The frontier model drought ended in a single midweek surge. On July 8 SpaceXAI put Grok 4.5 in developers’ hands at $2/$6. On July 9 OpenAI followed with the full public GPT-5.6 family and a new ChatGPT Work agent built on top of it. After weeks of government-gated previews and export-control drama, every major Western lab again had a publicly reachable frontier or near-frontier option, and the real fight shifted from access to price, latency, and which agent harness you trust.

The rest of the week was just as consequential for day-to-day engineering. npm v12 flipped install-time trust on its head, TypeScript 7 finally shipped the native Go compiler, Microsoft cut thousands of jobs, Apple took OpenAI to court over hardware secrets, and two separate research disclosures showed how AI coding agents can be tricked through symlinks and images. SambaNova raised another billion, AWS shipped an enterprise gateway for Claude Code, and SK hynix listed ADRs on Nasdaq in a $26.5 billion offering as AI memory stayed the market’s favorite shovel. Here is everything that mattered.

Top Stories This Week

OpenAI Ships GPT-5.6 Sol, Terra, and Luna for General Availability -

On July 9, OpenAI launched the GPT-5.6 family for general availability after the limited, government-coordinated preview that began in late June. Sol is the flagship, Terra is the lower-cost everyday tier competitive with GPT-5.5, and Luna is the fastest and cheapest option. API pricing is $5 input and $30 output per million tokens for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna. The models roll out across ChatGPT, Codex, and the OpenAI API, with Programmatic Tool Calling in the Responses API and an ultra setting that coordinates multiple agents in parallel for the hardest work.

OpenAI’s pitch is performance per dollar, not just raw score. On its published evals, Sol sets new highs on coding-agent and long-horizon workflow benchmarks while using fewer tokens and less time than Claude Fable 5 on several comparisons, and Terra and Luna are framed as ways to get near-frontier results at a fraction of the cost. As The Verge noted, the public launch closes the regulatory chapter that kept GPT-5.6 limited to vetted partners for about two weeks. For developers, the practical move is to pin explicit model IDs rather than relying on latest aliases, then A/B Terra and Luna on your real task mix before you assume the middle tier is always the better buy.

SpaceXAI Launches Grok 4.5 at $2/$6, Trained With Cursor -

On July 8, SpaceXAI released Grok 4.5, calling it its smartest model for coding, agentic tasks, and knowledge work, and saying it was trained alongside Cursor. Pricing is $2 per million input tokens and $6 per million output tokens, with claimed serving speeds around 80 tokens per second and roughly 2x the token efficiency of comparable leading models on some software engineering tasks. It is available the same day in Grok Build, in Cursor on all plans, and from the SpaceXAI console, with a limited free-usage window in Grok Build and Cursor. EU access was held back for mid-July.

As TechCrunch reported, Elon Musk framed Grok 4.5 as an Opus-class model that is faster, more token-efficient, and lower cost, later saying internal assessments put it roughly comparable to Opus 4.7. Independent coding-agent indexes have it below Fable 5 on some harnesses while estimating a lower cost per task, so the launch is more a price and distribution story than a settled benchmark crown. The Cursor integration is the developer-facing hook: after SpaceX’s June acquisition of Anysphere, Grok 4.5 lands inside an IDE many teams already use, not only behind a new API.

ChatGPT Work Turns ChatGPT Into a Long-Running Work Agent -

Also on July 9, OpenAI launched ChatGPT Work, an agent inside ChatGPT that can pull context from connected apps and files, stay with a project for hours, and turn a goal into finished sheets, slides, docs, and web apps. It is powered by GPT-5.6, folds Codex technology into non-coding workflows, and ships with a unified plugins directory for tools like Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars, and CRMs. Scheduled Tasks can keep work moving when you are away, Sites lets you publish interactive pages from ChatGPT, and the redesigned desktop app merges Chat, Work, and Codex on Mac and Windows, including for Free users.

