Google's Gemini 4 Carbon tops benchmarks as open models undercut frontier prices

Google's Gemini 4 Carbon, released last week, now ranks first on Artificial Analysis' Intelligence Index V5 with a score of 59. It leads Claude Opus 5.5 by one point, according to Tom's Hardware's report on 3 November. The same suite puts OpenAI's GPT-6 Astra at 57 and Meta's Muse Wonder Full at 50.

Raw intelligence is not the whole pitch. Gemini 4 Carbon (high) also carries a cost per task of $7.76 on the Intelligence Index. That is cheaper than Gemini 3.8 Flash (high), and several dollars below Claude Opus 5.5 and Sonnet 5.5. Claude Fable 5.1 (max with fallback) comes in at $12.99 per task, according to the same benchmarks.

Google is not selling it to everyone yet.

Tom's Hardware reports that access opens first to cyber defenders through the company's Fairwind program. Wider availability is promised sometime after. That sequencing matters. The pitch here is reliability under adversarial conditions, not consumer novelty.

Hallucination scores remain low across the board

Gemini 4 Carbon also tops Artificial Analysis' AA-Omniscience v1.1 test, which tracks hallucination and factual recall. The caveat is stark. Even the best model scored 54 on a -100 to +100 scale. That suggests frontier systems in 2026 are still far from consistently reliable when recalling information.

Keep that number in mind when vendors talk about agentic deployment. A model that invents facts one time in twenty is a different product from one that invents them one time in a thousand. The benchmark says nobody is near the second figure. The rankings move fast too. Tom's Hardware frames Gemini 4 Carbon as a comeback for Google after an irregular release year.

Meanwhile, the cheapest capable option is not a frontier lab's flagship at all.

DeepSeek V4.1 Flash 1017 costs roughly 13 times less per task than Gemini 4 Carbon. It scores just 32 on the same index, per Tom's Hardware.

Gartner: open weights do most of the work

That price gap is the backdrop to a Gartner document published last week and reported by The Register on 3 November. Analysts Mark Petty and Raymond Paquet argue that nine in ten agent workflow steps are routine. Open-weight models handle them at parity for a quarter of the price of the cheapest closed model, or as little as a twentieth against a frontier model run on every step.

"Current open models can be more verbose and retry more often, so the saving per task against a frontier model is closer to five or six times," the analysts wrote, according to The Register.

Gartner's strategic planning assumption is that by 2030, 90% of enterprise operational agent workflow steps will execute on open-weight models. The firm also argues the saving is secondary. "Years of workflow history are a large part of what an acquirer pays for in a SaaS company and every routine step sent to a frontier API moves a slice of it outside your perimeter," the analysts wrote. "Open-weights preserve your control. The cost saving is a bonus."

The Register notes that Gartner expects value to migrate to harnesses, the software that routes work between models. Frontier models, in that framing, become research and development spending. Open weights become the operating cost.

Inference volume climbs in China

Demand grows regardless of which model answers. China's average daily token calls exceeded 175 trillion, more than 1,000 times the level in early 2024. That figure comes from a report released on Saturday by the China Centre for Information Industry Development, which sits under the Ministry of Industry and Information Technology. The South China Morning Post reported on 2 November that inference data generated in China has for the first time surpassed the data used to train the systems.

Token calls, as the SCMP defines them, measure usage through public cloud providers. They are a proxy for consumption, not a quality measure. The report does not break out how much of that traffic is video, text or agents.

Video generation is one workload pushing that volume. Public data on model releases in that category is thin beyond vendor announcements in the recent headlines.

Specialists make their case

Il Sole 24 Ore reported on 1 November that Translated, a Rome-based language services company, released Lara 3 in October, a model built for professional translation. In tests the company organised, professional translators blind-reviewed 16,800 translations from English into 21 languages. 82% of Lara 3 output was judged usable without edits, against 69% for GPT-5.6 Sol, the best generalist in the comparison. The paper notes the tests were company-run and limited to those content types.

That is the specialist argument in miniature: fewer corrections, less editing time. It is also a claim from a vendor about its own benchmark. That is why the sample size and the scope matter more than the headline percentage.

Hardware and training work continues underneath.

RTL-Interact, a paper submitted to arXiv on 30 October, describes a multi-task post-training framework for chip design. It reaches 48.7% weighted Pass@1 on the CVDP benchmark with a 32B model, ahead of its teacher DeepSeek-V3.2. It also cuts average tool calls by 63.8% through a method the authors call Pattern-Aware Policy Optimization.

Power and patching

Data centres running all this inference face their own constraint. Ember's October 2026 analysis, reported by pv magazine on 3 November, finds that data centres already account for roughly 19% of peak electricity demand in Ireland. It projects that share could reach 24% by 2030, with data centre demand at 1.6 GW against a system peak of 6.5 GW. Load shifting could cut Irish peak demand by about 100 MW, or 1.5%, by 2030. Behind-the-meter generation and storage could deliver nearly five times more relief, with potential reductions of up to 2.7 GW in Texas.

Security teams have a more immediate problem.

Red Hat shipped kernels on 1 November fixing CVE-2026-98432, a use-after-free in the Linux kernel's IPv4 TCP code. The kernel CVE team rates it 9.8 out of 10, and a remote, unauthenticated attacker can reach it, The Hacker News reported on 3 November. Red Hat initially rated it 5.9 and has raised the score to 9.8. Canonical said on 2 November that Ubuntu fixes for 22.04, 24.04 and 26.04 LTS sit in its opt-in -proposed channel with no release date for the regular update.

Against that backdrop, the Gemini 4 Carbon numbers read less like a finish line and more like a snapshot. A one-point lead on an intelligence index, a $7.76 cost per task and a hallucination score of 54 describe a market where the frontier is expensive. The middle is cheap, and most production work, if Gartner is right, will not touch the frontier at all.