Google's Gemini 4 Carbon Tops the Rankings, Anthropic Achieves First Quarterly Profit

Google Releases Gemini 4 Carbon: Security Capabilities and Long-Horizon Tasks as Core Selling Points

Google has officially launched its new-generation frontier model Gemini 4 Carbon, highlighting industry-leading cybersecurity capabilities and long-horizon task handling. On the Artificial Analysis Intelligence Index B5.0, the model scored 59 overall, ranking first.

According to Google, this is its strongest model to date, with a success rate for generating secure code more than double that of Argon. In conjunction with the U.S. government's voluntary testing process, Google first opened the model to select cybersecurity defenders. This release cadence shows that the deployment path for frontier models in security-sensitive domains is taking shape: controlled testing under a government framework first, then gradually expanding access.

Artificial Analysis Intelligence Index V5: A Major Overhaul of Benchmark Evaluation

Artificial Analysis has released the Intelligence Index V5, the largest benchmark restructuring since V4 launched in January 2026. In the new version, 60 percentage points of evaluation weight are for the first time based on private test sets or hidden answers, with the goal of reducing score inflation.

The industry implications of this change are significant. For a long time, public benchmarks have been specifically optimized by model vendors, and the gap between leaderboard scores and actual capabilities has continued to widen. Shifting most of the weight to private test sets means the evaluation system is moving from "gameable" to "hard to game," placing higher demands on models' true generalization ability.

Under the same evaluation system, DeepSeek^Ad V4.1 Flash 1017 recorded 28 percentage points on Frontier Code 2.0, the highest score among all released open-weight models. Open-source models are approaching closed-source frontier models on coding tasks, which is changing the overall evaluation landscape—according to cross-verification from multiple sources, the gap between open and closed source has clearly narrowed in specific sub-capabilities.

Private Test Sets and the "Leaderboard Gaming" Problem: The questions and answers of public benchmarks (such as MMLU and HumanEval) have long been public, allowing model vendors to incorporate similar data into training in a targeted way, which leads to inflated leaderboard scores disconnected from actual generalization ability—the so-called "score inflation." Private test sets refer to evaluation question banks that are not publicly disclosed and whose answers are hidden, so models cannot train specifically for them in advance, making them more reflective of true capability. The cost is reduced transparency and reproducibility of evaluation, making it difficult for the community to independently verify, and private sets may be indirectly leaked over time. Industry Index V5 shifting 60 percentage points of weight to private test sets is a trade-off by evaluation organizations between "gameable" and "trustworthy," and it also means vendors can no longer rely on public leaderboards for marketing.

Anthropic Achieves First Quarterly Profit, Claude Code Becomes a Commercial Engine

Financial data Anthropic shared with investors this week shows that its AI coding tool Claude Code has achieved tremendous commercial success. The company recorded operating profit in the second quarter, with a net revenue retention rate of 900 percentage points.

This is Anthropic's first quarterly profit, a landmark achievement for an AI company that started as a research outfit. Coding tools have become its commercialization breakthrough, also confirming that developer scenarios are one of the AI application areas with the strongest willingness to pay today.

Anthropic also announced on Wednesday a new set of reporting tools to help customers gain deeper insight into the specific value generated by token spending. The feature uses internal Analytics and employee surveys to help enterprises evaluate their teams' work outcomes and costs. As enterprise procurement increasingly focuses on return on investment, converting "token consumption" into "value measurement" is a key step in retaining customers.

Net Revenue Retention (NRR): Net revenue retention measures the revenue contributed by the same cohort of customers one year later relative to one year earlier, typically excluding new customers and counting only expansion, downgrades, and churn among existing customers. An NRR above 100% indicates that existing customers are expanding usage and revenue is growing organically; reaching 900 percentage points (i.e., 9x) means the existing customer base's spending swelled to nine times the original within a year, which typically appears in early products with usage-based billing and rapid penetration. For AI coding tools billed by token, a high NRR both reflects increased depth of developer usage and suggests that revenue is highly dependent on a small number of heavy users, whose sustainability warrants attention.

NVIDIA Open-Sources Parakeet Series Speech Transcription Models

NVIDIA released Numitron 3.5 Transcribe on Wednesday, a free open-source speech transcription model supporting up to 40 languages. The model can handle recordings up to one hour long and can distinguish up to 8 different speakers.

Developers can combine the model with the Granary dataset to build targeted applications for scenarios such as business meetings and medical transcription. This 30-billion-parameter model can run directly on a single chip, covering all major hardware. For enterprises that need local deployment and are sensitive to data privacy, a speech model that runs on a single card lowers the barrier to adoption.

Speaker Diarization: Speech transcription must not only identify "what was said" but also determine "who said it." Speaker diarization uses voiceprint clustering to segment audio into clips from different speakers and labels each segment with a speaker number. When used together with speech recognition, it can output meeting minutes with speaker labels, a key capability for scenarios such as medical transcription and multi-person meetings. Being able to distinguish up to 8 speakers means the model must maintain stable clustering under overlapping speech, short utterances, and similar timbres—particularly difficult for long recordings (such as one hour)—which is also one of the model's main selling points for enterprise scenarios.

