Daily AI · 2026-07-31
Useful AI Daily - July 31, 2026
Today's useful AI signal is that access changes need an operating plan. OpenAI says GPT-5.6 pricing is lower for Luna and Terra, GitHub Models has been retired, and Copilot's latest editor updates make it easier to put more agent work into a repository. The practical move is to compare cost, permissions, and review capacity before changing a production workflow.
Try a small cost-and-access inventory before switching a model or enabling an agent feature. Watch for tools that retire faster than a migration plan, and for AI-assisted changes that become too large to review. Skip the assumption that a lower price or a new preview removes the need for a human owner.
The Short Version
- Try: rerun one fixed, non-sensitive task on your current setup and any new model option. Record cost per successful result, latency, retries, and the human time needed to verify it.
- Watch: GitHub Models is retired. Inventory any playground, inference API, BYOK credential, model name, or tutorial that still depends on it before a customer or teammate finds the breakage.
- Ordinary-user bridge: a lower AI price does not make a personal document safer to upload. Remove account numbers, health details, private photos, and unreleased work unless the product's data controls are clear.
- Builder signal: more capable coding agents raise the value of smaller, independently reviewable changes. Keep the approval path and required checks visible at every layer.
- Skip: do not turn on every available skill, agent, or command shortcut in a shared repository. Start in a disposable branch with a named reviewer and a rollback path.
5 Updates Worth Your Time
OpenAI says GPT-5.6 pricing is lower for Luna and Terra
- What changed
- OpenAI announced lower GPT-5.6 pricing for its Luna and Terra variants in a July 30 product post. The announcement frames the change as a price-performance improvement for enterprise AI workflows, but the right cost depends on the model, task, context size, retries, and review work around it.
- Why it matters
- Who should care: operators, creators, and small teams paying for repeat AI tasks. A cheaper listed model can still cost more when long inputs, tool calls, failed runs, or manual cleanup grow with it.
- Try, watch, or skip?
- Use a fixed sample of non-sensitive tasks and compare cost per verified result, not cost per request. Check the current product page and model identifier before changing a production default, then keep a budget cap and the old path available for the pilot.
GitHub Models is now retired
- What changed
- GitHub says GitHub Models retired on July 30. Its playground, model catalog, inference API, and bring-your-own-key option are no longer available, including to customers with active usage.
- Why it matters
- Who should care: developers, educators, and indie builders who used GitHub Models for a prototype, demo, prompt test, or internal tool. A quiet dependency can fail in a customer path long after the original experiment was forgotten.
- Try, watch, or skip?
- Search code, secrets, documentation, and onboarding steps for GitHub Models references. Name the replacement for each real use, test one critical path without cached credentials, and tell users what changes before removing the old instructions.
Copilot in VS Code adds more agent review and session controls
- What changed
- GitHub's July VS Code release summary covers versions 1.127 through 1.131. It adds a redesigned Agents window in public preview, worktree support for Copilot, Claude, or Codex sessions, multi-chat sessions, and more chat controls; Copilot vision is now generally available for images and PDFs in chat.
- Why it matters
- Who should care: people who use AI coding help but still need to inspect diffs. Separate worktrees and clearer review views can reduce accidental edits across concurrent experiments, while file attachments expand the data that can enter a chat.
- Try, watch, or skip?
- Use one disposable worktree and a small pull request. Keep input files non-sensitive, compare the final diff with your normal review, and disable any workflow that makes ownership or rollback less clear.
Visual Studio adds preview agent choices, built-in skills, and selected-code review
- What changed
- GitHub says Visual Studio's July update includes a preview Copilot agent based on the Copilot SDK, optional built-in .NET and Azure skills, selected-code review, and organization-level custom instructions. The built-in skills are off by default; organization instructions require Copilot Business or Enterprise.
- Why it matters
- Who should care: .NET teams and solo developers who want more useful context without losing control of their development environment. A useful skill is a narrow, inspectable instruction set, not a reason to give an agent broad access.
