Builder platforms leaned toward explicit limits this week instead of open defaults. Netlify tightened project privacy, and Vercel added spend budgets that can stop requests when they exceed a limit. OpenAI cut its cheapest model’s price 80% while raising the price of speed. An unreleased Anthropic model also found a serious flaw in a post-quantum signature scheme after two years of expert review.
Coding Agents: Reviews Gain Context as Models Turn Over
GitHub made agent skills and MCP support in Copilot code review generally available on July 29 for Pro, Pro+, Business, and Enterprise plans. A pull request is the page where a proposed code change sits before a team merges it. Copilot can now use your team’s own internal tools and standards in its review comments. MCP (Model Context Protocol) is the standard that lets Copilot pull in context from an outside system, like a ticket tracker, while it reviews.
Every MCP call a review makes stays read-only, so Copilot cannot write back to the connected system. That does not stop a review comment from quoting something sensitive it read there. That comment is visible to anyone who can see the pull request. GitHub’s own MCP server and a browser-automation MCP server are on by default, whether you configured MCP or not.
The model roster kept shifting too. GitHub retired Gemini 2.5 Pro and Gemini 3 Flash from Copilot and added Grok 4.5. Check any script or automation that calls a retired model by name.
A bigger governance change is coming on August 26. GitHub will start turning on newly generally available models by default for Copilot Business and Enterprise. That replaces the current system, which requires an admin to opt in first. Admins who want tighter control need to set an opt-out policy before that date.
Copilot’s IDE surfaces also got easier to run in parallel. VS Code’s Agents window, still in preview, can now start Copilot, Claude, or Codex sessions in a separate Git worktree. That is an isolated copy of your repository, so sessions do not overwrite each other’s files.
Cursor extended the same idea to a new surface. Cursor is now available on iPad on all paid plans. It ships with a full pull-request review screen and an inbox that tracks what several agents are doing at once. Reviewing and merging an agent’s work no longer requires a laptop.
AI Builders: Know Which Login Is Doing the Work
Lovable’s app-user connectors let each visitor connect their own external account to a deployed app. That differs from a shared app connector, which uses one connection for everyone.
Per-user connections can carry the visitor’s permissions through to the service, but they do not replace your app’s own authorization or protect a shared cache or export. Test a live sales dashboard with two accounts that have genuinely different permissions before handing it to more than one person.
xAI joined the AI app builder category this week with Build Mode, a new beta feature. It turns a chat description into a working website, game, or dashboard inside Grok. It is available now to SuperGrok Heavy subscribers on the web, iOS, and Android.
The pattern matches what Lovable, v0, and Bolt already do. Describe an idea, get a live preview, then publish it to a shareable link without touching code. A well-funded model vendor entering app generation is one more sign the category is becoming a standard feature. Anyone with a strong underlying model can now offer it, not just a handful of startups.
Infra & Deployment: Defaults Get Safer, Spend Gets a Ceiling
Netlify can make new projects private by default on eligible credit-based plans when a team enables the setting. That applies no matter how you deploy: a connected GitHub repo, a dragged-in folder, its command-line tool, or an AI agent. Only signed-in team members and invited collaborators can open the live URL until someone flips it public. The setting applies to new projects; existing projects do not change.
Vercel’s AI Gateway is the proxy service that routes an app’s AI calls to whichever provider is configured. It can now set spend budgets on an API key. Once a budget is exceeded, subsequent requests are blocked until it resets or is raised. An API key is the credential an app uses to bill its AI usage. The cap covers every provider and model running through that key, so give separate tiers separate keys when you need separate limits. In a July 26 post, independent developer Simon Willison argued that AI vendors need exactly this. He wanted a dollar limit that shuts a key off, not just a limit on how many requests it can make per minute. His argument followed an investigation into a relay market built on pooled credentials, abused free trials, and other forms of access fraud.
The same Gateway also picked up regional inference. You can pin a request to run in the US or the EU. That matters for data residency rules that require certain data to stay processed in a specific region. Pinning costs roughly 10% more. A request fails outright if no provider can serve your chosen model in that region, rather than silently running somewhere else.
Vercel rounded out access control with Vercel Passport, now generally available. It puts sign-in through Okta, Microsoft Entra ID, or another identity provider in front of a deployment, so you can keep access closed until a user is authenticated.
Simon Willison’s latest analysis found the same boundary problem in AI evaluations. An environment described as a simulation still reached real systems. A prompt is not isolation. If an agent can reach a repository, deployment account, or customer data, limit that connection as if the target were real.
