OpenAI cut its GPT-6 Sol and Luna API rates, and Anthropic estimates Opus 5.5 costs less to run on typical token-billed work. xAI kept Grok 4.7 at the same listed price as Grok 4.6; Vercel’s 40% offer is a temporary gateway promotion.

GitHub added separate ways to manage Copilot approvals and local access. Cursor introduced bots that review changes from pull request to production. Cloudflare and Lovable also made common setup and planning tasks easier to hand off to an agent.

Coding Agents: Lower Costs, Clearer Limits

Model costs and the temporary Grok offer

OpenAI expanded the GPT-6 family with Sol and Luna and cut their API rates by 50% compared with GPT-5.6’s promotional prices. Sol costs $2 per million input tokens and $10 per million output tokens. Luna costs $0.10 and $0.50. GPT-6 Astra remains at $10 and $50.

Tokens are pieces of text that an API counts and charges for processing. These are usage rates, so they do not automatically lower the price of a fixed subscription. If you are choosing a model for coding, compare the total bill and the amount of correction work on the same routine task before changing your default.

Anthropic estimates Claude Opus 5.5 costs about 40% less to run than Opus 5 on typical token-billed workloads. That is an estimate for total usage, not a 40% cut to every price or plan. Anthropic says Claude now works 3.3 times longer on each prompt and takes in 2.6 times more context per request than six months ago.

Claude Code users also have two time-limited actions. Activate the usage reset from Claude.ai’s Usage settings, not from Claude Code. Existing Pro and Max subscribers can claim a one-time cloud-session credit: $100 on Pro and $250 on Max. Claim it by October 7, as announced by Anthropic’s ClaudeDevs account. Our full briefing has the details.

xAI says Grok 4.7 launched at the same API price as Grok 4.6, with better performance on longer tasks.

Separately, Vercel is offering 40% off Grok 4.7 through September 27 on AI Gateway. That promotion does not change xAI’s base price.

GitHub adds separate approval and access controls

Copilot for JetBrains 1.18.0 added assisted approvals in public preview. GitHub says low-risk tool calls receive automatic approval, while higher-risk calls still prompt you. Its release notes do not define the risk categories or explain how to review approvals afterward.

Local sandboxing arrived in the Copilot app, also in public preview and off by default. It lets you limit the project files, network access, and saved GitHub credentials available to an agent. This setting is per project and covers local sessions, not cloud, remote-host, or Copilot CLI sessions. Our guide shows what to test.

To enable it, open the project’s settings in the Copilot app, choose Sandbox, and turn on Sandbox new sessions. Use /sandbox on to enable it for an active local session. Assisted approvals decide which tool calls need a prompt. Sandboxing limits what an agent can reach if a call is approved anyway.

Cursor reviews from pull request to production

Cursor’s Rollouts and Security Reviewer bots start before code ships. Rollouts follows a pull request through deployment. It needs connected source control, a deployment system, and a telemetry service that tracks app behavior. Before merge, it writes a monitoring plan.

After deployment, Rollouts compares live behavior with a pre-deploy baseline. It can alert the team, pause a staged release, or open a revert pull request for a person to approve. A staged release sends the change to a small group before a wider launch. Our briefing covers the setup.

Security Reviewer runs on every pull request before merge. It checks for injection bugs, where user input can alter database queries or commands, as well as missing login or permission checks and exposed secrets. It also flags unsafe data parsing, unvalidated redirects, vulnerable dependencies, and insecure infrastructure settings. It explains findings and proposes a one-click fix for you to review.

Both bots are currently available only on Cursor’s Teams and Enterprise plans. Rollouts will not help much if you have no deployment or metrics setup to connect. Security Reviewer has a simpler starting point, but read each proposed fix before applying it.

Infra & Deployment: Python and Bot Protection

Python Workers reached general availability on Cloudflare’s platform. You can run frameworks such as FastAPI, Django, and Flask, and connect Python code to Cloudflare’s other services. The runtime uses Pyodide, a version of Python compiled to WebAssembly, a compact format for running code in a sandbox. Some packages that rely on compiled C or Rust code still need a compatible version. Check the packages in our Python Workers briefing before building around one.

