Signals

Comprehensive 135-page survey from UIUC, Meta, Amazon, and Google DeepMind establishes agentic reasoning taxonomy. Identifies three capability layers: foundational (planning, tool use), self-evolving (feedback, memory, adaptation), and collective multi-agent reasoning.

GAIA benchmark reveals a persistent capability gap between humans and AI on tasks trivially easy for humans. On Level 1 (simplest tasks), humans achieve 92% while best AI achieves 75%. The gap widens at higher difficulty levels, signaling fundamental limitations in AI reasoning and tool use.

Writer Action Agent achieved 61% on GAIA Level 3 (hardest tier), surpassing Manus AI (57.7%) and OpenAI Deep Research (47.6%). GAIA tests multi-step reasoning, tool use, and web browsing - capabilities critical for general AI assistants.

Claude Opus 4.5 achieved 80.9% on SWE-bench Verified, becoming the first model to exceed the 80% threshold. This represents a 21x improvement from GPT-4's initial 3.8% score in October 2023, demonstrating rapid progress in AI coding capabilities.

NeoSignal model database pruned from 409 to 50 frontier models using smart de-duplication. Removed 161 duplicate versions and 198 lower-scoring models. Focus now on top-performing models from OpenAI, Anthropic, Google, xAI, and other leading providers.

AMD MI300X achieves highest per-GPU LLM inference throughput in MLPerf Inference v5.1, delivering 21,150 tokens/s per GPU on llama2-70b, outperforming NVIDIA H100 (15,610 tok/s), B200 (13,015 tok/s), and H200 (10,917 tok/s). Industry-standard benchmark validates AMD's competitiveness in AI inference.

Structured agent evaluation is rapidly professionalizing. Enterprises now require sandbox environments, human-in-the-loop testing, and specialized scoring methodologies before deploying AI agents. Harbor, Promptfoo, and Braintrust emerging as leading evaluation platforms.

Agent evaluation metrics shifting from single-attempt success to probabilistic measures. Pass@k (success in any of k attempts) and pass^k (success in all k attempts) becoming standard for measuring both capability ceiling and reliability floor.

Epoch AI's estimate of GPT-4o energy cost was corroborated by Sam Altman and matches Google's Gemini data. Average query uses less energy than running a lightbulb for 5 minutes.

While 36% of AI users report using it for work, only 18% have employer-provided AI access. This gap indicates substantial bottom-up workplace adoption through personal subscriptions and free tiers.

ChatGPT grew from under 400M weekly active users in January 2025 to nearly 800M by August, adding roughly 50M new users per month. Growth has since slowed but remains substantial.

The best open models runnable on consumer GPUs lag frontier AI by only ~1 year across GPQA, MMLU, and LMArena benchmarks, suggesting rapid capability democratization and regulatory implications.

Labs like OpenAI and Anthropic claim RL reasoning scaling cannot be sustained beyond 1-2 years due to compute infrastructure limits, suggesting the exceptional 2024-2025 capability growth could slow.

Installed AI compute from NVIDIA chips has more than doubled annually since 2020, with new flagship chips accounting for most compute within 3 years of release.

The ChatGPT mobile app was downloaded 1.9 billion times between October 2024 and September 2025, making it the most downloaded app of the year as usage shifts from web to mobile.

Between August and November 2025, Gemini monthly active users increased ~30% compared to only 5% for ChatGPT, though ChatGPT maintains the larger userbase with 35% vs 24% weekly usage among US voters.

58% of AI users report using LLMs for information lookup, far exceeding writing (32%), advice (30%), technical tasks (18%), and multimedia (17%). Most users treat AI as enhanced search, not autonomous agents.

ChatGPT daily users in India grew 7x over the past year, resulting in India having over twice the ChatGPT users of the US. Gemini users in India grew 15% in November alone.

Between April 2023 and March 2025, LLM API prices fell by more than 10x at equivalent performance levels, indicating a highly competitive market and significant efficiency gains.

DeepSeek V3 used 10x less compute than Llama 3 through MLA (multi-head latent attention), MoE innovations, and multi-token prediction, demonstrating 3x yearly algorithmic efficiency gains.

AI-native startups have captured 63% of the application market (up from 36% last year), earning $2 in revenue for every $1 earned by incumbents. Product velocity and PLG motions driving startup advantage.

Model Context Protocol (MCP), introduced by Anthropic and quickly adopted by OpenAI, Google DeepMind, and Microsoft, is becoming the universal specification for agents to access external APIs, tools, and real-time data.

The browser is emerging as the dominant interface for agentic AI, transforming from a navigation tool into a programmable environment for autonomous execution. Perplexity's Comet and OpenAI's Atlas leading this shift.

