GOOGLE / GOOGLE DEEPMIND · MODEL RELEASE TRACKER
Gemini 4 Carbon
A Gemini 4 family model (Gemini 4 Carbon). Reported to have an Artificial Analysis Intelligence Index score of 60 at the xhigh reasoning level, a token generation rate of about 57 tokens/second, and regular list pricing of $4 per million input tokens and $20 per million output tokens. It had a promotional price in effect through the end of November and an undiscounted per-task cost (xhigh) reported as $7.76.CURRENT SNAPSHOT4/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
RELEASE TIMELINE
Published, source-backed release events only.
Mistral launches Vibe Desktop app for macOS and Windows
Mistral released Vibe Desktop, a macOS and Windows app that lets its AI agent Vibe run multiple tasks in parallel, combine local Vibe CLI features with cloud sessions from Vibe Code Web, use Git worktrees to isolate parallel work, and access local files; a Linux version is planned. The company positions the release as catching up to rivals whose desktop agents (notably Anthropic’s Claude) already have deeper system-control capabilities, and Mistral’s Large 4 model still trails top competitors in head-to-head comparisons.
Google DeepMind launches Gemini 4 Neon — a 1M‑token multimodal frontier model
Google DeepMind announced Gemini 4 Neon, a "frontier-level" multimodal model with a 1,000,000‑token context window. Neon is available via the Gemini API, Google AI Studio, Google Antigravity, Gemini for business, and paid Gemini app subscriptions; Google published pricing ($3 per million input tokens, $0.15 per million cached input tokens, $15 per million output tokens) and highlighted high rankings on multiple internal and external benchmarks (e.g., #3 on the Artificial Analysis Intelligence Index at xhigh with score 61, #2 on the Vals Index, and top placements on several coding, finance, legal and agent benchmarks).
Google releases Gemini 4 Argon for robotics control, simulation, and policy improvement
Google released Gemini 4 Argon, a general-purpose physical-reasoning model aimed at robot control, Real2Sim scene construction, multimodal labeling of robot video, and iterative policy improvement. Google’s demos show Argon guiding a Boston Dynamics Spot and robot arms in MuJoCo twins, reconstructing scenes from teleoperation video, and synthesizing domain-randomized MuJoCo episodes (reported to raise success from 56% to 75% on one setup); Google reports Argon scores 87% on the ERQA benchmark, and makes it available to Google AI Ultra subscribers and via the paid Gemini API as gemini-4-argon, while publishing only limited per-demo failure data and relying on heavy prompt scaffolding in examples.
VERIFIABLE FACTS
Every value stays attached to a source and date.
Announced Oct 28, 2026.
Described as Google’s strongest model with state-of-the-art cybersecurity capabilities and long-horizon / sustained agentic workflows (able to run for days to pursue complex, multi-step engineering tasks); reported improvement across coding, agentic, and scientific tasks, with uplift in generating secure code and reduced misalignment and fabrication rates (per Google).
Ranks #1 on the updated Artificial Analysis Intelligence Index v5.0 with a score of 59.
Reported scores: SEC-bench Pro (finding vulnerabilities) 88.2%, ExploitGym (exploitation) 54.8%, CWE-bench (patching) 71.0%.
A promotional price for Carbon was reported as being in effect through the end of November (amount not specified in the excerpt).
Accessible via the Gemini API paid tier in Google AI Studio and Android Studio. Google AI Ultra subscribers can try Carbon in Google Antigravity, AI Studio, the Gemini app, and AI Mode in Search. Gemini for Business users can access it via the API or model dropdown, Google Antigravity, or CodeMender.
Described as combining frontier visual understanding with reasoning and exhaustive retrieval capabilities, indicating multimodal (text + vision) functionality.
In an independent evaluation of drone/surveillance targeting scenarios, Gemini 4 Carbon reportedly recognized lethal consequences and steered responses toward empty waypoints to avoid facilitating attacks.
Described as the strongest Gemini 4 model with state-of-the-art cybersecurity and long-horizon capabilities; designed for sustained, multi-day agentic workflows to tackle complex, multi-step engineering tasks (e.g., root-cause analysis and patching subtle bugs across large codebases and low-level systems).
Ranks #1 on Artificial Analysis Intelligence Index v5.0 with a score of 59.
Reported to resist prompt-injection attacks better than Argon; on Gray Swan's indirect prompt-injection benchmark Google reports Carbon reduced attack success from 0.7% (Argon) to 0.0% in the tested cases, though reporting notes this does not mean prompt injection is solved.
Reportedly rolled out more broadly in the week before the blog post published on 2026-11-17 (i.e., mid-November 2026).
Built for tasks that push a model's intelligence; reported by Google to enable thorough exploration, deep reasoning, and extended verification loops.
10.9% on SemiAnalysis's GDPx benchmark
Ranked 2nd (behind Claude Opus 5.5) on SemiAnalysis's GDPx leaderboard
Described by Google as its strongest model with frontier performance in cybersecurity and long‑horizon agentic workflows across complex knowledge work, software engineering, and scientific reasoning; built for tasks that push model intelligence such as complex multi‑step coding, debugging, and open‑ended research.
Google reports Carbon ranks #1 on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench and performs strongly on the Multimodal Knowledge Agent (MKA) benchmark, providing high quality at a lower cost than comparable models in those tests.
Carbon uses more tokens than Gemini 4 Argon to reason more deeply and run extended verification loops; as a result, total task cost may be higher in practice despite the same per‑token price.
Google is previewing a '/swarm' feature in Antigravity that turns a single prompt into a coordinated team of agents; Carbon is positioned to play deep‑reasoning roles (architecture, coordination, auditing) within such multi‑agent workflows.
Reported to rank #1 on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench; also reported to have topped the MathArena math benchmark in independent media reporting.
Independent reporting notes a common failure mode of exceeding time limits and a tendency to almost never give up on a task; an example cited a script running for more than eleven hours without checking progress.
The article describes Gemini 4 Carbon as built for tasks that push the limits of a model’s intelligence.
The article states Carbon has the ability to 'explore thoroughly, reason deeply, and run extended verification loops.'
The article reports builders used Carbon to simulate live physics, including an example that simulated 1,048,576 particles in real time with WebGPU.
The article describes a project where Carbon turned a single reference image into a 3D design (a 19th‑century optics lab with 92 instruments).
Media reporting described Gemini 4 Carbon as a leading model in recent comparisons, noting it outperformed Mistral's Large 4 in head-to-head comparisons.
60
About 57 tokens per second
$4 per million input tokens; $20 per million output tokens
$7.76 (undiscounted per-task figure referenced for xhigh)
55 points
About 36 minutes (for the benchmark referenced in the excerpt)
WHAT CHANGED