Turnstile Spin lets a coding agent set up Cloudflare’s free CAPTCHA alternative. The full setup needs a browser widget and a backend call to Cloudflare’s verification service. Without that backend check, the widget can appear to work while the site still accepts bot submissions. Turnstile’s dashboard can flag widgets without a matching check and offer a Spin fix for you to review. Our setup guide has the steps.

AI Builders: Free Chat and Narrower Access

Lovable’s daily free chat allowance lets you discuss an idea or question an existing app without spending credits, up to the daily limit. Generating images or video still costs credits. So does handing work to Lovable’s Plan or Build modes. The current terms run through October 31, but Lovable has not said what follows. Our briefing explains the line between discussion and app changes.

Microsoft’s new Copilot Code feature lets people describe an app, tracker, dashboard, or workflow in natural language. It is a Microsoft 365 Copilot product, separate from GitHub Copilot, and uses the same underlying technology. Code is rolling out to Microsoft’s Frontier early-access program at the end of September, with wider access planned in the coming weeks. Microsoft says Code uses usage-based billing, where charges vary with use rather than a fixed monthly license, so check the cost model before building a workflow around it.

UpGuard’s Supabase security study examined roughly 300,000 domains with signs of Supabase use. It found 16,326 databases with readable tables. More than half had table-schema indicators of personal information. The researchers assessed schemas rather than reading every row, and did not attribute the findings to specific AI app builders such as Bolt or Base44.

If your app uses Supabase, check which users can read each table and review its row-level security rules, the rules that limit which records an account can access. Test signed-out and ordinary-user access with fake records before storing real customer data.

Workflow: Four Voices, Three Questions

Four people across three stories focus on different questions. Ed Zitron’s newsletter post estimates that finance has supplied $100 billion to $150 billion in debt for AI data centers outside the largest cloud providers. He argues many companies building those facilities lack experience with AI data centers or power procurement. Treat the range as Zitron’s estimate, not an audited total. Our briefing covers his wider argument and limitations.

Ethan Mollick’s essay calls the gap between what current models can do and what people actually use a capability overhang. He argues that much of today’s ability is still sitting unused. Ben Thompson uses overhang to discuss different pressures on frontier labs, including whether releasing models more slowly addresses risks from models already in use. These are related concerns, not a direct disagreement about the same claim. See our comparison for the distinction and links to both original essays.

Nathan Lambert says Chinese models’ share of open-model usage on OpenRouter rose from about 70% to over 80%. OpenRouter is one model-routing service, not a measure of all AI use. His briefing also names products from Cursor, Perplexity, DoorDash, and Airbnb that use Chinese open-weight models. Lambert’s briefing defines these models as downloadable weights people can run themselves. See our earlier briefing for more on the shift.

For builders, the useful question is which model fits a real task, license, and data-handling need. Try an existing capability on a routine piece of work and measure the result, cost, and review time before making a larger change.

What mattered

Lower token costs help only when your tool bills by usage and the result is good enough. Agent controls help only when you set their boundaries and check what they can reach. Test a narrow sandbox on a project, review each production-monitoring plan, and confirm database permissions with fake data before shipping.

The offers affect different costs: API usage, a temporary gateway discount, or a one-time cloud credit. Check which applies to your workflow before changing plans or model defaults. Compare models on the same task and starting files. Across several runs, track token spend, corrections, retries, and whether you would ship the result. That catches a cheaper model that needs more review.

What was noise

Google’s Gemini 3.8 Live Avatar is available in Gemini Enterprise and supports an API for conversational agents. It is a specialized enterprise feature, not a default choice for most solo builders.

Meta’s Horizon Create and Horizon Studio use AI to help make games for the Horizon ecosystem, with early access available by request. They target game creators rather than general app building.

What to watch

  • Whether GitHub explains which JetBrains agent actions count as low risk and how users can review automatic approvals.
  • Whether Cursor expands Rollouts or Security Reviewer beyond Teams and Enterprise.
  • The September 27 end of Vercel’s Grok discount, the October 7 Claude Code credit deadline, and Lovable’s October 31 chat terms.

End of article