The AI Model Landscape

Founded January 2021 by Dario and Daniela Amodei, who left OpenAI over disagreements about safety versus commercialization. Incorporated as a Public Benefit Corporation. Raised approximately $67 billion across 17 rounds, with a $380 billion post-money valuation as of February 2026. Revenue grew from $1 billion annualized in December 2024 to $14 billion by February 2026, with roughly 85% coming from enterprise customers.

Defining innovation: Constitutional AI, training models against a written set of principles (23,000 words, led by philosopher Amanda Askell) rather than relying solely on human feedback. In early 2026, Anthropic refused Pentagon demands to remove prohibitions on mass surveillance and autonomous weapons, resulting in a presidential order to cease federal use. The case is in federal court. Ironically, the confrontation boosted consumer adoption to over one million new signups per day.

Opus 4.6 (February 5, 2026): 1-million-token context window, 14.5-hour autonomous task horizons, Agent Teams. Sonnet 4.6 (February 17, 2026): near-Opus performance at 60% lower cost, default for most users. Haiku 4.5: fastest and cheapest, powers free tier.

Consistently wins blind writing tests for natural prose. Dominates coding benchmarks at 74-81% on SWE-bench Verified. Lowest hallucination rate among frontier models. First on financial reasoning benchmarks.

Clean conversational interface on web, iOS, Android, and desktop. Key features: Artifacts (interactive outputs created in chat), Projects (dedicated workspaces per engagement), Cowork (January 2026 desktop agent with file system access and browser automation). Model Context Protocol (MCP) has become an industry standard adopted by OpenAI, Google, and Microsoft, with 97 million monthly SDK downloads.

Named clients include AIG and Newfront. Claude for Financial Services launched July 2025 with industry-specific data integrations. Consulting partners Slalom, PwC, Deloitte, and Infosys are building insurance-specific agents for claims, compliance, and underwriting. Sonnet 4.6 achieved 94% accuracy on insurance-specific computer use benchmarks. Pre-built connectors for S&P Capital IQ, FactSet, Morningstar, PitchBook, and Snowflake.

Founded December 2015 as a nonprofit. Created a capped-profit subsidiary in 2019, then restructured in October 2025 as a Public Benefit Corporation. February 2026 mega-round of $110 billion from Amazon ($50B), NVIDIA ($30B), and SoftBank ($30B) established an implied valuation of approximately $840 billion. Revenue projected at $25 billion annualized by early 2026, though the company still burns roughly $17 billion per year and does not expect profitability until 2029-2030.

Microsoft holds approximately 27% of OpenAI with exclusive cloud provider status through 2030. CEO Sam Altman is the only original leader still active. Only two of eleven founding members remain.

GPT-5.4 (March 5, 2026): up to 1 million tokens via API, configurable reasoning depth, native computer use. The GPT-5 family ships near-monthly updates. Excels at coding (57.7% on SWE-bench Pro), mathematics (94.6% on AIME 2025), and broad knowledge work, matching or exceeding industry professionals in 83% of cases across 40+ occupations. Multimodal capabilities span vision, voice, DALL-E 4 image generation, and Sora video generation.

The broadest ecosystem of any AI platform. Custom GPTs, integrations with Gmail, Slack, SharePoint, GitHub, and Shopify. Code Interpreter for data analysis. Canvas for collaborative editing. Memory stores persistent facts across conversations. Testing advertising on the free tier.

Insurify launched the first ChatGPT insurance comparison app in February 2026 leveraging 196 million auto insurance quotes. Experian launched an Insurance Marketplace app across 37+ carriers. Common use cases include claim summarization, FNOL transcript processing, FAQ automation, underwriting support, and policy comparison. 92% of Fortune 500 adoption means most carriers are already somewhere on the ChatGPT learning curve.

Google's AI story begins with the 2014 DeepMind acquisition (Demis Hassabis, 2024 Nobel laureate) and the foundational 2017 "Attention Is All You Need" transformer paper. In 2023, Google consolidated Brain and DeepMind under Hassabis. Planned 2026 capital expenditure of $175-185 billion is the highest AI infrastructure commitment by any single company. Alphabet surpassed $400 billion in annual revenue in 2025.

The journey has included stumbles: the February 2024 image generation fiasco, AI Search recommending to "add glue to pizza," and the departures of Ethical AI leaders Timnit Gebru and Margaret Mitchell. Hassabis credits these crises with forcing Google to rediscover startup roots.

Gemini 3.1 Pro (March 2026) leads on 13 of 16 major benchmarks, including 94.3% on GPQA Diamond and 77.1% on ARC-AGI-2. Native multimodal design processes text, images, video, and audio simultaneously. Context windows extend to 1-2 million tokens. NotebookLM answers strictly from uploaded documents, available as a core Workspace service in 180+ regions.

For Google Workspace organizations, the integration is seamless: AI features in Gmail, Docs, Sheets, Slides, and Meet. NotebookLM for source-grounded research. Deep Think mode for multi-hypothesis reasoning. Google AI Studio for free experimentation, Vertex AI for enterprise deployment with data residency controls.