The competitive read is straightforward: this is OpenAI’s answer to Anthropic’s Claude Cowork and the broader “digital coworker” wave, aimed at knowledge workers who want agentic help without living in a terminal. Rollout starts with Pro, Enterprise, and Edu on web and mobile, then Plus and Business over the following days. Enterprise admins get spend controls, plugin governance, and a Compliance API. If you already use the Codex desktop app, updating it turns it into the new ChatGPT desktop app, with an option to keep Codex as the default view.

npm v12 Turns Install Scripts Off by Default -

On July 8, npm v12 became generally available and tagged latest, flipping the install-time defaults GitHub previewed in June. Dependency lifecycle scripts and implicit node-gyp builds no longer run unless explicitly allowed. Git dependencies and remote HTTPS tarball dependencies no longer resolve unless explicitly allowed. Teams can prepare with npm approve-scripts --allow-scripts-pending and commit the resulting allowlist in package.json. As Socket explained, skipped scripts warn and continue by default, which means native modules like sharp or better-sqlite3 can install cleanly and then fail at runtime if nobody approved the build.

The same release begins winding down 2FA-bypass granular access tokens. Sensitive account, package, and org actions are expected to require interactive 2FA in early August 2026, and direct publishing from those tokens is expected to shrink to staged publishing around January 2027. For CI, the migration path is trusted publishing with OIDC or staged publishing with a human approval step. After a year of install-time worms and Phantom Gyp style attacks, npm is finally matching the safer defaults pnpm already shipped. Expect broken pipelines this week if you upgrade without an allowlist.

TypeScript 7 Ships as a Native Go Compiler -

On July 8, Microsoft announced TypeScript 7 as generally available, the native Go port of the compiler and language service that the team has been building for more than a year. Install is the usual npm install -D typescript, and Microsoft says full builds typically land between 8x and 12x faster thanks to native code speed, shared-memory multithreading, and new optimizations. Editor support moves to the language server protocol, with a dedicated VS Code extension and automatic enablement in Visual Studio based on the workspace.

The caveat for framework teams is the programmatic API. TypeScript 7.0 prioritizes the compiler and editor path; a stabilized API for tools that embed TypeScript, the ones Vue, Svelte, Astro, and MDX template checkers rely on, is planned for the 7.1 cycle. Next.js, React, and plain Node services can move now. Teams whose editor tooling still imports TypeScript as a library should stay on 6.x until 7.1, or run a split stack. After years of type-check CI as the slowest job in the pipeline, this is the rare language release that can change wall-clock build time overnight.

Apple Sues OpenAI Over Alleged Trade Secret Theft -

On July 10, Apple sued OpenAI in Northern California federal court for trade secret misappropriation and breach of contract, accusing OpenAI of a leadership-directed campaign to extract confidential hardware information as it builds consumer devices. The complaint names OpenAI Chief Hardware Officer Tang Yew Tan, formerly Apple’s VP of product design for iPhone and Apple Watch, and former Apple engineer Chang Liu, and it notes that more than 400 former Apple employees now work at OpenAI. Apple alleges recruiting practices that solicited designs and prototypes, misuse of supplier relationships, and downloads of confidential specs for unannounced products.

OpenAI said it has no interest in other companies’ trade secrets and remains focused on building its own technology. The filing lands weeks before OpenAI’s widely reported confidential IPO preparations and in the middle of a deteriorating commercial relationship that once put ChatGPT on Apple platforms. For the broader industry, it is the clearest signal yet that the AI talent war has moved from signing bonuses into discovery. Every frontier lab that has been hiring silicon and hardware teams out of Big Tech will be reading the complaint for what Apple thinks it can prove.

Microsoft Cuts 4,800 Jobs as Xbox Begins Its Biggest Reset -

On July 6, Microsoft said it is immediately eliminating 4,800 jobs, about 2.1 percent of its workforce, after a voluntary retirement program earlier in the year. Xbox takes the heaviest hit: CEO Asha Sharma told staff the division will cut about 3,200 roles through fiscal 2027, including roughly 1,600 on day one, amounting to about one-fifth of Xbox headcount. Compulsion Games and Double Fine will become independent again, while Ninja Theory and Undead Labs move to new ownership. Commercial sales roles are also reduced.

Chief People Officer Amy Coleman framed the cuts around a fast-changing industry and said AI is not replacing the laid-off workers, even as it changes how work gets done. The timing matches Microsoft’s weak 2026 stock performance and heavy AI infrastructure spend, with Xbox revenue already shrinking and console hardware costs rising. For engineers, the message is familiar across Big Tech this year: payroll is the financing layer for the AI capex cycle, and gaming studios remain easier to restructure than cloud capacity commitments.

