What Q2 2025 Reveals About the AI Landscape: Key Trends and Model Rankings

Industry Overview

The report, based on extensive hourly API performance tests and millions of human votes, aims to help engineers and enterprises understand AI capabilities and make strategic decisions. It emphasizes that 2025 is the "year of agents," with innovation moving from foundational research to influencing every organization’s operations.

Five Major Trends

Continued improvement of language model intelligence

Tools and connectors enabling integrated intelligent workflows

China leading in language and video capabilities

Coding agents rapidly spreading in development workflows

Video models achieving breakthroughs and quickly improving quality

AI Value‑Chain Integration

Vertical integration varies across players. Google is the most vertically integrated, from TPU accelerators to the Gemini model. NVIDIA, OpenAI, and Meta also have strong multi‑layer presence, while companies like SambaNova and Cerebras focus on niche domains. Large tech firms such as OpenAI, Google, and Alibaba have comprehensive coverage across all AI modalities (language, image, video, speech), whereas smaller challengers concentrate on specific modalities.

Language Model Progress

The Artificial Analysis Intelligence Index v2 (including MMLU‑Pro, GPQA Diamond, etc.) shows xAI Grok 4 leading with a score of 73, surpassing OpenAI o3‑pro (71), Google Gemini 2.5 Pro (70) and DeepSeek R1 (68). OpenAI remains the overall leader, but Google Gemini and DeepSeek are rapidly catching up, while Meta Llama and Mistral have slipped.

According to the AI Adoption Survey (H1 2025, N=591), usage rates are 83 % for OpenAI GPT/o, 80 % for xAI Grok, and 46 % for Google Gemini.

Language Model Rankings

Top inference models include xAI Grok 4, OpenAI o3‑pro, and Google Gemini 2.5 Pro. Chinese labs such as DeepSeek, MiniMax, and Alibaba follow closely. Open‑source models like DeepSeek R1 0528 narrow the gap with proprietary models, but Grok 4’s release re‑establishes a lead.

Efficiency Gains

In Q2 2025, the cost per million tokens dropped dramatically: high‑intelligence models fell from $0.26 to $0.063, driven by models such as DeepSeek R1 0528 and Gemma 3n E4B Instruct. Output speed increased, with Gemini 2.5 Flash‑Lite becoming the fastest high‑intelligence class.

However, inference models now consume up to ten times more tokens than non‑inference models (≈10 M output tokens vs. 78 M inference tokens), and agents require 20 × more requests, pushing compute demand upward. GPT‑4‑level intelligence can be achieved at 1/100 of GPT‑4’s cost, but new deep‑research query workloads may cost up to ten times more.

Agent Workflow Emergence

Q2 saw 12 major coding agents released, including OpenAI Codex and Gemini CLI. Agents are defined as LLM‑driven autonomous systems that use tools to perform tasks such as coding, research, desktop automation, customer support, and sales. Benefits include dynamic planning, system integration, natural collaboration, and error recovery. GitHub Copilot and Cursor dominate the market with usage rates of 84 % and 53 % (N=955).

Training now emphasizes agents and long‑term tool usage, exemplified by Claude 4 series and Kimi K2.

Image and Video Model Innovation

Q2’s focus shifted to video models, with Seedance 1.0 achieving ~150 ELO on text‑to‑video benchmarks, and Kling 1.6 Pro reaching ~200 ELO. Midjourney V1 and Kling 2.1 Pro remain the only variants available for image‑to‑video conversion. Open‑source video models lag behind; Alibaba Wan 2.1 is the open‑source SOTA, while LTX Video v0.9.7 13B ranks 16th.

Image editing models are competitive: GPT‑4o leads, while FLUX.1 Kontext [max] and HiDream‑E1.1 rank in the top five. Chinese labs such as Bytedance SeeDream 3.0 match GPT‑4o, and HiDream Vivago 2.0 rivals Google Imagen 4. Seedance 1.0 leads video innovation, with Google Veo 3 being the only US‑based SOTA video model supporting audio (priced at $0.75 / s for 720p).

Speech and Audio Model Development

Advances in the AI stack make speech agents more natural, powerful, and cheap. Leading text‑to‑speech models include MiniMax Speech‑02‑HD, Cartesia Sonic‑2, and Nari Labs Dia. Open‑source models like Kokoro 82M and Sesame CSM 1B reduce costs, with third‑party inference approaching frontier performance.

End‑to‑end speech models remain limited to OpenAI GPT‑4o, Google Gemini 2.0 Flash, and Amazon Nova Sonic, offering low latency and emotional understanding but with beta APIs. Speech‑to‑text is dominated by OpenAI Whisper, challenged by GPT‑4o transcription and ElevenLabs Scribe. Third‑party Whisper APIs (e.g., Fal, Fireworks) provide the lowest cost and fastest speed.

Accelerator Market Dynamics

Inference demand is accelerating, with multi‑node inference becoming common. NVIDIA Blackwell is widely available; GB200 launches as the first rack‑scale accelerator; AMD announced MI355X. OpenAI reports compute shortages delaying products like Gemini 2.5 Pro.

System performance now exceeds chip performance, supporting >10 T parameter training. Multi‑node inference battles feature DeepSeek (open‑source) and NVIDIA Dynamo. Geopolitical tensions rise, with US bans on H20 and MI308, while Huawei pushes NVL72; Chinese chip manufacturing advances toward Hopper‑class unknowns.

NVIDIA leads frontier training, with AMD, Google, and Groq offering alternatives. The Artificial Analysis System Load Test shows an 8× B200 system achieving ~39 K tokens/s, 8× H200 only ~13 K tokens/s, and per‑query output speed 3.5× faster under high load.

Conclusion

The Q2 2025 State of AI Highlights Report underscores a maturing AI industry where agents, efficient inference, and multimodal models drive strategic decisions, while hardware acceleration and geopolitical factors shape the competitive landscape.