Enterprise generative AI spending hit $37 billion in 2025, up 3.2x from $11.5B in 2024, making it the fastest-scaling software category in history and representing 6% of the global SaaS market.

ChatGPT maintains massive consumer adoption with 800-900M weekly active users. Only 9% of consumers pay for multiple AI subscriptions, indicating 'winner take most' market dynamics rather than fragmentation.

Despite the agent hype, only 16% of enterprise and 27% of startup deployments qualify as true agents. Most production architectures remain simple, built around fixed-sequence or routing-based workflows.

Anthropic's Claude Code achieved a $1 billion run rate in just six months, demonstrating the massive demand for AI-powered coding tools among prosumer and technical users.

Healthcare AI captured $1.5 billion in 2025 (tripling from $450M), representing nearly half of all vertical AI spend. Ambient scribes market alone reached $600M, minting two new unicorns (Abridge and Ambience).

Anthropic has unseated OpenAI as the enterprise LLM leader, growing from 12% market share in 2023 to 40% in 2025. OpenAI fell from 50% to 27% over the same period.

xAI's Grok Code Fast 1 has surged to the #1 position on OpenRouter with 572.7B tokens processed weekly, more than 3x the second-place model. This dethroned mimo-v2-flash which dropped from #1 (170.9B) to #9 (77.6B), signaling a major shift toward specialized coding models.

xAI's Grok has emerged as a significant consumer AI player, growing from zero to 9.5M DAU and 38M MAU by December 2025. Animated Companions and integrated X app features driving adoption.

Google Gemini is accelerating growth with 155% YoY desktop user growth compared to ChatGPT's 23%. Gemini Pro subscriptions growing nearly 300% YoY versus 155% for ChatGPT Plus.

Code generation has become AI's first true 'killer use case' with $4 billion in spending (55% of departmental AI). 50% of developers now use AI coding tools daily, with Cursor capturing significant share from GitHub Copilot.

Text embedding models have emerged as a distinct category in the AI stack, with MTEB standardizing evaluation across 8 task types. The top performers (gte-Qwen2, NV-Embed-v2, voyage-3-large) achieve ~70% overall scores with vector dimensions ranging from 768 to 4096, enabling specialized retrieval and semantic search applications.

AgentBench evaluation reveals GPT-4 class models achieve ~44% overall success rate across 8 real-world agent environments, while open-source alternatives score 15-30% lower. Operating System and Database tasks show the largest capability gaps, highlighting the challenge of autonomous agent development.

NVIDIA H100 GPU rentals have reached record low prices on marketplace platforms, with Vast.ai offering $1.49/hr and multiple providers below $2/hr. This represents a 75%+ reduction from hyperscaler on-demand pricing, fundamentally shifting GPU compute economics.

Alibaba's gte-Qwen2-7B-instruct has claimed the top position on MTEB with ~70% overall score, excelling in retrieval (nDCG@10) and semantic textual similarity tasks. The 7B parameter model with 3584 dimensions outperforms NVIDIA's NV-Embed-v2 and Voyage AI's voyage-3-large.

OpenAI announced a significant 40% price reduction for GPT-4o API access, with input tokens now at $2.50/1M and output tokens at $10.00/1M, intensifying price competition in the LLM market.

A critical prompt injection vulnerability (CVE-2025-12345) has been discovered in LangChain versions prior to 0.2.0, affecting chain-of-thought and agent implementations.

Anthropic's Claude Opus 4.5 has achieved a new state-of-the-art score of 93.2% on MMLU, surpassing GPT-5.1 and demonstrating significant advances in broad knowledge understanding.

CoreWeave is the only cloud provider to achieve Platinum tier in SemiAnalysis ClusterMAX 2.0 rankings, citing exceptional GPU availability and performance.

Claude Code has emerged as the most adopted AI coding agent, with developers citing superior code understanding, autonomous task completion, and seamless IDE integration as key differentiators.

AWS has launched P5 instances featuring NVIDIA H200 GPUs, now generally available in US East and West regions with EU availability expected in Q1 2026.

Microsoft and OpenAI announced an expanded partnership bringing GPT-5 and o1 models exclusively to Azure for enterprise customers, with dedicated capacity guarantees.

Gemini 3 Pro has reached 1490 ELO on LMArena, surpassing Claude Opus 4.5 and GPT-5.1 to claim the top position in human preference rankings.

Google has announced TensorFlow 2.x will enter maintenance mode in Q2 2026, with new development focused on JAX and TensorFlow 3.0 architecture.

DeepSeek R1 has emerged as the leading open-source model for mathematical reasoning, outperforming many closed-source alternatives on MATH and GSM8K benchmarks.

PyTorch has surpassed 100 million weekly downloads on PyPI, cementing its position as the dominant deep learning framework for research and production deployments.

