Google Launches Gemini 4 Neon and Stitch Loop as Cloudflare Open-Sources Clairic and Vercel Debuts Multi-Agent Board
Production AI infrastructure pivoted toward lower-cost steerability, specialized extraction, and swarm coordination on November 18 as Google rolled out $3/M Gemini 4 Neon alongside 24/7 Stitch Loop agents, Cloudflare’s Workers AI Superintelligence neolab released open-weight Clairic and Clairic-flash with zero output fees, and Vercel shipped Board for multi-agent software factories.

At a Glance
- What changed: Google DeepMind launched 1M-context Gemini 4 Neon across Google AI Studio, Antigravity, the Gemini app, Vercel AI Gateway, and GitHub Copilot while Google Labs unveiled 24/7 evergreen product agent Stitch Loop (plus DeepMind’s Genie 4 in Project Genie); Cloudflare’s new Workers AI Superintelligence team open-sourced redesigned 27B Clairic and 9B Clairic-flash extraction models on Hugging Face and Workers AI; and Vercel launched Vercel Board for harness-agnostic multi-agent coordination.
- Who can use it: Gemini 4 Neon is live in the Gemini API, Gemini for business, Google Antigravity, the Gemini app (Pro/Ultra), Vercel AI Gateway (google/gemini-4-neon), and GitHub Copilot (Pro+, Max, Business, Enterprise); Clairic and Clairic-flash are live on Workers AI and downloadable as open weights on Hugging Face; Stitch Loop opened its global waitlist (priority for Google AI Ultra); and Vercel Board (@vercel/board) is open in experimental preview.
- What it costs: Gemini 4 Neon lists at $3.00/M input tokens ($0.15/M cached input—a 95% discount) and $15.00/M output tokens, cutting cost per task by 32% vs. Gemini 4 Carbon at default high thinking; Cloudflare charges $0.00 for output tokens under System One pricing—billing only $0.10/M input tokens for 9B Clairic-flash and $0.24/M input for 27B Clairic (discounted to $0.10/M through December 1, 2026).
- Why should I care: Instead of routing every workflow through $8–$15/M frontier giants or fragile Markdown todo files, engineering teams can now pair a #1-ranked coding orchestrator ($3/M Neon) with sub-second $0.10/M open-weight PDF/image extraction (Clairic), continuous post-merge PR loops (Stitch Loop), and structured task-lease queues (Vercel Board).
While frontier AI vendors spent early November competing over raw context sizes, November 18, 2026 marked a decisive pivot toward production unit economics, specialized open-weight inference, and autonomous post-merge coordination. Google DeepMind rolled out Gemini 4 Neon at $3.00 per million input tokens alongside Google Labs’ 24/7 repository optimization agent Stitch Loop; Cloudflare formally debuted its Workers AI Superintelligence “neolab” with open-weight Clairic (27B) and Clairic-flash (9B) document-extraction models that charge zero output-token fees; and Vercel introduced Vercel Board to coordinate multi-agent software factories across heterogeneous coding harnesses.
Google Gemini 4 Neon: $3/M Coding Orchestration, Cyber Hardening & Day-One Ecosystem Rollout
Positioned inside the Gemini 4 lineup between heavy long-horizon Gemini 4 Carbon and ultra-fast Gemini 4 Hydrogen, Gemini 4 Neon targets high-frequency engineering, scientific analysis, and multimodal enterprise work across a 1-million-token context window (accepting text, code, spreadsheets, PDFs, images, audio, and video) with four reasoning tiers: low, medium, high, and xhigh (Model Pricing & Inference):
- Benchmark Lead at 32%–80% Lower Task Cost: Priced at $3.00/M input tokens ($0.15/M cached input—a 95% discount) and $15.00/M output tokens, Neon claims #1 overall on the Artificial Analysis Coding Agent Index (58.6), #3 overall on the Artificial Analysis Intelligence Index (61 at
xhigh), and #2 on Agent Arena Index (1336 across 4,200+ sessions) as well as #2 on mid-task steerability. At its defaulthighthinking tier, Neon delivers 32% lower cost per task than Gemini 4 Carbon and 50%–80% lower task cost than rival frontier models—jumping from 12.8% on mini-SWE to a state-of-the-art 39.0% on SWE-sweep inside Google Antigravity when orchestrating parallel Hydrogen subagents. - Frontier Cyber Defense & Anti-Fabrication Controls: Neon posts state-of-the-art scores on SEC-bench Pro (90.5%), WebExploitBench v1.2 (51.26%), and hard-tier AutoBaxBench (76%) at roughly half the per-challenge cost of larger frontier models, while matching a 0.0% indirect prompt-injection attack success rate on Gray Swan’s IPI benchmark at
xhigh, recording a 99.2% agent honesty pass rate, and trimming task refusals to just 1.98% on Terminal-Bench 4.0. - Broad Developer & Enterprise Availability: Beyond Google AI Studio, Android Studio, Google Antigravity, Gemini for business, and the Gemini app (for Google AI Pro and Ultra subscribers, who also gained real-time Genie 4 interactive world generation in Project Genie on November 18), Neon shipped immediately on Vercel AI Gateway (
google/gemini-4-neon) and GA in GitHub Copilot (Pro+, Max, Business, and Enterprise across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, CLI, mobile, andgithub.com).