Ant Group's Ling Open-Sources Ling 3.1 Flash: Targeting Personal Agent Scenarios

Ling, the large model under Ant Group, has officially open-sourced Ling 3.1 Flash 1019, scoring 44 on the Intelligence Index V4.3.2. The new version targets Personal Agent scenarios, with its score on the complex knowledge work DA Briefcase rising to 1524 ELO.

Ling also open-sourced the base model Ling 3.1 Flash Base with approximately 56 billion total parameters, adopting a hybrid linear MoE architecture. The base model achieved the highest score on 32 of 47 Benchmark tests across 8 domains.

Personal Agent is a new focal point of recent open-source model competition: compared with general conversation, personal assistant scenarios place higher demands on long context, tool calling, and multi-step task planning, and are closer to actual use. Ling's positioning in this direction echoes DeepSeek's breakthrough on coding tasks—the open-source camp is shifting from "following" to "staking out positions" in segmented scenarios.

Hybrid Linear MoE Architecture: MoE (Mixture of Experts) splits a model into multiple expert subnetworks, activating only a few of them in each forward pass, thereby controlling inference compute despite a large total parameter count. The compute of traditional attention mechanisms grows quadratically with context length, making long-context scenarios costly; linear attention uses approximation methods to reduce complexity to nearly linear. Hybrid linear MoE uses linear attention in some layers and retains standard attention in others, balancing long-context efficiency and expressiveness. This explains why a 56-billion-parameter base model can still be deployed for long-context, multi-step tool-calling scenarios such as personal assistants.

Google Earnings and Chrome Update: AI Search Ads Go Global, Browser Gets Smarter About Forms

Alphabet's third-quarter revenue reached $127.6 billion, up 25 percentage points year over year, but growth in its core search advertising business continued to slow. On Wednesday's earnings call, Google announced it would expand the trial of advertising in its AI search experience to international markets. Google's Chief Business Officer said high-quality ads are gradually being rolled out at scale in AI Overviews within AI Mode. Meanwhile, Google Services revenue reached $96.3 billion, up 11 percentage points year over year.

The slowdown in search advertising growth and the accelerated global expansion of AI search ads form a noteworthy contrast: Google is migrating ad inventory from traditional search results pages to AI-generated answer interfaces to hedge against growth pressure on the old model.

The Chrome browser has also received a major update, using AI technology to greatly improve the intelligence of form autofill. The updated Chrome can extract information such as flight confirmation numbers, order numbers, and addresses from users' Gmail and automatically suggest key information. The feature is a further extension of Google's Personal Intelligence; sensitive data is encrypted, and users can choose in Chrome settings whether to enable enhanced autofill.

Industry Observations: Benchmark Restructuring, Open-Source Rise, and Commercialization Validation

Looking at these stories together, three main threads emerge.

First, the credibility of the evaluation system is being rebuilt. Artificial Analysis shifting most of its weight to private test sets—the restructuring of the Artificial Analysis Intelligence Index itself is a response to "leaderboard gaming culture"; and DeepSeek's lead on Frontier Code 2.0 shows that competition on real capabilities has begun.

Second, open-source models continue to make breakthroughs in segmented scenarios. Ling targets Personal Agents, DeepSeek focuses on coding, and NVIDIA bets on multilingual speech transcription—the open-source camp no longer pursues comprehensive benchmarking but chooses to be strong at single points.

Third, the commercialization path is gradually becoming clear. Anthropic achieved its first quarterly profit through coding tools, while Google is trying to turn AI search into a new advertising vehicle. Beyond technical capability, whoever can first embed AI into users' payment habits will be able to close the commercial loop sooner.

One-line Summary

Google releases Gemini 4 Carbon, leading in security and long-horizon tasks, as benchmark restructuring and the open-source rise proceed in parallel.

Paragraph Summary

Google launched its new-generation frontier model Gemini 4 Carbon, highlighting cybersecurity and long-horizon task capabilities and ranking first on the Artificial Analysis Intelligence Index. Meanwhile, Artificial Analysis released Intelligence Index V5, for the first time basing 60 percentage points of weight on private test sets to reduce score inflation. Anthropic achieved its first quarterly profit thanks to Claude Code, with a net revenue retention rate of 900 percentage points. NVIDIA open-sourced a speech transcription model supporting 40 languages and distinguishing 8 speakers, while Ant Group's Ling open-sourced the Ling 3.1 Flash series targeting Personal Agents. Google's earnings show slowing search advertising growth, and it is migrating ads to AI search interfaces. Overall, rebuilding benchmark credibility, open-source positioning in segmented scenarios, and a gradually clearer commercialization path constitute this week's three main threads.