Developer Tools & Platforms

AWS Ships a Claude Apps Gateway for Enterprise Claude Code -

On July 8, AWS announced the Claude apps gateway for AWS, a self-hosted control plane for Claude Code and Claude Desktop. Instead of handing every developer a personal cloud credential and pushing settings laptop by laptop, enterprises deploy a stateless gateway on ECS, EKS, or EC2, back it with PostgreSQL, authenticate through their existing OIDC identity provider, and enforce model access, tool permissions, and per-user or per-group spend caps centrally. Inference can route to Amazon Bedrock or Claude Platform on AWS, including cross-Region setups. For platform teams that want Claude Code at scale without secret sprawl, this is the missing enterprise middle layer, and it mirrors the Google Cloud Claude apps gateway pattern that landed the prior week.

GPT-5.6 Becomes the Preferred Model in Microsoft 365 Copilot -

Also on July 9, OpenAI and Microsoft made GPT-5.6 the preferred model series in Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat, and Cowork. The update is the distribution half of the GPT-5.6 launch: the same model family that just opened in the API becomes the default intelligence layer inside the office suite millions of enterprise users already pay for. Microsoft will serve the models natively and also reach them through the OpenAI API. For teams standardizing on Copilot, the model upgrade arrives without a separate procurement cycle, which is often how frontier capability actually reaches corporate desktops.

Security

GhostApproval: Symlink Tricks Bypass AI Coding Agent Approvals -

On July 8, Wiz publicly disclosed GhostApproval, a trust-boundary pattern affecting six major AI coding assistants: Amazon Q Developer, Anthropic Claude Code, Augment, Cursor, Google Antigravity, and Windsurf. A malicious repo plants a symlink that looks like a harmless project file, such as project_settings.json, but resolves to something like ~/.ssh/authorized_keys. When the developer asks the agent to set up the workspace, the agent follows the link. In several tools the approval UI still showed the innocent filename even when the agent internally knew the real target, which Wiz frames as symlink following plus UI misrepresentation.

AWS, Cursor, and Google shipped fixes. Cursor assigned CVE-2026-50549. Anthropic argued the scenario sits outside its threat model once a user trusts a directory and approves an edit, and pointed to earlier symlink warnings in Claude Code. Augment and Windsurf had not fully closed the issue at disclosure. The practical advice is blunt: resolve symlinks before prompting for approval, show the canonical path when a write leaves the workspace, and treat “Accept” dialogs that hide the real destination as broken consent.

Ghostcommit Hides Prompt Injection Inside Images Reviewers Never Open -

On July 11, researchers from the University of Missouri-Kansas City’s ASSET group detailed Ghostcommit, an attack where a pull request adds an AGENTS.md convention that points at a PNG. The malicious instructions live as readable text inside the image: read .env, encode each byte as an integer, and emit the result as a module constant. Text-based PR reviewers and bots that skip images wave it through. Later, when a developer asks a vision-capable coding agent for a routine feature, the agent follows the convention, opens .env, and writes the secrets into source as a number list that secret scanners do not treat as credentials.

In end-to-end runs, Cursor driving Claude Sonnet leaked a full .env on the first try. Claude Code, running the same model family, refused. That split is the story: the harness mattered more than the weights. Defenses need multimodal PR review that actually opens images, plus runtime monitoring when an agent touches credentials it had no reason to read. If your agent auto-loads convention files and can see images, treat unreviewed binary assets in those conventions as untrusted code.

Funding & Industry Deals

SambaNova Raises $1B at an $11B Valuation -

On July 8, SambaNova raised $1 billion in a Series F first close at an $11 billion valuation, led by General Atlantic with Intel Capital, BlackRock, Vista, QIA, T. Rowe Price, Battery, and others participating. The round comes about five months after a $350 million Series E and the SN50 chip unveil, and it pairs with news that JPMorgan Chase selected SambaNova as an inference-infrastructure partner for on-premises serving on SN40L and SN50 systems. CEO Rodrigo Liang said the money is going into supply chain and delivery capacity for the next year. It is another reminder that inference silicon and private enterprise racks are raising at neocloud scale, even as model vendors race each other on list price.

SK Hynix Lists ADRs on Nasdaq in a $26.5B Offering -

On July 10, SK hynix listed American Depositary Receipts on Nasdaq after an Opening Bell ceremony in Times Square. As Yonhap reported, the offering covered about 177.9 million ADRs priced at $149 each, raising roughly $26.5 billion, with each ADR representing one-tenth of a Seoul-listed common share. The company said proceeds support new fabrication and advanced packaging capacity in South Korea, and framed the listing as a way to broaden its US investor base as demand for high-bandwidth memory keeps AI accelerator supply constrained. It is one of the clearest public-market bets on the picks-and-shovels layer underneath the model price war.

