Updates for 5/19

Announced "Grok Build," a coding agent positioned against Claude Code and Codex (available via the $300/mo SuperGrok Heavy tier)

Until last month, xAI's Grok was mostly known as a chat tool with real-time X data access — not somewhere you'd go for multi-file agentic coding. For the past six months, Claude Code, Cursor, and Codex had been the main players carving up the coding-agent market. Grok Build is xAI's first move into that space.

Right now this only affects SuperGrok Heavy subscribers paying $300/month — regular Grok and free users won't see a difference today. But as a third serious contender enters the ring, benchmark comparisons and pricing pressure will only increase. xAI's real-time X data could give it an edge for coding tasks tied to live trends. For developers in the $20–40/month range, this is background noise for now.

Partnered with Dell to deliver Codex in hybrid and on-prem environments, enabling enterprises to deploy the AI coding agent securely across data and workflows

Until now, Codex was cloud-only — you had to send your code to OpenAI's servers. For the past year, that was the deal-breaker for financial, healthcare, and defense teams who couldn't move code outside their firewall. Lots of POCs stalled here. The Dell partnership changes that, opening up hybrid and on-prem delivery.

If 'we can't send code to the cloud' was the thing stopping your organization from adopting Codex, that excuse just got much thinner. Being able to say 'it runs entirely on Dell infrastructure' is exactly what approval processes need. For individual developers or small teams already running in the cloud, nothing changes today. On-prem support also becomes a new differentiator when comparing Codex vs. Cursor vs. GitHub Copilot.

Composer 2.5 released — built on Kimi K2.5 and trained on 25x more synthetic tasks; matches Opus 4.7 / GPT-5.5 benchmarks at a fraction of the cost

Until last month, heavy Cursor users often switched to Max mode or a different tool when they needed top-tier reasoning — the built-in Composer wasn't benchmark-competitive at the high end. For the past few months, a Kimi K2-based migration had been rumored, but the production model stayed the same. Composer 2.5 on Kimi K2.5, trained on 25x more synthetic data, is the official move.

Opus 4.7 / GPT-5.5-level benchmark performance inside Cursor Pro at $20/month is the real headline here. Claude Max starts at $100/month; Codex adds per-token costs on top. For developers who want heavy multi-file editing at a flat monthly rate, Cursor's value case just got stronger. Real-world feel depends on the task — large-scale refactors across multiple files are where you'll likely notice the difference first.

Copilot Cowork now generally available — an autonomous agent that understands org context via "Work IQ," runs tasks in the background from the cloud, supports reusable Skills, and integrates with M365, Power BI, Dynamics 365, monday.com, Miro, and more

For the past year, Copilot was a 'you ask, it answers' assistant. Meeting summaries in Teams, formula suggestions in Excel — useful, but always reactive. Background execution features existed in preview, but 'hand it off and come back later' workflows weren't production-ready for most teams. As of today, Copilot Cowork is generally available.

If your organization pays the $30/user/month M365 Copilot add-on, weekly reports, customer follow-ups, and Dynamics updates can now be handed to an agent while you do something else. The 'Work IQ' context engine should make it more org-aware than generic ChatGPT for internal tasks. That said, if you're not on M365 Copilot, today's announcement doesn't touch you at all — this is firmly enterprise territory.

Malta government to give all citizens free ChatGPT Plus for one year upon completing a national AI course (national-scale rollout)

Until last month, ChatGPT adoption happened through individual subscriptions or enterprise contracts — no government had bought it in bulk and distributed it nationally. Public AI procurement was mostly analytics and infrastructure tools; conversational AI for all citizens didn't exist as a model. Malta (population ~500K) changed that by making Plus free for one year to any citizen who completes a national AI course.

Malta alone doesn't move the market. But if this works, other governments will follow. It's the first concrete example of AI shifting from 'a SaaS you pay for yourself' to 'a public good funded by the state.' For product and policy people watching how AI adoption scales beyond tech workers, this is an important precedent. For most readers, nothing changes today — but watch for other countries repeating this model.

Acquired Stainless, a NY startup (founded 2022) whose SDK-generation tooling is used by OpenAI, Google, and Cloudflare

Stainless is the startup that takes an API spec and auto-generates polished, multi-language SDKs — OpenAI, Google, and Cloudflare all use it for their own official client libraries. Until last month, Anthropic was also an external customer, relying on Stainless tooling for Claude's SDKs. Bringing that capability in-house is the move today.

If you build on the Claude API, expect SDK updates and multi-language support to get faster over the next 6–12 months. The day-to-day dev experience will improve incrementally, not overnight. For engineers not on Claude, nothing changes. The interesting side note: OpenAI and Google were also Stainless customers — some of the tooling that generates their SDKs now sits under Anthropic's roof.

To brief major treasuries and central banks on cyber-defense vulnerabilities in the global financial system uncovered by Claude Mythos Preview (The Decoder)

Last week, Claude Mythos Preview made headlines for finding 271 unknown Firefox vulnerabilities autonomously via Mozilla's pipeline — including bugs 20 years old. That same model has now been pointed at global financial infrastructure, and Anthropic is preparing to brief major treasuries and central banks on what it found. Until a month ago, an AI independently discovering security holes and carrying them to governments was largely hypothetical.

For financial institutions and government security teams, this signals that AI-driven audits at national-infrastructure scale are here, not coming. The pace at which AI replaces manual penetration testing may be moving faster than most security roadmaps assume. For everyday users and small businesses there's no direct impact today — though 'global financial infrastructure gets patched' is good for everyone in the background.

NVIDIA shipped "Vera," its first CPU designed for AI agents, to Anthropic, OpenAI, SpaceXAI, and Oracle Cloud (a dedicated CPU now joins the GPU-centric landscape)

AI compute has been GPU-dominated since the transformer era. CPUs handled server management and auxiliary work; the interesting AI tasks happened on GPUs. Over the past few months, as Blackwell GPU supply stabilized, the question of 'what's next to bottleneck?' was getting louder in infrastructure conversations. Today, NVIDIA started shipping 'Vera,' an agent-optimized CPU, to Anthropic, OpenAI, SpaceXAI, and Oracle Cloud.

Nothing changes for you today — but if you run agents on cloud infrastructure, latency and cost improvements could reach you in 6–12 months as these CPUs work their way into production data centers. Large enterprise IT teams doing infrastructure procurement now have a new category to evaluate. For individual developers and most smaller teams, this stays background noise for a while.

Elon Musk's $134B lawsuit dismissed by jury verdict (claims time-barred), removing a reported pre-IPO restructuring threat

Elon Musk's $134 billion lawsuit against OpenAI had been hanging over the company's IPO plans and corporate restructuring for more than six months. The claim — that OpenAI's shift to for-profit violated founding commitments — had the potential to block or reshape the restructuring if it succeeded. Today a jury dismissed all claims on statute-of-limitations grounds.

For potential OpenAI investors and enterprises considering long-term contracts, one layer of legal uncertainty just cleared. If an IPO moves forward, having this resolved matters for valuation and investor confidence. This doesn't change API pricing or model availability today — and if you're an individual user or developer on a small plan, it's essentially irrelevant. For executives deciding whether to build long-term on OpenAI infrastructure, it's a meaningful data point.