Enterprise teams are increasingly adopting multi-agent frameworks like CrewAI and AutoGen for complex workflows, moving beyond single-agent implementations to orchestrated agent teams.

NVIDIA H200's 141GB HBM3e memory requires updated CUDA drivers and framework versions. Teams should verify compatibility before migration from H100.

NVIDIA's B200 Blackwell GPUs are shipping to hyperscalers, promising 2.5x performance gains over H100 for AI training workloads.

AMD's MI300X is seeing increased adoption as enterprises seek alternatives to NVIDIA's supply-constrained GPUs, with 192GB memory enabling larger model deployments.

Tool invocation has seen consistent upward trend throughout 2025. Initially concentrated in GPT-4o-mini and Claude 3.5/3.7, tool use has diversified to Claude 4.5 Sonnet, Grok Code Fast, and GLM 4.5. Enabling tool use is becoming table stakes for enterprise adoption.

Open source models have stabilized at ~30% market share, with DeepSeek leading at 14.37T tokens. Chinese OSS models grew from 1.2% to nearly 30% of total usage, reshaping competitive dynamics.

vLLM has become the de facto standard for LLM inference, with major cloud providers and inference platforms adopting it for production deployments.

Reasoning-optimized models (o1-class, GPT-5, Claude 4.5, Gemini 3) now account for over 50% of all LLM token usage, up from negligible share in early 2025. This reflects a fundamental shift from single-pass text completion to multi-step deliberation inference.

Coding-optimized models now dominate OpenRouter's top 10: Grok Code Fast (#1), Claude Sonnet (#2), Claude Opus (#3), and Kwaipilot's kat-coder-pro-v1 (#4, 118.4B tokens). Mistral's Devstral 2512 (#8, 81.3B) adds to the coding focus. This reflects the broader industry shift where programming surpassed roleplay as the dominant LLM use case.

Programming queries grew from 11% to over 50% of LLM usage in 2025. Anthropic's Claude dominates with 60%+ of programming spend, while average programming prompts now exceed 20K tokens.

Anthropic's Claude 4.5 Sonnet and Opus have claimed the #2 and #3 positions on OpenRouter with 176.7B and 121.2B tokens respectively. Combined, Claude models process nearly 300B tokens weekly, demonstrating strong enterprise adoption and programming use case dominance.

LangGraph has matured into a production-ready framework for complex agent orchestration, with ThoughtWorks recommending 'Adopt' status in their Technology Radar.

Asia's share of global LLM spend more than doubled from 13% to 31%, with Singapore (9.21%), China (6.01%), and South Korea (2.88%) leading. North America retains 47%, but global distribution is increasingly decentralized.

New agent-focused frameworks are gaining traction: FastMCP simplifies MCP server implementation, LiteLLM provides unified multi-provider access, and Pydantic AI offers type-safe agent development. Model Context Protocol (MCP) is becoming the standard for LLM context and tooling.

ThoughtWorks Technology Radar Vol.33 elevated LangGraph and vLLM to 'Adopt' status, recommending them for broad production use. vLLM has become the default inference engine for Azure AI Model Catalog.

SemiAnalysis launched InferenceMAX, an open-source nightly benchmark comparing GPU inference performance across NVIDIA (H100, H200, B200, GB200 NVL72) and AMD (MI300X, MI325X, MI355X). Endorsed by Jensen Huang, Lisa Su, OpenAI, and Microsoft. First multi-vendor benchmark with TCO and power efficiency metrics.

NVIDIA B200 delivers ~3x power efficiency vs H100 for GPT-OSS 120B FP4 (2.8M vs 900K tok/s/MW). Similar gains seen in AMD CDNA3 to CDNA4 (MI355X 3x better than MI300X). Blackwell 20% more energy efficient than MI355X due to lower TDP (1kW vs 1.4kW).

NVIDIA GB200 NVL72 with TRT-LLM Dynamo achieves 4x better TCO per million tokens than single-node servers for DeepSeek R1 at 30 tok/s/user. Rack-scale inference with disaggregated prefill, wide expert parallelism, and multi-token prediction (MTP) delivers 2-3x throughput gains.

AMD MI355X delivers lower TCO per million tokens than NVIDIA B200 for GPT-OSS 120B FP4 summarization at interactivity below 225 tok/s/user. MI300X also beats H100 on GPT-OSS 120B MX4 across all interactivity levels. B200 leads on LLaMA 70B FP4 and high-interactivity workloads.

Anthropic published definitive guidance distinguishing 'workflows' (predefined orchestration) from 'agents' (dynamic self-directed systems). Key patterns: prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer. Recommends simple patterns over complex frameworks.

Devin emerged as the first commercial AI software engineer, achieving $2B valuation and 53.8% on SWE-bench Verified. As the pioneering commercial coding agent, it established the market category now contested by Claude Code, Cursor, and others.