AgentdexModel Pulse #6 · Thursday, Oct 29, 2026
Today in AI: smarter routing, fewer tokens, and AI finding the bugs first
The seven AI stories from the last day or two that matter most for AI coders and indie game devs. Each one links to its primary source.
Sponsored Sponsor this issue: reach AI coders and solo founders in Model Pulse.
Oct 28, 2026
NEW: GitHub Copilot Fusion picks the models for each task, and lands near the top at about half the cost
At GitHub Universe on Oct 28, GitHub launched Copilot Fusion, a runtime router that decides which models handle each prompt and how. GitHub's docs describe three patterns: Single (one model does the task), Cascade (an efficient model drafts and a judge accepts it or escalates to a stronger model) and Critique (several models review a draft in parallel and the first model revises it). You choose an effort level: Standard, High or Max. GitHub says Fusion Max scored within two points of Claude Code with Opus 5.5 Max in the Artificial Analysis Coding Agent Index, at about half the cost and 40% less time per task. Artificial Analysis puts it at $5.59 per task, against $11.42 for Claude Code with Opus 5.5 (max). Fusion is generally available in VS Code, Copilot CLI, the GitHub Copilot app and JetBrains IDEs. You pay each model's standard rate with no extra Fusion fee, though High and Max can call several models per task, and paid plans get 30% off model costs through Nov 30. GitHub also announced Copilot projects for agent work across multiple repos, session continuity across the app, VS Code, the CLI and the web, and Copilot code review Max. One warning from the docs: if Fusion throws away a draft, file edits that draft already made are not undone, so review changes before you commit.
Oct 28, 2026
MODELS: Gemini 4 Carbon takes #1 on the Artificial Analysis index, but cyber defenders get it first
Google announced a preview of Gemini 4 Carbon on Oct 28, two weeks after Gemini 4 Argon. It is the new #1 on the Artificial Analysis Intelligence Index v5 with a score of 59. Artificial Analysis puts it second on its Coding Agent Index at 59 in Antigravity CLI, behind Claude Code with Opus 5.5 at 62, and at $7.76 per index task, about 23% less than Opus 5.5 (max). On AA-Omniscience its accuracy rises 10 points over Argon to 58%, but its hallucination rate rises from 4% to 11%. Google reports state-of-the-art security scores: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym and 71.0% on CWE-bench for patching, plus #4 on WebDev Arena. Access starts this week with a group of cyber defenders through Google's Fairwind Program. Developers, businesses and consumers come later, with no date given. The introductory price is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterward. You cannot build on it yet, but plan to test it once it reaches the API, and watch that hallucination rate.
Source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
Oct 27, 2026
SECURITY: Bugs found with Claude drive another Chrome 156 patch, so update your browser
Anthropic's Frontier Red Team says an unreleased internal research model found critical vulnerabilities in widely used browsers and operating systems and wrote five attack chains. One was a one-click exploit that ran code outside Chrome's sandbox on macOS. Anthropic says Google assigned 62 CVEs to its reports, all rated critical or high and all patched as of Chrome 156, including 13 sandbox escapes. On Oct 27 Google shipped Chrome 156.0.8078.73/.74 with 52 more security fixes. Anthropic researcher Xinyang Ge is credited with all four critical ones, and two are marked "assisted by Claude". The same notes also credit OpenAI Codex Security for bugs. Anthropic says it reported everything under coordinated disclosure, which gives vendors 90 days to patch, and it recommends scanning code at pull-request time. Put together with Gemini 4 Carbon's security focus, AI is now finding bugs faster than most teams ship patches. Update Chrome (Help → About Google Chrome), and run AI security review on your own code before attackers run it for you.
