI’s X Subscription, which bundles Grok, Grok Bot, Cursor, and X Premium. Grok usage in Cursor shares one weekly allowance with the Grok apps, so heavy coding leaves less for chat. Models from other providers cost extra on Premium and above, billed at the provider’s published API rates. Existing Cursor subscribers can keep their plans. Moving to the bundle replaces any included third-party usage with those separate charges.

These arrangements affect different bills. Google’s API rates apply to direct model requests, Copilot charges its own credit balance, and the X Subscription includes Grok usage alongside separately charged models. Compare the total cost of your usual work before replacing a plan that already includes the models you use.

Conversation history, upgrades, and review

  • Claude Code 2.1.299 compacts conversations of 200,000 tokens or more after about 54 idle minutes. Compaction replaces the history with a shorter summary, so exact error messages may be lost. Version 2.1.302 adds an Idle compaction toggle in /config, the command that opens its settings.
  • OpenCode 2.0 adds background agents and commands, letting other work continue while a longer task runs. It does not support 1.0 plugins, so check any extensions you depend on before upgrading. Separately, version 1.19.1 ended session sharing, and existing conversation links no longer open.
  • Anthropic’s Code Review is generally available for Team and Enterprise plans. Once an administrator connects the GitHub app and selects repositories, it reviews new pull requests automatically. These proposed code changes still need a person to assess the findings.
  • Claire Vo’s roughly 20-task benchmark ranked Meta’s Muse Wonder first on three agent tasks but found its design output weak and its agent runs slow. Part of the scoring came from another AI model grading the results.

AI Builders: personal apps, regional processing, and ad budgets

Lovable expanded the models and tools available inside apps, including voice conversations, Claude models, and web research. You can ask for a feature such as a voice tutor or a research panel with linked sources. Lovable chooses a model or tool, or you can name one in the chat.

Google Play is rolling out a way to run personal Android apps built with Google AI Studio on your own device, without a developer account. It starts in select markets, including the US, for users 18 and over. Sharing with friends and family is planned for the coming months; public distribution still requires a professional developer account and the usual reviews.

Regional processing changes which features work

For new requests, Lovable’s EU inference setting for Enterprise workspaces controls where supported models process prompts and produce replies. Already-running requests can finish outside that boundary, and direct connections to other providers bypass the setting. Requests fail if no permitted endpoint is available. Some Microsoft-hosted models can run in Norway or Switzerland, so the setting does not promise processing only within EU member countries. App storage is a separate choice, and web search and browser automation still process data in the United States.

Lovable’s announcement also says Claude remains available while building, but not through built-in Lovable AI in published apps with this setting enabled. Published apps using unsupported models through Lovable AI need a supported replacement and must be published again for those requests to work. Built-in video generation and voice conversations become unavailable, and AI usage consumes more credits. Check the models your published app uses before enabling the setting.

Campaign creation uses credits before ad spend begins

Lovable also plans and drafts ads and search content from the chat that builds your app. Creating plans and ad assets consumes regular Lovable credits. Google Ads and Meta bill ad spend directly, and Lovable says it adds no service fee to that spend. No campaign launches until you approve its budget, targeting, and goals, according to Lovable.

Peter Yang argues that cheap AI clones make traditional software harder to sell, and forecasts revenue moving to products agents can use directly. His own buying rule is to skip software without an API or MCP connection for his agent. MCP is a standard that lets an assistant call a product’s functions. For an app builder, this raises a product question: can an assistant perform your app’s main task, or must someone click through its screens?

Infra & Deployment: agent files, data handling, and saved releases

Cloudflare open-sourced Void, a toolkit from VoidZero, the company behind Vite, which prepares web app code for browsers. Running void deploy builds a Vite app, applies database changes, creates the resources it needs, and publishes it to Cloudflare. A separate Void Platform lets a team host internal apps in one place without giving each builder access to the company’s cloud account. The tools and Platform ship together as open source, so teams can adapt the Platform code.

Vercel added a free route to NVIDIA’s Nemotron 3.5 Lightning. Vercel warns that NVIDIA serves the requests and may retain prompts to improve its models. Use public or invented data when you try it.

Vercel Drives is generally available for Vercel Sandbox, the separate environments where agents run code. Files stored on a drive survive when the agent moves to a new environment. An agent can resume with the files it saved instead of rebuilding its workspace.

Token Pass is in early access for Pro teams at $30 a month and includes $80 in monthly credit for GLM 5.3 and Kimi K3. Requests stop when the credit runs out, with no automatic overage charge, and unused credit does not roll over.

October 23: check which older releases you need

The Vercel retention deadline is six days away: deletion of deployments older than 30 days starts October 23 unless affected Pro teams opt out. Settings already at 30 days or shorter stay unchanged. Deployments are saved app releases; losing an older one can remove a version you could restore after a bad update. The policy protects the deployment serving your production address, certain recent releases, and the latest preview for each active branch. Check retention settings before the deadline if you need older versions.

A GitHub outage that reached Vercel

On October 12, Vercel reported failed builds for GitHub-connected projects and linked them to an upstream GitHub incident. Vercel closed its incident at 20:11 UTC after processing delayed deployment status updates on pull requests. A GitHub outage can block a new release without taking down an existing site. If a release fails, check whether the live version still works before retrying the deployment.

End of article