Layoffs: Microsoft, and a Viral “Claude Fable 5 Is Enough” Claim

  • Microsoft: On July 6, Microsoft cut 4,800 jobs immediately, or about 2.1 percent of its workforce. Xbox will shed about 3,200 roles through FY27, including 1,600 on day one, and spin out or rehome Compulsion Games, Double Fine, Ninja Theory, and Undead Labs. Commercial sales teams were also reduced. The company said AI is changing how work gets done but is not directly replacing the eliminated roles.
  • Unnamed services firm (Reddit claim): On July 9, Times of India covered a viral Reddit post from a developer who said their mid-sized services company cut a team of about 70 down to eight in one round, with leadership telling the remaining staff that Claude Fable 5 would be enough to handle the work left. The company was not named, and the claims have not been independently verified. Commenters split between panic over AI replacing engineers and skepticism that “layoffs due to AI” is cover for fewer clients and weaker revenue. Either way, the post hit a nerve in the same week Fable 5 moved off included subscription access and onto usage credits.

The Numbers That Matter

  • $5 / $30 GPT-5.6 Sol API price per million input and output tokens
  • $2 / $6 Grok 4.5 API pricing
  • 8x to 12x Typical full-build speedup claimed for TypeScript 7’s native Go compiler
  • 4,800 Microsoft jobs cut on July 6, about 2.1 percent of the workforce
  • $1 Billion SambaNova Series F first close, at an $11 billion valuation
  • $26.5 Billion SK hynix Nasdaq ADR raise at $149 per ADR

Quick Hits

  • Microsoft Layoffs - July 6. 4,800 jobs cut immediately; Xbox begins a multi-year reset and studio spinouts.
  • Fable 5 Billing Cliff - July 7 to 8. Anthropic’s post-export-control included Fable 5 access for Pro and Max ends; usage credits at $10/$50 take over.
  • Grok 4.5 - July 8. SpaceXAI ships its Cursor-trained coding and agent model at $2/$6.
  • npm v12 - July 8. Install scripts, git deps, and remote tarballs become opt-in; GAT 2FA-bypass wind-down begins.
  • TypeScript 7.0 - July 8. Native Go compiler goes GA with typical 8x to 12x faster builds.
  • Claude Apps Gateway for AWS - July 8. Self-hosted control plane for enterprise Claude Code identity, policy, and spend.
  • SambaNova Series F - July 8. $1 billion at an $11 billion valuation; JPMorgan named as inference partner.
  • GhostApproval Disclosure - July 8. Wiz details symlink approval bypasses across six AI coding assistants.
  • GPT-5.6 GA - July 9. Sol, Terra, and Luna open across ChatGPT, Codex, and the API.
  • ChatGPT Work - July 9. OpenAI’s long-running productivity agent ships with plugins, Sites, and a unified desktop app.
  • GPT-5.6 in Microsoft 365 Copilot - July 9. GPT-5.6 becomes the preferred model in Word, Excel, PowerPoint, Chat, and Cowork.
  • “Claude Fable 5 Is Enough” Layoffs - July 9. Viral Reddit post claims a services firm cut ~70 developers to 8; company unnamed, unverified.
  • Apple v. OpenAI - July 10. Apple files a trade-secret suit over alleged hardware-theft via recruiting and ex-employees.
  • SK hynix ADR - July 10. $26.5 billion Nasdaq listing tied to HBM demand for AI accelerators.
  • Ghostcommit - July 11. Researchers show image-embedded prompt injection that steals .env secrets through coding agents.

The week closed the access gap and opened a reliability gap. GPT-5.6 and Grok 4.5 put frontier-class models back in every developer’s selector, ChatGPT Work tried to own the non-terminal half of agentic work, and npm v12 finally stopped treating every dependency script as trusted by default. At the same time, GhostApproval and Ghostcommit showed that the approval dialogs and PR bots wrapped around those models are still easy to fool, Apple reminded the industry that hardware ambitions come with discovery risk, and Microsoft’s cuts showed again who pays for the AI buildout. Even an unverified Reddit claim that Claude Fable 5 let a services firm keep eight developers out of seventy was enough to set off another round of “is this the future” anxiety. Next week watch for Gemini 3.5 Pro’s reported July 17 window, Grok 4.5’s EU rollout, and whether npm v12 allowlists become the new default chore in every JavaScript monorepo. See you then.