The week belonged to Anthropic, in two very different ways. On June 30 it launched Claude Sonnet 5, a cheaper and more capable mid-size model aimed squarely at the agent workloads that now dominate developer spending. The same day, the US government lifted the export controls that had kept its most advanced models, Fable 5 and Mythos 5, dark for 19 days, and Fable 5 came back online worldwide on July 1. After last week’s government-gated GPT-5.6 preview, the state’s hand on frontier AI is still very much visible, only this time it loosened its grip.

The rest of the week filled in around those two stories. GitHub Copilot became a native agent inside JetBrains IDEs, Meta claimed its next model had caught up to GPT-5.5, and a Cloudflare failure took npm and a string of other services offline right in the middle of the workday. Mastra shipped scheduled agents, CloudNativePG hit 1.30, Together AI raised $800 million, and a SharePoint zero-day landed on CISA’s catalog with a holiday-weekend deadline. Here is everything that mattered.

Top Stories This Week

Anthropic Ships Claude Sonnet 5 as a Cheaper Way to Run Agents -

On June 30, Anthropic released Claude Sonnet 5, calling it the most agentic Sonnet model yet, one that can make plans, drive browsers and terminals, and run autonomously at a level that just a few months ago required larger and more expensive models. It replaces Sonnet 4.6 from February as the default model for Free and Pro plans, and it is available across Max, Team, and Enterprise, in Claude Code, and via the claude-sonnet-5 API. It launched with introductory pricing of $2 per million input tokens and $10 per million output through August 31, before moving to $3 and $15.

The framing matters as much as the model. As TechCrunch noted, shipping agentic capability is now table stakes, so the real contest is who can run agents cheaply and reliably without a human watching. Sonnet 5 promises performance close to Opus 4.8 for much less money, undercutting GPT-5.5 and Gemini 3.1 Pro while staying pricier than Gemini 3.5 Flash. One catch worth budgeting for: Sonnet 5 uses a new tokenizer that produces about 30 percent more tokens for the same input, so the effective cost per task runs higher than the headline price, and it drops support for temperature, top_p, and top_k. It is a drop-in upgrade for 4.6, but check your token budgets before you flip the switch.

US Lifts Export Controls and Fable 5 and Mythos 5 Come Back Online -

The other half of Anthropic’s week was about getting models back. On June 30, Anthropic said the US Commerce Department had lifted the export controls it imposed on June 12, which had forced the company to suspend all access to Fable 5 and Mythos 5 for 19 days because it could not verify user nationality in real time. Fable 5 returned to users worldwide on July 1 across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork, with Pro, Max, and Team plans getting it for up to half of weekly usage limits through July 7. Mythos 5, which ships with fewer safeguards, was restored to a set of approved US organizations under the Glasswing defensive cybersecurity program following government sign-off on June 26.

As Ars Technica reported, the curbs came off about three weeks after the administration flagged the models as national security risks, and only after Anthropic trained a new safety classifier that reliably blocks the specific jailbreak, documented by Amazon researchers, that let the model surface software vulnerabilities. The episode is now the template for how frontier AI ships in 2026: models pass through Washington first, and the details of who can use them get negotiated behind closed doors. Anthropic is still working with Amazon, Microsoft, Google, and other partners on a shared framework for grading jailbreak severity, which suggests this will not be the last model to blink offline for a while.

GitHub Copilot Becomes a Native Agent in JetBrains IDEs -

On June 30, GitHub and JetBrains made Copilot a first-class option in the JetBrains AI Assistant agent picker. Copilot had already been reachable inside AI Assistant through the Agent Client Protocol (ACP), but this integration bundles it in directly, so it shows up in the agent list automatically for anyone with a valid Copilot subscription, with no endpoint to configure. You open the agent picker, select GitHub Copilot, choose your model and reasoning depth, and hand off multistep work that Copilot reasons through, runs commands for, and iterates on with you.

As JetBrains described it, the point is stability and zero setup: the agent is co-built, bundled, and updated with AI Assistant, and Copilot CLI slash commands like /remote and /chronicle now work straight from the chat. Copilot authenticates over OAuth through your GitHub account and still needs its own subscription, separate from JetBrains AI. The deeper story is the ACP ecosystem maturing. Cursor joined the same registry in March, and now the biggest coding assistant is a default citizen inside a rival IDE. For developers, the walls between coding tools keep getting thinner, and where your agent runs matters less than which one you trust.

