Weekly AI Dev News Digest: May 16 - May 22, 2026

Cursor released Composer 2.5 on Monday, and the pricing is the most aggressive move yet in agentic coding. The model runs on Moonshot's open-source Kimi K2.5, with the bulk of compute spent on Cursor's own reinforcement-learning post-training, around 25x more synthetic coding tasks than Composer 2. It scores 79.8% on SWE-Bench Multilingual against Opus 4.7's 80.5%, and beats Opus on CursorBench v3.1. The standard tier costs $0.50 per million input tokens and $2.50 per million output, roughly a tenth of frontier API rates; the fast tier that runs by default in the IDE is $3 and $15. Cursor doubled included usage for the launch week and disclosed a separate deal with SpaceXAI to train a much larger model from scratch on Colossus 2, with around 10x the total compute and no release date. (Cursor Composer 2.5)

One licensing wrinkle matters here. Kimi K2.5 ships under a Modified MIT license that requires visible attribution for any service over 100M monthly active users or $20M in monthly revenue, and Cursor's reported $2B-plus ARR clears that bar several times over. Cursor first disclosed the Kimi base in March after pushback on Composer 2, and this time it named Moonshot in the opening line of the announcement. (Pasquale Pillitteri analysis)

The race around it filled out fast. Codex got a midweek update that turns goals on by default with their own storage, makes codex remote-control behave like a foreground command with daemon-style start and stop, and adds permission profiles, plugin discovery, and extension hooks (Codex Release Notes). xAI's Grok Build, the terminal agent it launched on May 14, plans multi-step projects and spawns up to 8 concurrent sub-agents, each in its own Git worktree so parallel edits do not collide; it leans on Grok 4.3's 2M-token context for large refactors and runs through SuperGrok Heavy at roughly $300 a month, with a $99 promo tier for the first six months (xAI Grok Build). Musk had openly admitted xAI was behind on coding agents, and this is the catch-up bid. Anthropic's May 19 Claude Code update was a reliability push: stronger plugin dependency handling, /resume for background sessions, an agent view for managing parallel sessions from one place, more PowerShell support on Windows, Opus 4.7 as the new fast-mode default, and a fix for the 75-second startup hang behind captive portals (Claude Code Changelog). The model under most of Google's demos, Gemini 3.5 Flash, arrived the same week, beating Gemini 3.1 Pro on most coding and agentic benchmarks while running about 4x faster in tokens per second; Google also previewed Gemini Omni Flash, a unified model that takes image, audio, video, and text in and generates editable video out (IO Collection).

Model | Input / M tokens | Output / M tokens | SWE-Bench Multilingual Composer 2.5 (standard) | $0.50 | $2.50 | 79.8% Composer 2.5 (fast, IDE default) | $3.00 | $15.00 | 79.8% Claude Sonnet 4.6 | $3.00 | $15.00 | comparable Claude Opus 4.7 | ~$15.00* | ~$75.00* | 80.5% GPT-5.5 | ~$5.00* | ~$30.00* | comparable

*Frontier list prices per Cursor's published comparisons; verify with the vendor.

Six months after launch, Google rebuilt Antigravity into a standalone agent platform: a desktop app, a new CLI written in Go, an SDK for self-hosting custom agents, and Managed Agents in the Gemini API that spin up a sandboxed agent from a single call. Antigravity 2.0 runs parallel subagents, scheduled background tasks, and voice commands, with cross-platform terminal sandboxing and credential masking built in. The cost of all that focus: Google is killing Gemini CLI and the Gemini Code Assist IDE extensions for consumer tiers on June 18, with enterprise customers keeping access (Antigravity) (TNW).

Anthropic pushed in the same direction from the other end. Its Managed Agents now support self-hosted sandboxes in public beta and MCP tunnels in research preview. The orchestration loop, meaning context management and error recovery, stays on Anthropic's infrastructure, while tool execution moves into the customer's own sandbox so files, repos, and packages never leave the perimeter. Supported runners are Cloudflare, Daytona, Modal, and Vercel, or any sandbox client. MCP tunnels let an agent reach a private MCP server over one outbound, end-to-end-encrypted connection with no inbound firewall rules. Amplitude, Clay, Rogo, and Mason are the named early users (Claude Managed Agents).