- Try, watch, or skip?
- Enable one built-in skill on a low-risk repository, review its scope, and test selected-code review against a bug you already understand. Keep organization-wide instructions short, versioned, and easy for developers to inspect or disable.
Stacked pull requests enter public preview
- What changed
- GitHub introduced stacked pull requests in public preview. A stack breaks a large change into dependency-ordered pull requests that can be reviewed independently while existing branch protections and required checks still govern what reaches main.
- Why it matters
- Who should care: any builder whose AI tool produces a change too large for one reviewer to understand. Smaller review units preserve context without asking a teammate to approve a giant diff because the agent made it quickly.
- Try, watch, or skip?
- Split one multi-step feature into three layers: data or contract, implementation, then UI or integration. Require tests for each layer and make the final merge depend on human review of the whole behavior, not only green checks.
Tool Worth Trying Today
GitHub Stacked Pull Requests: a five-layer change test
Use a small feature to test whether stacked pull requests make AI-assisted work easier to review. Start with a contract or data layer, then implementation, tests, UI, and documentation. Review each layer on its own before deciding whether the stack improves clarity.
Best for: Teams and indie builders whose changes have real dependencies, required checks, and a reviewer who needs to understand more than the final generated diff.
Watch out: This is a public preview rolling out to repositories. Confirm availability in your repository, keep branch protections enabled, and do not let a chain of small pull requests hide a risky combined behavior.
Privacy / Cost Watch
- Do not upload sensitive personal, customer, legal, unreleased, health, financial, hiring, research-participant, or private photo or document data to a new AI tool until its terms, retention settings, admin controls, data location, and deletion path are clear.
- Image and PDF chat attachments can carry more than the visible text. Remove private pages, embedded metadata, access tokens, customer names, and source material you are not permitted to share before using an AI review feature.
- Lower model pricing does not eliminate cost risk. Set a cap for experiments, count retries and human verification time, and review who can add a model, skill, key, or external connection to a shared workflow.
One Practical Workflow
Run a 25-minute AI access-and-cost migration check
- List every AI model, endpoint, key, skill, attachment type, and repository integration used by one real workflow.
- Mark each item as keep, replace, disable, or investigate. Treat a retired service, an undocumented key, and a broad attachment permission as items that need an owner today.
- Run one non-sensitive test task through the current or replacement path. Record model identifier, input size, elapsed time, retries, cost estimate, output quality, and named human verifier.
- Review the generated diff or result in a small unit. Check source facts, permissions, tests, and rollback before approving any customer-facing, financial, legal, security, access, or publishing action.
- Write the decision in the repository or team runbook: chosen path, budget cap, permitted data, owner, fallback, and the date to recheck it.
Builder Note
AI makes it cheap to create more code, copy, and configuration. It does not make a large change easier to understand. Design the delivery path around independently testable layers, clear ownership, and visible approvals. A product that can explain what changed, who approved it, and how to roll it back earns more trust than one that merely ships faster.
Ignore For Now
Broader agent access without a migration and review plan
Skip a new skill, model, connector, or command shortcut when nobody can name the data it sees, the action it can take, the fallback after retirement, and the person who reviews the outcome. Start narrow. A productive AI workflow stays useful when a provider changes price, access, or availability.
Bottom Line
Bottom line: lower prices and richer agent interfaces are reasons to measure, not reasons to expand blindly. Keep access, cost, and approval visible; migrate retired dependencies before they become outages; and make every AI-assisted change small enough for a person to review with confidence.
Sources
- OpenAI: Advancing the price-performance frontier with GPT-5.6
- GitHub Changelog: GitHub Models is now retired
- GitHub Changelog: GitHub Copilot in Visual Studio Code, July 2026 releases
- GitHub Changelog: GitHub Copilot in Visual Studio - July update
- GitHub Changelog: Stacked pull requests are now in public preview