Architecture, Pricing & Benchmarks Across November 18’s AI Stack Releases
The comparison matrix below details token pricing, latency and benchmark metrics, licensing, and deployment surfaces across Google, Cloudflare, and Vercel’s November 18 releases:
| Release / Model | Core Architecture & Performance | Pricing & Licensing Terms | Availability & Integration Surfaces |
|---|---|---|---|
Google Gemini 4 Neon (google/gemini-4-neon) |
1M multimodal context; 4 thinking levels (low–xhigh); #1 Coding Agent Index (58.6); 39.0% SWE-sweep in Antigravity; 90.5% SEC-bench Pro |
$3.00/M input, $0.15/M cached read (-95%), $15.00/M output; 32% lower cost/task vs. Gemini 4 Carbon | Gemini API / AI Studio, Antigravity, Gemini app (Pro/Ultra), Vercel AI Gateway & GitHub Copilot (Pro+, Max, Business, Enterprise) |
| Cloudflare Clairic (27B) & Clairic-flash (9B) | Parallel slot autoregressive generation + arithmetic/forecast heads; up to 4 images or PDF pages/call; 155–358 ms p50 text & 305–453 ms p50 image latency | Open-weight on Hugging Face; System One billing charges $0.00 output tokens ($0.10/M input on 9B; $0.24/M on 27B—$0.10/M promo until Dec 1) | Workers AI (@cf/cloudflare/clairic & @cf/cloudflare/clairic-flash) and self-hosted open weights via Hugging Face |
| Google Labs Stitch Loop | Evergreen 4-step product agent (Suggestions, Solutions, sandboxed Build PRs, Reflect); pairs Gemini 4 Carbon + DeepMind CodeMender for security | Experimental Google Labs release; priority onboarding for paid Google AI Ultra subscribers | Connects GitHub + 180+ analytics/log integrations, Slack (@stitch), Stitch CLI & MCP servers (stitch.google.com/loop/waitlist) |
Vercel Board (@vercel/board) |
Harness-agnostic channels, threads, task leases (claim/renew/complete), JSON schema validation & serverless event watchers |
Experimental primitive on Vercel; pairs with free Parallel web search on AI Gateway through Jan 31, 2027 | CLI & JS SDK across Claude Code, Codex, OpenCode, eve, fx, Mastra & LangChain; two-way Slack & GitHub sync + Sandbox policies |
Cloudflare’s Workers AI Superintelligence Neolab Open-Sources Clairic & Clairic-flash
Declaring that Cloudflare “has become a neolab,” the company formally introduced its Workers AI Superintelligence frontier research group (the team behind Clef and Clairic) and released redesigned 27B Clairic and 9B Clairic-flash models under open weights on Hugging Face alongside managed endpoints on Workers AI (Open-Source Releases). Rather than burning general-purpose reasoning tokens to parse invoices, commercial leases, or support tickets, Cloudflare replaced its single-pass preview design with parallel slot autoregressive generation—allowing every requested schema field to advance simultaneously in the serving batch while retaining typed confidence scores, character spans, probabilities, and specialized heads for arithmetic, calculated dates, and forecasting.
The redesigned family accepts up to four images or multi-page PDF pages per request, cutting errors on Cloudflare’s internal receipt benchmark from 207 down to 0 (576/576) and hitting 1,263/1,263 on tracked customer assertions—outpacing directional internal runs of Liquid AI’s D1.2 + OCR and Scaledown Extract while returning complete HTTP responses in 155–393 ms p50 on 9B Clairic-flash and 358–691 ms p50 on 27B Clairic. Crucially, Cloudflare preserved its System One input-only pricing: output token generation is completely free, with @cf/cloudflare/clairic-flash billed at $0.10/M input tokens and 27B @cf/cloudflare/clairic discounted from $0.24/M to $0.10/M input tokens through December 1, 2026.
Post-Merge Autonomy & Swarm Coordination: Google Stitch Loop Meets Vercel Board
Completing the stack shift (Cloud & Developer Infrastructure), both Google Labs and Vercel targeted what happens after a single coding agent finishes a prompt. Google Labs’ Stitch Loop connects repositories on GitHub with more than 180 live telemetry, log, and user-feedback integrations (plus custom MCP servers and @stitch in Slack) to run continuous 24/7 improvement cycles across three default loops—cybersecurity (combining Gemini 4 Carbon with DeepMind’s CodeMender), latency, and accessibility—alongside custom conversion goals. Every candidate fix is built inside an isolated cloud sandbox and submitted as a human-reviewed GitHub pull request with visual screenshots or videos before Stitch Loop monitors post-merge telemetry to verify whether the metric improved.
