Google Leads AI Model Rankings as OpenAI Cuts Prices and Voice Tools Advance

AI competition is shifting on two fronts: model performance and the cost and setting of deployment. Google says Gemini 4 Carbon has taken the top position on the Artificial Analysis Intelligence Index, while OpenAI has cut the price of a mid-tier model by half. At the same time, new speech systems and smaller models are aimed at customer calls, local devices and other specific business uses.

Those moves point to a market where a single ranking tells only part of the story. Buyers also have to weigh speed, accuracy, data handling and the work a model can complete in a particular setting. The latest releases give companies more options, while leaving open how results on tests and demonstrations will translate into day-to-day performance.

Google’s lead comes with a cybersecurity focus

Google DeepMind announced Gemini 4 Carbon on Oct. 28, two weeks after Gemini 4 Argon. Google describes Carbon as its strongest model to date. It ranks first on the Artificial Analysis Intelligence Index, with a score of 59, according to the company.

Carbon’s stated area of strength is cybersecurity. It can identify software flaws, assess whether they could be exploited and draft fixes. Google’s Chrome team says Carbon performed more than 10 times better than another lab’s large frontier model at writing high-quality security fixes. That comparison is a company-reported result, and the source does not provide further detail on the test conditions.

The emphasis matters for organizations that use AI to review code: finding a weakness is only one part of the task, since teams also need to judge its severity and produce a workable correction. Google’s claims suggest it is positioning Carbon around that full sequence. But rankings and vendor comparisons do not establish how reliably the model will perform across different codebases or operating conditions.

OpenAI’s price cut targets routine computer work

OpenAI released GPT-6.1 Terra on Oct. 26 at half the price of its predecessor. The company says the mid-tier model handles computer tasks, including filling out forms and navigating software, more than three times faster than its flagship GPT-6 Astra and GPT-6.1 Sol models.

OpenAI also reports an improvement on an internal benchmark: the error rate fell to 16.7% from 53.6% for the previous model, GPT-5.6 Terra. Those figures make the release notable for businesses weighing automated office tasks, though they come from OpenAI’s own testing. The source does not specify how the benchmark’s tasks compare with the mix of work found in a typical workplace.

GPT-6.1 Terra is available to paid ChatGPT subscribers and developers through OpenAI’s API. The lower price could make it easier to test automation across more workflows. The practical calculation will still depend on the cost of correcting errors, the time saved and whether the system can complete each task reliably.

Moonshot targets research-heavy work

Moonshot AI launched Kimi K3.1 Pro on Oct. 23. The Beijing-based lab says its latest model has improved at coding, 3D creation and financial research. Demonstrations included a linked financial model, an investment committee deck and an industry report with sources.

The model can read roughly 1,500 pages at once, according to the source article. That capacity could help with large document collections, such as a deal file, by letting teams ask questions across material that might otherwise need to be reviewed in separate batches. Moonshot plans to make the model free to download in early November. The article provides no independent assessment of its research accuracy or the reliability of its citations.

Speech models change how calls are handled

Block introduced an early version of Dessa on Oct. 23. The speech model was trained only on recordings of human conversation, without text, and is designed to listen and speak at the same time. Block says it can laugh mid-sentence and recognize sarcasm and frustration. The company frames the work around safety, arguing that a voice system should be able to detect when a caller is distressed.

Dessa runs locally on a MacBook Pro M4 Max or better, and early access is limited to a small group of AI safety researchers. A demonstration drew praise online as well as concerns about scams. Its local operation and conversational design make it a different proposition from a cloud-based call tool, but wider use will depend on how well it handles real conversations and difficult situations.

Rime’s Prism, launched Oct. 28, is designed for enterprise call centers. It drafts a response as text so rules can screen the wording before the system speaks, while it continues listening to the caller. Rime says it responds in about half a second.

On the telecom portion of τ-Voice, an Artificial Analysis benchmark based on simulated customer service calls, Prism scored 82.5, compared with 45.6 for OpenAI’s GPT-Live-1. Rime says companies can fine-tune Prism using their own call recordings, although self-serve tools remain in beta. The score offers one point of comparison on simulated calls; companies will still need to assess how the system performs with their own scripts and customer interactions.

Smaller models bring more work onto local devices

Nvidia released Nemotron 3.5 Bolt E4B on Oct. 23. The free model is designed to run on a PC without an internet connection, supporting voice assistants and simple multistep tasks on local hardware.

Liquid AI introduced LFM3 on Oct. 27, a family of six free, small models that handle text, images, audio and quick decisions. The company says they run on phones and laptops and can be fine-tuned on a single graphics chip. Mercedes-Benz and Shopify tested the models before release; Liquid identifies real-time fraud detection as one possible use.

Running models on local devices can reduce reliance on cloud access and may suit tasks that need quick responses. It also makes device capacity and model performance central to deployment choices. Across these releases, the competitive test is becoming more specific: the top index score matters, but so do the cost of routine automation, the quality of speech handling and the ability to operate where the work happens.