The AI Rankings Just Got Reshuffled

Artificial Analysis’s Intelligence Index V5 reshuffled the frontier rankings after changing its benchmark mix. Gemini 4 Carbon scored 59, ahead of Opus 5.5 at 58, while GPT-6 Astra scored 57 and rose from sixth to second. GPT5-1 Sol moved from ninth to third, Sonnet 5.5 fell from second to fourth, and GPT6-1 Terra rose from tenth to fifth. The new results suggest that rankings depend heavily on which capabilities an index measures and how it weights them.

Carbon scored nine points above Gemini 4 Argonne and ranked first on six of the eight benchmarks cited. It improved by eight points on Terminal Bench 4.0, 12 on Exploit Gem, and seven on OS World. On the AA Coding Index, it tied Sonnet 5.5 for second, three points behind Opus 5.5. Artificial Analysis put Carbon’s cost at $7.76 per task, 23% below Opus 5.5 and 20% above Astra. These results show a strong broad improvement, but they do not make Carbon the leader on every measure.

V5 weights knowledge work, coding, science, and general knowledge and hallucination equally, at 25% each. It adds benchmarks for mergeable real-world coding and agentic scientific research, revises or upgrades existing tests, and removes others. Artificial Analysis said private datasets now make up 60% of its benchmarks, up from 45%, to make it harder for models to maximize scores by learning the test material. Speakers noted that benchmarks can lose their power to distinguish models once saturated, while training on test content can inflate results. They also questioned whether scientific discovery deserves equal emphasis with other capabilities and argued that category-specific rankings, including results at different effort settings, can be more useful than a single headline score.

Carbon’s release status leaves its ranking provisional in practical terms. Google announced it two weeks after Argonne, before general availability, and said it was collecting feedback and refining guardrails. The planned rollout starts with selected cyber defenders, then trusted partners, before wider access. The host said the public should wait to judge the model until people can use it. Arena render and game-design tests were generally less impressive than Opus 5.5, although the host questioned whether 3D design is a suitable general test of language models. Peter Gostev described Carbon as a real improvement over Argonne, while Chubby Inax said Google had an unreleased model that was better still. Sundar Pichai predicted faster model progress and launches over the next 12 months, citing RSI; the host said a new base model could let Google iterate through reinforcement learning without another lengthy pre-training run.

The reshuffle also highlights a growing gap between frontier and other models. Kimmy K3.1 Pro dropped from fourth to ninth but remained the leading open-weight model at 46; Artificial Analysis said the gap between open-weight and frontier models had widened from three points to 12. Kyle Chan noted Kimmy’s 67% on Terminal Bench 4.0, the highest result, but only 34% on Terminal Bench Science, compared with Astra’s 63%. He said Chinese models can match or surpass US frontier systems on some agentic tasks, while a substantial gap remains in advanced scientific reasoning and research workflows. The new weighting also made Astra’s index results better reflect its reported performance than the previous version did.

Cloud results underscored both AI demand and the scale of investment behind it. Microsoft reported Azure revenue of $33 billion, up 48% year over year, with OpenAI contributing $10 billion, about one-third of Azure revenue; Microsoft also began reporting Azure separately. Google Cloud doubled year over year to $30.7 billion, nearly closing the gap. Google’s overall revenue grew 25%, but Search growth slowed to 12% and YouTube to 9.2% for a third consecutive period. Pichai said AI Mode queries weighed on results because they monetize at a lower rate. Google missed on bottom-line earnings largely because of a markdown on its SpaceX stock, and shares fell as much as 10% after hours. Dan Ives said the cloud result mattered more to investors than weaker Search growth.

Major technology companies kept their capital expenditure forecasts steady, although quarterly spending came in below expectations. Google guided to a 50% increase next year, reaching $300 billion, despite undershooting this year’s figures so far. The account pointed to possible constraints in components, power, labor, and community support; Luke Longo instead attributed the slowdown more to financing. Separately, the DOJ is examining whether NVIDIA’s proposed $13 billion Hugging Face acquisition could give it control of an important AI-model distribution point and affect competition. The deal, announced in September, is expected to close in the first half of 2027 pending regulatory approval. NVIDIA said Hugging Face would remain free to operate and partner broadly. Reuters reported that a wider, multi-year antitrust investigation also covers other acquisitions, venture investments, sales practices, and possible pressure on customers. No charges have been filed, the investigation could end without action, and the account made no claim that NVIDIA acted improperly.