Source: https://www.anthropic.com/research/discovering-vulnerabilities-with-claude
Oct 28, 2026
EFFICIENCY: Particle-1 does the same coding work with about 45% fewer tokens
Mixpanel founder Suhail launched Particle on Oct 28 with its first model, Particle-1, post-trained from DeepSeek V4.1 Flash to finish tasks with fewer tokens. According to Particle's write-up, it uses about 45% fewer tokens on average. On four coding and data benchmarks (Terminal-Bench 2.1, ReactBench, ADE Bench and DeepSWE 1.1) it used 28.4% to 52.2% fewer output tokens and cost 34.4% to 51.6% less per task than the base model, with pass rates within the base model's confidence intervals. Particle says reasoning tokens account for about 90% of the savings on three of the benchmarks: the model deliberates less and stops re-checking once a task is done. It is priced the same per token as DeepSeek V4.1 Flash ($0.30 per million input tokens and $1.20 per million output tokens), and you can use it through the Particle API, OpenRouter and Vercel AI Gateway. These are Particle's own benchmarks. If you run agents on open-weight models, test it in your own harness before you rely on those numbers.
Oct 28, 2026
TOOLS: Vercel CI beta goes after the pile of agent pull requests
Vercel introduced Vercel CI on Oct 28, in beta on all plans. You define tasks in Turborepo and Vercel runs them with no further configuration, skipping tasks your change does not affect. Vercel says CI ran 2 to 6 times faster in its own repositories, and that the beta has already run nearly 8 million tasks. Agents can drive it through the Vercel MCP server and CLI. CEO Guillermo Rauch pitches it as a fix for "the AI coding bottleneck": he argues CI that is "slow, flaky, goes down" holds back agents, and that agents can now inspect, abort, trigger and replay CI runs from the CLI. If your agents open more PRs than your CI can keep up with, try it on one repo.
Oct 27-28, 2026
GAME DEV: Valve won't ban AI tools, and its advice against AI clones is to keep building
At Steam Indie Connect in Busan on Oct 27, Valve's Tom Giardino and Marc Wilkinson answered developers worried about AI clones and low-effort games during this month's Steam Next Fest. PC Gamer reports that one developer described an extortion threat: "If you don't pay, we'll clone your game with AI and release it first." Giardino said Valve will not restrict AI use: "We don't want to be gatekeepers telling developers which tools they should or shouldn't use." His advice was to build a community and to keep "adding features, personality, and creativity" rather than relying on "a single mechanic that someone can easily copy." Wilkinson said Next Fest shows demos almost at random for the first two days so every game gets a fair shot: "We recognize that the trade-off is some low-quality games get exposed." Output Lag reports that Giardino called extortion a matter for investigators and asked developers to report cases to Valve. The quotes come from a translated transcript by Korean outlet Inven Global. For indie game devs: report extortion attempts, build your community early, and be careful about who gets your pre-release builds.
Oct 27-28, 2026
DEALS: General Compute raises $100M for fast inference, and Cognition reportedly weighs a $70B-$80B round
General Compute raised a $100 million Series A on Oct 27, co-led by Blackbird and Abstract, with Fuse Venture Partners and Village Global also in. It runs an inference cloud on chips from Cerebras, SambaNova and Positron AI, selling dedicated bare metal or a guaranteed tokens-per-second SLA. Its first major deployments go live in Q1 2027, with more than 50 MW in the pipeline. TechCrunch reports an $800 million post-money valuation, up from $60 million after its May seed, and says a customer deal worth $100 million in annual recurring revenue cleared the way for the round. In AI coding, Business Insider reported on Oct 28 that Cognition, the maker of Devin, has received offers valuing it at $70 billion to $80 billion and is seriously considering a round before year-end. Unconfirmed: the report cites unnamed sources, the talks are early and Cognition has not announced anything. That would come weeks after it raised at $48 billion in September. Money keeps flowing to faster tokens and to agents that ship code.
Source: https://www.generalcompute.com/blog/general-compute-100m-series-a
Ship faster this week
Grab the free Agentdex starter kit with prompts and agent configs for Cursor, Claude Code and Codex.
Get the free kitWe store your email with Resend and use it only for Model Pulse. See privacy. Missed the last issue? Read Model Pulse #5.