SIGNAL IDUNA rolled out Gemini Enterprise to 10,000+ employees and sales partners, reporting 30% reduction in information search time and escalation rates dropping from 27% to 3%. Generali Italia uses Vertex AI for model evaluation. American Family Insurance showcased AI transformation at Google Cloud Next '25. Financial services platform Rogo reported hallucination dropping from 34.1% to 3.9% after switching to Gemini 2.5 Flash.

Founded August 2022 by four engineers from OpenAI, Google Brain, DeepMind, and Meta AI. CEO Aravind Srinivas (PhD, UC Berkeley, age 31) built the company on a thesis: Google Search is broken. Raised approximately $1.5 billion at a $20 billion valuation. Revenue at roughly $200 million annualized. In February 2026, Perplexity discontinued advertising entirely, committing to subscription-first. Faces 10+ copyright lawsuits from publishers including The New York Times, Dow Jones, and BBC.

Not a single model but a retrieval-augmented generation pipeline routing to GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro, and its own Sonar models. "Best" mode auto-selects per query. Model Council (February 2026) runs queries across three or more models, synthesizes outputs, and shows where they agree and disagree. Perplexity Computer orchestrates 20+ models with specialized sub-agents. Excels at real-time fact retrieval with citations (93.9% on SimpleQA), averaging 1.9-second response times with an estimated 1-2% hallucination rate.

Focus modes target specific source types: Web, Academic, Finance (SEC filings, earnings data), Social, and Video. Pro Search enables multi-step reasoning with 3x more sources. Deep Research delivers comprehensive cited reports in 2-4 minutes. Finance features include real-time stock quotes, portfolio tracking, SEC filing analysis, and sector comparisons from FactSet, S&P Global, and 40+ data tools.

No confirmed major insurance company deployments have been publicly announced, but capabilities align well with insurance research needs: claims analysis, regulatory research, competitive intelligence, and market surveillance. The Finance focus mode and Document Review feature (auditing contracts for logical consistency, factual accuracy, and contradictions) are directly applicable. GSA agreement offers Enterprise Pro to all U.S. federal agencies.

Meta's AI strategy is counterintuitive: because Meta monetizes through advertising ($135+ billion annual revenue), giving away AI models costs nothing directly while commoditizing the model layer that competitors sell. Planned 2026 capital expenditure of $115-135 billion. Most significant departure: Yann LeCun, Turing Award winner and founding FAIR director, left in November 2025 after being marginalized by new leadership.

Critical nuance: Llama models are "open weights," not open source. The license prohibits use by companies with 700+ million monthly active users without Meta's permission. Internal signals suggest a potential pivot toward closed models (Project Avocado).

Llama 4 (April 2025) introduced Mixture of Experts architecture. Scout: 17 billion active parameters, 10-million-token context window, fits on a single GPU. Maverick: 17 billion active out of 400 billion, 1-million-token context, 128 experts. Both are natively multimodal and support 200 languages. Maverick delivers comparable performance to GPT-4o at an estimated one-tenth the inference cost when self-hosted.

No hosted consumer platform (beyond meta.ai). Access through third-party providers: Together AI, Groq, AWS Bedrock, Azure. Free to download from llama.com or Hugging Face. Requires technical capability to deploy: the 70B model needs approximately 43GB of VRAM, the 405B model requires 230+ GB.

Even executives who never run a Llama model should understand why it matters. Open-weight models create pricing pressure on proprietary vendors and provide negotiating leverage. Self-hosted models run entirely within the security perimeter, eliminating data sovereignty concerns. Fine-tuning enables training on insurance-specific language without sharing data externally. Enterprise adopters include Goldman Sachs, AT&T, and Block/Cash App. A reported mid-sized insurer invested $50,000 in GPU infrastructure and recovered costs in three months.

Liang Wenfeng co-founded High-Flyer, a Chinese quantitative hedge fund managing $8-14 billion. Starting in 2021, he stockpiled thousands of NVIDIA A100 GPUs before U.S. export restrictions, then spun off DeepSeek in July 2023 with zero venture capital funding. The entire operation is bankrolled by the profitable hedge fund. Fewer than 200 employees (compared to OpenAI's 3,000+), with a flat bottom-up culture. Many researchers are fresh university graduates recruited for passion over experience.

The key architectural innovation, Multi-head Latent Attention (reducing memory requirements by 90%+), originated from a young researcher pursuing a personal interest, enabled by this flat structure.

DeepSeek-V3.2 (December 2025): 671 billion total parameters with only 37 billion activated per token, 128K context window, hybrid reasoning modes. DeepSeek-R1 (January 2025): the dedicated reasoning model trained for $5.6 million that matched OpenAI's o1 on mathematical reasoning. API pricing: $0.56 per million input tokens and $1.68 per million output, approximately 10-30x cheaper than equivalent models from competitors.

DeepSeek teaches something no other platform in this comparison addresses: frontier AI can be built for $5.6 million by a team of fewer than 200 people, under sanctions. For executives making AI investment decisions, this is the most consequential strategic question in the current landscape: if this quality of AI can be built at this cost, what does that mean for every assumption underlying your technology roadmap? DeepSeek-R1 briefly became the #1 app on the U.S. App Store while erasing $1 trillion in U.S. market capitalization in a single day.