At the orchestration layer, Vercel Board (npx skills add vercel/board) gives multi-agent fleets across Claude Code, Codex, OpenCode, eve, fx, Mastra, and LangChain a shared message store with task leases, task dependencies, versioned JSON schemas (security-finding@1), granular user/project grants, credential-free @vercel/sandbox network policies, and serverless watchers that wake deployments only when a Slack thread or GitHub PR updates. Already powering Vercel’s internal security swarm and v0’s software factory—where agents now merge over 50% of pull requests—Board illustrates alongside Vercel’s newly announced equity-free Vercel X San Francisco accelerator ($9M+ in partner compute/API credits) how fast agentic engineering is maturing from solo IDE chat into governed, multi-model production pipelines.
Why It Matters
Just twenty-four hours after Microsoft Ignite and OpenAI Reset frontier 1M-context coding benchmarks, Google, Cloudflare, and Vercel answered the more urgent operational bottleneck facing production AI teams: runaway per-task token bills, uncoordinated multi-agent swarms, and post-deployment code decay. By positioning Gemini 4 Neon as a high-steerability $3/M orchestrator that directs fast Hydrogen subagents at 32% lower task cost than Carbon—while Cloudflare decouples structured PDF/receipt extraction into sub-second open-weight System One models with zero output-token fees and Vercel replaces ad-hoc agent logs with lease-backed message boards—enterprise architectures are rapidly splitting into specialized, auditable, cost-bounded tiers.
What Is Confirmed
Confirmed across six primary disclosures published November 18, 2026 by Google, Cloudflare, Vercel, and GitHub: (1) Gemini 4 Neon offers a 1M-token multimodal window, four thinking tiers (low, medium, high, xhigh), $3.00/M input ($0.15/M cached) and $15.00/M output pricing, #1 ranking on the Artificial Analysis Coding Agent Index (58.6), #3 overall on the Artificial Analysis Intelligence Index (61 at xhigh), 39.0% SOTA on SWE-sweep in Google Antigravity, 90.5% SOTA on SEC-bench Pro, and 0.0% Gray Swan indirect prompt-injection success at xhigh, rolling out across AI Studio, Antigravity, Gemini app, Vercel AI Gateway (google/gemini-4-neon), and GitHub Copilot (Pro+, Max, Business, Enterprise); (2) Google Labs’ Stitch Loop connects GitHub and 180+ product/MCP integrations to run 24/7 cybersecurity (Gemini 4 Carbon + CodeMender), latency, and accessibility PR loops alongside DeepMind’s Genie 4 rollout for Google AI Ultra; (3) Cloudflare’s Workers AI Superintelligence team open-sourced 27B Clairic (@cf/cloudflare/clairic, $0.24/M input, $0.10/M promo through Dec 1) and 9B Clairic-flash (@cf/cloudflare/clairic-flash, $0.10/M input) on Hugging Face with $0 output-token charges, parallel slot autoregressive generation, and up to 4 images/PDF pages per call; and (4) Vercel launched experimental Vercel Board (@vercel/board) with two-way Slack/GitHub sync, task leases, JSON schema validation, and zero-credential Sandbox policies.
What Remains Unclear
Google Labs has not yet announced general-availability seat or compute pricing for Stitch Loop once teams graduate from the global waitlist, Cloudflare noted that public standardized multi-page extraction benchmarks do not yet exist (making its head-to-head comparisons against Liquid AI D1.2 + OCR and Scaledown Extract directional internal evaluations), and Vercel Board remains an experimental primitive without finalized standalone consumption tiers.
Sources & Underlying Evidence (6)
We link directly to the primary documents supporting our reporting and distinguish company claims from independently checked facts.
- Introducing Gemini 4 NeonPublisher: Google Official Blog (Google DeepMind) • Type: Official AnnouncementSource date: Nov 18, 2026
- An updated Clairic and Clairic-flash are now open-weight, from the new Workers AI Superintelligence teamPublisher: Cloudflare Blog • Type: Official AnnouncementSource date: Nov 18, 2026
- Introducing Stitch Loop: A new way to build and evolve productsPublisher: Google Official Blog (Google Labs) • Type: Official AnnouncementSource date: Nov 18, 2026
- Introducing Vercel BoardPublisher: Vercel Blog • Type: Official AnnouncementSource date: Nov 18, 2026
- Gemini 4 Neon now available on AI GatewayPublisher: Vercel Changelog • Type: Product ChangelogSource date: Nov 18, 2026
- Gemini 4 Neon in GitHub CopilotPublisher: GitHub Changelog • Type: Product ChangelogSource date: Nov 18, 2026


