Gemini 4 Carbon Beats Claude Opus 5.5 by One Point and Leads New Hallucination Index
Google's new model scores 59 points on the Intelligence Index V5 and costs $7.76 per task, making it the cheapest option among the frontier models tested
Pontos-Chave
- Gemini 4 Carbon (high) scored 59 points on the Intelligence Index V5, one point ahead of Claude Opus 5.5 (max with fallback) and two ahead of GPT-6 Astra (max)
- The new model's cost per task is $7.76, lower than Gemini 3.8 Flash (high) and about half that of Claude Fable 5.1, which costs $12.99 per task
- In the AA-Omniscience v1.1 test, which measures hallucinations and information recall, the best model reached only 54 points on the -100 to +100 scale
Google has unveiled Gemini 4 Carbon, its newest frontier model, which takes the lead on the Artificial Analysis Intelligence Index V5 with 59 points. The performance puts the system one point ahead of Anthropic's Claude Opus 5.5 and two points above OpenAI's GPT-6 Astra, in a tight race for the top of the cognitive capability ranking.
The launch represents a strategic recovery for Google, which had come off a year marked by uneven releases in the foundation model segment. The company is seeking to consolidate its position in a market where leadership changes hands frequently and where cost efficiency has become as decisive as raw intelligence.
Economic efficiency as a differentiator
Beyond the top score, Gemini 4 Carbon stands out for its operational cost. The model has a cost per task of $7.76 on the Intelligence Index, slightly lower than Gemini 3.8 Flash and significantly cheaper than direct competitors. Claude Opus 5.5 and Sonnet 5.5, both in maximum configuration with fallback, cost several dollars more per task, while Claude Fable 5.1 reaches $12.99 — nearly double that of Google's new model.
This combination of high performance and reduced cost positions Gemini 4 Carbon on the so-called Pareto frontier, a concept developers pursue when seeking the best possible balance between intelligence and price. For companies operating models at scale, the difference of a few dollars per task can represent substantial savings in large-scale deployments.
Hallucinations under scrutiny
The new model also takes the lead in the AA-Omniscience v1.1 test, which evaluates systems' tendency to invent information and their ability to retrieve knowledge accurately. Gemini 4 Carbon has the lowest hallucination rate among the frontier models evaluated, a differentiator that could become decisive as governments and institutions seek to maximize safety and reliability in the use of artificial intelligence.
However, the numbers reveal a still concerning reality for the industry. Even the top performer scored only 54 on the scale ranging from -100 to +100, indicating that the most advanced systems of 2026 remain far from consistent reliability in information retrieval. The metric shows that, despite advances, the tendency to hallucinate remains a structural challenge for the entire industry.
Competitive context
Gemini 4 Carbon's lead, though narrow, signals a repositioning of Google in a scenario where Anthropic and OpenAI had been alternating in the lead. The single-point difference on the Intelligence Index demonstrates that frontier models have reached a level of technical parity, in which factors such as cost, latency and reliability now define companies' choices.
The AA-Omniscience v1.1 test, by measuring both factual knowledge and propensity for errors, introduces a qualitative dimension to the benchmark race. For regulated sectors such as healthcare, finance and legal, a model's ability to recognize the limits of its own knowledge may be more valuable than additional points on general intelligence indices.
The -100 to +100 scale used in the test reflects a spectrum ranging from completely hallucinated answers to accurate and verifiable retrievals. The fact that no model has surpassed 54 points suggests that the industry still needs to resolve fundamental issues of calibration and epistemic honesty before promoting critical large-scale deployments.
Practical impact
For engineering teams, the combination of 59 points on the Intelligence Index V5 with a cost of $7.76 per task makes Gemini 4 Carbon a natural candidate to replace more expensive models in production pipelines, especially where task volume amplifies the unit cost difference. The lead in AA-Omniscience v1.1, even with a score of 54, reinforces the need to maintain verification and fallback layers in the architecture.
Operationally, the one-point margin over Claude Opus 5.5 and two over GPT-6 Astra confirms technical parity at the frontier, shifting the decision to cost, latency and reliability. The maximum score of 54 on the -100 to +100 scale of the hallucination test imposes caution on regulated sectors, where epistemic calibration continues to require active governance and human validation.