Meta Says Its In-Training Watermelon Model Has Caught GPT-5.5 -

On July 2, at an internal town hall, Meta superintelligence chief Alexandr Wang told staff that the company’s next model, codenamed Watermelon, has caught up to OpenAI’s GPT-5.5 on closely watched benchmarks. Wang said Watermelon, the successor to the Muse Spark family Meta released in April under the codename Avocado, is still in training and uses an order of magnitude more compute than its predecessor. He did not name the benchmarks, there is no public preview or release date, and Meta declined to comment, so this is a leadership claim, not a verified result.

Still, it is a notable signal from a lab that has trailed OpenAI, Google, and Anthropic at the frontier despite enormous spending. As Tech Times pointed out, the reference point has already moved, since OpenAI previewed GPT-5.6 in late June, so matching GPT-5.5 is catching up to where the race was, not where it is. The part developers should watch is the open-weight angle: Meta has committed to releasing future Muse models with open weights, and Wang posted on July 3 that a near-term Muse Spark update will bring major coding and agentic gains. If a GPT-5.5-class model ships as open weights, teams running on-premises get a frontier-adjacent option they cannot buy today.

A Cloudflare Failure Takes npm and Other Services Offline -

Starting late on July 1 and running into July 2, Cloudflare hit a major network performance degradation across Eastern and Western North America that lasted around 10 hours and needed two rounds of remediation before it cleared. Because a lot of developer infrastructure sits behind Cloudflare, the blast radius reached well past Cloudflare’s own products. The npm registry, whose registry.npmjs.org endpoint runs on Cloudflare, went down for package installs, publishes, and logins, and developers on Hacker News tracked overlapping trouble at GitHub and other platforms during the same window.

The practical lesson is the familiar one, delivered on a slow holiday week: shared infrastructure means shared failure. When one CDN or one registry stumbles, CI pipelines stall, deploys hang, and a “productive coding day” evaporates for reasons that have nothing to do with your code. It is a good prompt to check whether your build can survive a registry outage, whether through a cache, a mirror, or a lockfile that does not force a network round trip for every install. None of this was an attack, just the internet’s concentration risk showing up again on schedule.

Mastra 1.48 Adds Scheduled Agents With Heartbeats -

On July 1, the TypeScript AI framework Mastra shipped 1.48 with heartbeats, a way to schedule agents to run on a recurring cron. A heartbeat can deliver into an existing conversation thread as a normal signal or run on its own, and the schedules persist across restarts, exposed through @mastra/core, @mastra/server, and @mastra/client-js. Setting one up is a few lines: give it an agentId, a cron expression, and a prompt, and the agent checks in on its own timetable. The same release adds file-based agents you define by dropping directories under src/mastra/agents/, a public Observational Memory Extractor API, and stronger human-in-the-loop recovery that can resurface a suspended run after a page refresh.

It is a small release on paper but it points at where agent frameworks are going. Agents that only respond to a user request are giving way to agents that wake up on a schedule, do work in the background, and report back, which is exactly the shape of the “digital coworker” pitch every lab is now making. Cron for agents sounds mundane until you realize it is the primitive that turns a chatbot into a service.

Developer Tools & Platforms

CloudNativePG 1.30 Ships Declarative Roles and Safer Failover -

On June 29, CloudNativePG 1.30.0 landed, the CNCF operator for running PostgreSQL on Kubernetes. The headline additions are the new DatabaseRole custom resource, which lets you manage Postgres roles declaratively in a GitOps workflow instead of running SQL by hand, and a lease-based primary election primitive that makes failover safer by reducing the chance of a split-brain during a promotion. The release also bumps the default PostgreSQL image to 18.4, adds support for Kubernetes 1.36, and backports security fixes like search_path pinning to the 1.29 and 1.28 maintenance lines. With 1.30 out, the 1.28.x series reaches end of life, so teams on it should plan an upgrade. If you run stateful Postgres on Kubernetes, the declarative role management alone is worth the look, since it pulls one more piece of database config into the same version-controlled model as the rest of your cluster.

