Gemini 4 Carbon tops Artificial Analysis' V5 index, but the video-model race is elsewhere

The newest number is 59. That is Gemini 4 Carbon's score on Artificial Analysis' Intelligence Index V5, according to Tom's Hardware's report on 3 November. It is one point clear of Claude Opus 5.5 and two ahead of OpenAI's GPT-6 Astra. Google's model also topped the firm's AA-Omniscience v1.1 hallucination test.

Those are language-model scores. The topic here is AI video generation, and on that front the picture is thinner and stranger than the headline suggests. The freshest dated items are a market report, a product launch and a string of vendor releases. The models doing the generating mostly arrive through secondary mentions. The most recent concrete date is 3 November, when Precedence Research pushed out an AI anime generator market study and Kling3.io announced Kling 4.0, per the context headlines. Neither is a peer-reviewed benchmark.

What the benchmark actually measured

Tom's Hardware's piece is worth reading closely because it separates the intelligence score from the cost figures. Gemini 4 Carbon (high) costs $7.76 per task on the Intelligence Index test. That is slightly cheaper than Gemini 3.8 Flash (high), and well under Claude Fable 5.1 at $12.99 per task. It is still more expensive than OpenAI's GPT 5 Astra and roughly six times the per-task cost of GPT 6.1 Sol (max). Against the open-weight DeepSeek V4.1 Flash, the gap is about 13 times, though that model scores 32.

Google is giving cyber defenders first access through its Fairwind program, with wider availability later. The hallucination result deserves a caveat the article itself supplies: even the best model scored 54 on a scale running from -100 to +100. A leader on recall is not a reliable narrator.

"Even the best scored only 54 on the -100 to +100 scale, suggesting even the best frontier models in 2026 are a far cry from being consistently reliable at accurately recalling information."

That framing matters for anyone buying video tools. Buyers judge generated clips by eye, not by a leaderboard. The video-adjacent items lean on vendor claims rather than independent tests. The Register reported on 3 November that Gartner analysts Mark Petty and Raymond Paquet expect open-weight models to handle 90% of enterprise agentic workflow steps by 2030. That would cost a quarter of the price of the cheapest closed model, and as little as a twentieth of a frontier model run on every step. The analysts admit current open models "can be more verbose and retry more often," so the real saving is closer to five or six times.

That price pressure is the backdrop for every generation model launch. If routine work migrates to open weights, the closed labs sell the expensive tail: research, hard reasoning, and the branded creative tools that sit on top.

The specialist argument, and who is making it

Il Sole 24 Ore argued on 1 November that specialised systems still have room beside generalists. Its example is Translated, the Rome-based language services firm, which released Lara 3 in October for professional translation. Translated's own tests had human reviewers blind-check 16,800 translations from English into 21 languages. Reviewers judged 82% of Lara 3 output usable without edits, against 69% for GPT-5.6 Sol. The paper notes the company organised the tests and limited them to those contents. That is the right caveat to attach to any vendor benchmark.

The same logic applies to video. A model that nails typography and layout control is worth more to a brand team than a generalist that scores higher on an index. Ideogram's 4.5 model is one example, focused on type and layout rather than broad capability.

Underneath all of it sits the electricity bill. Ember's October 2026 analysis, covered by pv magazine on 3 November, found data centers already account for about 19% of peak demand in Ireland, and projected 24% by 2030. Load shifting could cut Irish peak demand by roughly 100 MW, or 1.5%. Behind-the-meter generation and storage could deliver nearly five times more relief, with potential reductions of up to 2.7 GW in Texas. A large data center needs 200 MW to 500 MW and can be built in 18 to 24 months; grid expansions take four to 10 years.

Physical AI is moving too. Tutor Intelligence said on 2 November it raised $75 million in Series B funding, led by Andreessen Horowitz, for robots trained with real-world data and remote human tutors. CEO Josh Gruenstein said the company "can hardly build robots fast enough."

None of this settles which video model wins. It does explain why the release cadence keeps accelerating while independent evaluation lags.