The Model Context Protocol’s move to a stateless core, previewed in last week’s issue as landing July 28, actually shipped on schedule. Claude and Vercel’s MCP tooling already support the new spec. Stateless means a server no longer has to remember a session between requests. That makes it easier to run on serverless infrastructure that spins up only when needed. Existing MCP clients keep working without a forced migration.
Product: Cheaper by Default, and a Capability Surprise
OpenAI cut GPT-5.6 Luna’s API price 80%. Terra’s price fell 20%, effective July 30. Luna, the fastest and cheapest tier, now costs $0.20 per million input tokens (a token is roughly a word fragment). That is down from about $1. OpenAI also replaced its old Priority Processing option with Fast mode. GPT-5.6 Sol now runs up to 2.5 times faster at twice the price, with no change to the model’s actual intelligence.
Vercel’s AI Gateway picked up a matching idea the same week. A single speed: fast setting now requests the fast tier of any model that has one, including Anthropic’s Opus models. Claude Code can reach that tier with the /fast command, and GPT-5.6 Sol supports it too. The Sol fast tier is now 2.5x, up from 1.5x before this week. If a model has no fast tier, the setting simply does nothing. Pay extra for speed only when someone is actively waiting on the answer. It is less useful for background work nobody is watching.
Cost was not the only place a model surprised people this week. Anthropic said its unreleased Claude Mythos Preview security model found a serious flaw in HAWK. HAWK was a candidate for an official US post-quantum digital-signature standard. A digital signature is cryptographic proof that a message or file really came from whoever signed it and was not altered. Post-quantum means designed to stay secure even against future quantum computers.
HAWK had already survived two rounds of expert review over two years. Mythos then found a mathematical shortcut in a third round of testing. That shortcut would make an attack on the smallest HAWK-256 parameter set about 67 million times cheaper than its designers expected. Anthropic found the flaw within about 60 hours of analysis.
Anthropic is careful to say this breaks no deployed system. HAWK was still a candidate, not a shipped standard. Catching exactly this kind of flaw is what additional review rounds exist to do.
The real signal is how fast a purpose-built model found something two years of human cryptographers had already looked at. If you rely on a cryptography library or standard, treat this as a reminder to track its review status. It is not a reason to worry about anything already in production.
OpenAI added a similar reminder about evaluation setup. On the ARC-AGI-3 puzzle benchmark’s public test set, GPT-5.6 Sol scored 13.3% using the official harness. Two settings that let the model reuse past reasoning raised the score to 38.3% while using six times fewer output tokens. Check the harness before treating someone else’s leaderboard as a fair test of your use case.
Workflow: The Middle of the Job Is Getting Thinner
Rich Mironov argues that faster AI coding does not shrink the product manager’s job. It reshapes it into a barbell: heavier at both ends, thinner in the middle. Mironov, a product consultant with four decades in Silicon Valley, writes that engineering used to absorb most of a team’s time. That time went into turning a validated plan into working software. As AI tools close that gap, what remains concentrates on deciding what is worth building and making sure people actually adopt it.
His framing holds up on both ends. When shipping the wrong idea now costs days instead of months, more judgment upfront pays for itself. A working product still needs someone who can explain why a buyer should care.
Where his post goes quiet is the middle he names as engineering’s territory. As agents write more of the code, a smaller group of engineers absorbs responsibility for correctness and security. That responsibility now moves at a much faster pace.
If you run a small team or build solo, treat AI’s speed as a reason to spend more time, not less, on deciding what to build. Do that before you open a coding tool.
What mattered
The through-line was not another model launch. It was the move from informal trust to explicit controls: private defaults, spend budgets, regional routing, approval policies, and per-user permissions. Builders should treat those controls as part of the product, not cleanup after launch.
What was noise
Cursor also launched a discounted India-only pricing tier at ₹649 a month. This is not a decision point unless you build from India. There, it is a cheaper way to try agentic coding day to day.
OpenAI opened free access to its most advanced models for researchers at select universities, expanding toward 100,000 people over the next year. That is meaningful for scientific funding. It is not a change to how you build or ship a product.
What to watch
GitHub’s default model enablement policy takes effect August 26. Models you have not explicitly configured will start following your enterprise’s default setting instead of staying off. If you rely on new models being unavailable until someone turns them on, set your opt-out policy now. Do it before the switch flips.
Watch whether Anthropic widens access to Claude Mythos beyond its current partner group. Also watch whether other post-quantum candidates face the same scrutiny that just exposed a flaw in HAWK. A model found the flaw in 60 hours after human reviewers spent two years on it. That is worth tracking as an ongoing trend, not a one-time headline.
This week’s practical move is auditing defaults, not chasing features. Check your Copilot model policies before August 26. Confirm your Netlify and Vercel projects are as private or capped as you think. Test any per-user connector with two accounts that see different things.
End of article