Google Cloud Leans Into Agentic Data and Claude Tooling -

Google Cloud had a busy start to July. On July 1 it brought Conversational Analytics in BigQuery to general availability, an agent that lets business and technical users query data, run multi-step analyses, and build visual reports in natural language, reaching across native BigQuery tables, Iceberg, and cross-cloud sources like Databricks Unity and AWS Glue. The same day it announced the GA of Nano Banana 2 Lite and a public preview of Gemini Omni Flash on the Gemini Enterprise Agent Platform, aimed at fast, cheap image and video generation. And on July 2, in a notable cross-vendor move, Google published a guide to the Claude apps gateway for Google Cloud, a self-hosted service that sits between local Claude Code clients and Google Cloud so enterprises can broker and secure that traffic on their own infrastructure.

AWS Launches Interconnect Last Mile With AT&T -

On June 30, AWS announced Interconnect last mile, a managed connectivity service that lets customers connect branch offices, data centers, and remote sites to AWS in a few clicks. The first partner is AT&T, in a US gated preview. You choose a Region, bandwidth, Direct Connect Gateway ID, and partner subscriber ID, and AWS pre-provisions capacity and automates the fiddly parts like BGP peering, VLAN configuration, and ASN assignment, backed by an SLA and zero-downtime maintenance. It is a networking, not an AI, launch, but it targets a real source of friction: getting private, reliable pipes into the cloud without weeks of telecom coordination. Partners can integrate through a published open API on GitHub.

Security

SharePoint Deserialization Zero-Day Hits CISA’s Catalog With a July 4 Deadline -

On July 1, CISA added CVE-2026-45659, a remote code execution flaw in Microsoft SharePoint Server, to its Known Exploited Vulnerabilities catalog with a July 4 deadline for federal agencies. The bug is a deserialization of untrusted data with a CVSS score of 8.8, and it needs only low Site Member permissions, the standard level for anyone who can contribute to a SharePoint site, including guest users on externally shared sites. That is a low bar for a foothold into enterprise document stores, HR portals, and intranets.

The wrinkle, as Threat-Modeling.com documented, is the disclosure timeline. Microsoft shipped the patch during the May 2026 Patch Tuesday cycle but forgot to publish the associated security bulletin, only adding it weeks later on May 21 with an assessment of “less likely” to be exploited. Any team relying solely on Microsoft’s published bulletins, rather than independently scanning patch levels, may have quietly skipped this fix. CISA has now confirmed active exploitation. If you run on-premises SharePoint, patch to the fixed builds now, then audit Site Member accounts and logs for signs of compromise rather than assuming the update alone closes the book.

Funding & Industry Deals

Together AI Raises $800M and Kling AI Pulls In Nearly $3B -

On July 1, Together AI raised an $800 million Series C at an $8.3 billion valuation, led by Aramco Ventures with Nvidia, Vista, General Catalyst, and others joining. The neocloud, which rents Nvidia GPU clusters and specializes in running open-source models cheaply, more than doubled its valuation from the $3.3 billion of its Series B 16 months ago, and says it plans to grow capacity roughly 50-fold over five years against more than $1.1 billion in annual bookings. It is the clearest sign yet that the money chasing AI inference infrastructure has not slowed. The same window saw Kuaishou’s Kling AI raise nearly $3 billion at an $18 billion valuation on July 2, backed by Alibaba, Tencent, and Baidu, positioning the Chinese video model as one of the few in the category with real revenue behind the hype after OpenAI shut down Sora’s consumer app.

Datadog Buys Adaptive ML and Schneider Buys Cognite -

On June 30, Datadog acquired Adaptive ML, a startup building a reinforcement learning operations (RLOps) platform, to fold into Datadog AI Research. The bet is that Datadog’s firehose of real-world observability signals gives it the production-scale data needed to post-train specialized agents for monitoring and security, an area where the algorithm was never the hard part but scale was. Terms were not disclosed, and Datadog shares climbed on the news. Separately, Schneider Electric agreed to acquire industrial data and AI firm Cognite for $3.1 billion in all cash on June 30, integrating it with AVEVA inside its industrial automation business, a sign that industrial AI is drawing big strategic checks alongside the consumer and developer tooling that usually grabs headlines.