The on-prem pitch came with hardware behind it this time. OpenAI announced a Dell partnership to connect Codex with the Dell AI Data Platform and AI Factory, aimed at companies that want agentic coding inside their own infrastructure rather than shipping code to OpenAI's cloud; OpenAI says Codex now has 4 million weekly developers and is used for report prep, lead qualification, and incident response beyond coding (OpenAI x Dell). Dell built out the rest of that story at Dell Technologies World, where it now runs Gemini 3 Flash on PowerEdge servers in confidential computing, hosts Codex, runs Palantir Foundry and Reflection AI models on premises, and offers SpaceXAI's Grok as a hybrid assistant, with 5,000 customers reportedly deploying the AI Factory (Next Platform) (Dell Blog). Google added a coding stack of its own: an Android Studio migration agent that rewrites React Native, web, or iOS apps to native Kotlin, a stable Android CLI any agent can drive, and Google AI Studio gaining Kotlin support, Firebase, and one-click Cloud Run deploy with export straight into Antigravity (Dev Keynote). And Apple approved Replit's first iPhone update in four months on May 18, ending a dispute over how AI-built apps get previewed on iOS and shipping Replit Agent 4 (TUAW).

MCP turned up as a shipping feature in four unrelated places in one week. Adoption like that, all at once and across products that never coordinated, is usually what a default looks like before anyone calls it one. Google previewed WebMCP, an open standard that lets sites expose structured tools, JavaScript functions and HTML forms, so browser agents call them directly instead of scraping the DOM; the experimental origin trial starts in Chrome 149, and Gemini in Chrome will support it (Chrome at IO 26). Red Hat shipped a developer-preview MCP server inside RHEL 10.2 and 9.8, putting the protocol into the base operating system, alongside an optional goose assistant in the command line (Red Hat). Zendesk shipped both an MCP client and an MCP server at its Relate conference, so its agents can call external tools and outside systems can read Zendesk data, and paired it with outcome-based pricing that only charges for resolutions a separate evaluation model confirms (Zendesk). Anthropic's MCP tunnels, above, make four.

Provenance standardized in the same stretch, with even less fanfare. During the keynote Google said OpenAI, ElevenLabs, and Kakao are adopting its SynthID watermark, and OpenAI confirmed the same day, layering SynthID on top of the C2PA Content Credentials it already attaches to images from ChatGPT, Codex, and the API (OpenAI Provenance) (TechCrunch). The two systems are built to complement each other: metadata gets stripped by social platforms, while SynthID survives screenshots, cropping, and re-encoding. With OpenAI on board, every major closed-model provider now signs images on the same stack. The open-weight side, FLUX, Llama, and Stable Diffusion forks, cannot be forced to participate, which is the gap that remains. Internet Pros has a solid overview of where C2PA and SynthID stand in 2026, including the EU AI Act Article 50 labeling rules now in force for the closed providers (Internet Pros).

Intuit is laying off about 3,000 of its roughly 18,200 employees across seven countries. In a May 20 memo, CEO Sasan Goodarzi framed it as reducing complexity to sharpen focus on big bets, AI first among them (Reuters). The developer-relevant detail is the contracts sitting next to that decision: Intuit has signed multi-year deals with both Anthropic and OpenAI to put their models inside TurboTax, QuickBooks, Credit Karma, and Mailchimp, and to surface its own tax and accounting tools inside Claude and ChatGPT. US staff get 16 weeks of base pay plus two weeks per year of tenure, with a last day of July 31. The memo went out hours before Q3 earnings, and the stock fell about 5% that morning.

Meta began cutting around 8,000 roles, about 10% of staff, the same day. One tracker counts more than 140 tech firms shedding over 111,000 jobs in 2026, a growing share of them citing AI efficiency, even as WEF executives warn that AI is increasingly the stated reason for cuts that were already planned. The mood outside the industry showed up at a podium. Former Google CEO Eric Schmidt was booed repeatedly at the University of Arizona commencement on May 15, and the jeering got louder when he turned to AI and jobs. "I know what many of you are feeling about that. I can hear you. There is a fear," he said (NBC News). Two things were tangled at that event: anxiety about the job market these graduates are entering, and a separate organized protest over a lawsuit from his former partner that a judge sent to arbitration in March. Either way, it is a sharp read on how the incoming cohort feels about the Silicon Valley AI pitch.

The talent moved even as the headcount shrank. Andrej Karpathy, an OpenAI founding member and former Tesla AI director, said he has joined Anthropic. "I think the next few years at the frontier of LLMs will be especially formative," he wrote, adding that he wanted to get back to R&D after running his AI-education startup Eureka Labs (Karpathy). It lands loudly in the OpenAI-versus-Anthropic frame, in the same week as the IPO news below.