Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5

artificialanalysis.ai · Oct 29, 2026

Google’s Gemini 4 Carbon ranked first on Artificial Analysis’s Intelligence Index v5 with a score of 59, led by results in complex reasoning, knowledge, and scientific tasks. The model placed first in four of the index’s nine evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA. It supports text, image, video, and speech inputs, produces text output, and has a one-million-token context window.

Gemini 4 Carbon scored 59 on Artificial Analysis’s Coding Agent Index when used with Antigravity CLI, three points below Claude Opus 5.5’s score of 62. It ranked second on Terminal-Bench 4.0 at 64% and FrontierCode v1.1 at 51%. On AA-Briefcase, its 1,609 Elo rating exceeded Gemini 4 Argon’s 1,483 but remained below Grok 4.7, Muse Wonder, and Kimi K3.1 Pro.

At standard pricing of $4 per million input tokens and $20 per million output tokens, Artificial Analysis estimates Carbon costs $7.76 per Intelligence Index task, compared with $10.04 for Claude Opus 5.5 and $6.25 for GPT-6 Astra. Google is rolling out Carbon to selected users rather than making it publicly available, and offers a 50% launch discount with no confirmed end date.

Read original ↗

What people are saying

The thread treated Carbon’s number-one ranking less as proof of a breakthrough than as a test of whether Google and Artificial Analysis deserve trust.

What readers liked

Gemini remains unusually good at turning hard science and higher-level mathematics into clear explanations, giving learners a practical advantage…
Google’s search and Maps integration makes Gemini genuinely useful for research-heavy planning, where answers can lead directly to sources, places…

Concerns raised

Carbon’s top ranking feels hollow while public access lags behind the announcement, turning a claimed frontier win into marketing that users cannot…
Argon has produced elaborate, wholly invented answers rather than small factual slips, a failure mode that makes confident output dangerous for…

View discussion ↗

Read full discussion analysis

Source: Hacker News · Based on 15 sampled comments · Analyzed Oct 29, 2026

Discussion overview

The thread treated Carbon’s number-one ranking less as proof of a breakthrough than as a test of whether Google and Artificial Analysis deserve trust. Several commenters argued that Gemini has long been stronger than its reputation in practical niches: scientific and mathematical explanations, Google Search and Maps-assisted research, and difficult handwriting transcription. Some developers also reported Argon solving coding problems that competing models missed, while earlier Gemini Flash releases earned praise for a rare combination of speed, accuracy, and low API cost.

But the dominant mood was impatient and skeptical. Carbon is being celebrated before most people can use it, and recent Google rollouts have trained users to discount launch claims until they can run their own tasks. Firsthand Argon reports described severe message caps, sprawling fabrications, uneven instruction following, and prices that no longer justify choosing Gemini over rivals. The ranking itself drew the sharpest suspicion: v5 changed its benchmark mix just as a major closed model arrived on top, reduced visibility for some open models, and relies heavily on private tests while offering consulting to labs. Defenders replied that public evaluations rapidly saturate and invite benchmark gaming, so refreshed and private tests may be necessary. Even they did not erase the demand for reproducible scores and real-world access.

Points of disagreement

The central fight is whether Artificial Analysis v5 exposes genuine frontier capability or launders a preferred closed model into first place. Skeptics see conveniently timed methodology changes, opaque private evaluations, and disappearing open-model results as evidence that the leaderboard cannot be independently trusted. Defenders argue that old public benchmarks are routinely trained against or optimized for, making unsaturated private tests and regular index revisions essential rather than suspicious.

What readers liked

Gemini remains unusually good at turning hard science and higher-level mathematics into clear explanations, giving learners a practical advantage over flashier general chatbots.
Google’s search and Maps integration makes Gemini genuinely useful for research-heavy planning, where answers can lead directly to sources, places, and next steps.
On difficult historical handwriting, Argon reportedly extracts faded nineteenth-century text at a level that can beat painstaking manual reading, making costly OCR still economically compelling.
Earlier Gemini Flash APIs won loyalty by delivering fast, accurate, inexpensive responses, showing Google can hit an excellent production tradeoff when the model is well tuned.

Concerns raised

Carbon’s top ranking feels hollow while public access lags behind the announcement, turning a claimed frontier win into marketing that users cannot independently test.
Argon has produced elaborate, wholly invented answers rather than small factual slips, a failure mode that makes confident output dangerous for transcription and research.
Recent Gemini models are criticized for weak instruction following, bloated reasoning, long-context degradation, and pricing that loses the old Flash value proposition.
Artificial Analysis v5 looks compromised to skeptics because private benchmarks dominate, methodology shifted near the launch, and lower-scoring open models became harder to compare.
Tight usage caps and multi-hour cooldowns make a premium personal subscription feel unusable for sustained AI work, regardless of benchmark standing.