Layoffs: Microsoft, TikTok, and Cisco

The Numbers That Matter

  • $2 / $10 Claude Sonnet 5’s introductory price per million input and output tokens through August 31
  • ~30% More tokens Sonnet 5’s new tokenizer produces for the same input, raising effective cost
  • 19 days How long Fable 5 and Mythos 5 were offline before export controls were lifted
  • ~10 hours Duration of the Cloudflare North America degradation that hit npm on July 1 and 2
  • 18.4 Default PostgreSQL version in CloudNativePG 1.30, alongside Kubernetes 1.36 support
  • $800 Million Together AI’s Series C, at an $8.3 billion valuation
  • $18 Billion Post-money valuation for Kling AI’s near-$3 billion round
  • $3.1 Billion Schneider Electric’s all-cash acquisition of Cognite
  • July 4 CISA deadline to patch the SharePoint zero-day CVE-2026-45659

Quick Hits

  • CloudNativePG 1.30 - June 29. DatabaseRole CRD and lease-based failover for Postgres on Kubernetes, default PG 18.4.
  • Claude Sonnet 5 - June 30. Anthropic’s most agentic mid-size model ships at $2 per million input tokens.
  • Fable 5 and Mythos 5 Controls Lifted - June 30. US lifts export curbs after 19 days; Fable 5 returns globally July 1.
  • GitHub Copilot Native in JetBrains - June 30. Copilot becomes a first-class agent in the AI Assistant picker.
  • Datadog Buys Adaptive ML - June 30. RLOps startup joins Datadog AI Research to post-train observability agents.
  • Schneider Buys Cognite - June 30. $3.1 billion all-cash deal for industrial data and AI.
  • AWS Interconnect Last Mile - June 30. Managed connectivity to AWS launches in gated preview with AT&T.
  • Cisco WARN Filings - June 30. 471 Bay Area cuts detailed, software engineer the most-affected title.
  • Microsoft Layoffs Reported - June 30. Under 2.5% of the workforce, spanning Xbox, sales, and consulting.
  • Together AI Series C - July 1. $800 million raised at an $8.3 billion valuation for AI inference.
  • Mastra 1.48 - July 1. Heartbeats bring cron-scheduled agents plus file-based agent definitions.
  • Conversational Analytics in BigQuery GA - July 1. Natural-language querying agent across BigQuery and other clouds.
  • Nano Banana 2 Lite GA - July 1. Fast, cheap image generation lands on Gemini Enterprise Agent Platform.
  • Cloudflare Outage Hits npm - July 1 and 2. A 10-hour North America degradation knocks npm installs and others offline.
  • SharePoint Zero-Day on KEV - July 1. CISA sets a July 4 deadline for CVE-2026-45659.
  • TikTok Trust and Safety Layoffs - July 1. Cuts across Singapore, Indonesia, and Malaysia as AI takes over moderation.
  • Claude Apps Gateway for Google Cloud - July 2. Self-hosted broker between local Claude Code clients and Google Cloud.
  • Kling AI Raises Nearly $3B - July 2. Kuaishou’s video model hits an $18 billion valuation with Alibaba and Tencent backing.
  • Meta’s Watermelon Claim - July 2. Alexandr Wang tells staff the in-training model has caught GPT-5.5 on benchmarks.

The through line this week was the state and the switch, in every sense. Anthropic flipped Fable 5 back on the moment Washington let it, and shipped Sonnet 5 the same day to make agents cheap enough to run everywhere. Meanwhile the plumbing underneath all of it wobbled, with a single Cloudflare fault taking npm down and reminding everyone how few places the internet actually lives. Add a SharePoint bug that Microsoft nearly forgot to disclose, layoffs that keep landing hardest on engineers, and yet another billion-dollar round for the companies that rent out GPUs, and you get a fair snapshot of mid-2026: more capable models, more concentrated infrastructure, and less slack in the system every week. Next week Microsoft’s cuts should become official and GPT-5.6’s broader rollout may finally start to open up. See you then.