Skip to main content
Silicon Epoch
Chapter 03 · The frontier · Updated July 2026

Frontier Labs: The Global Ecosystem Leaderboard.

The battle for AGI has divided the global ecosystem into capital-intensive U.S. hyperscalers, open-weights Chinese disruptors, and sovereign European champions. Below: flagship models, valuations, and roadmap targets.

Jump to Lab:OpenAIAnthropicGoogle DeepMindxAIMeta AI · MSLDeepSeekAlibaba QwenMistral AICoherePerplexity AIElevenLabsRunway
Lab 01 of 12
OpenAI

OpenAI

San Francisco, USA · founded 2015

CEO: Sam Altman

CapEx / Valuations:

$6.6B Oct 2024 round at $157B valuation; Microsoft >$13B total backing.

The pioneer of ChatGPT. Pivot to personal agentic AGI router systems, funded by Microsoft and sovereign capital.

Flagship models · mid-2026
  • GPT-5.6 Sol (June 26, 2026)[79]
    Flagship model. Features 'ultra' multi-agent mode and 'max' reasoning effort. Reaches 91.9% on Terminal-Bench 2.1 and excels at genomic analyses. Phased rollout restricted to vetted partners. Billed at $5.00 / $30.00.
  • GPT-5.6 Terra[79]
    Balanced workhorse model. 2x cheaper than GPT-5.5 ($2.50 input / $15.00 output). Optimizes price and System 2 tradeoffs.
  • GPT-5.6 Luna[79]
    High-speed, latency-sensitive tier at lowest cost ($1.00 input / $6.00 output). Legacy endpoints map to this model.
  • o3 Reasoning Series[38]
    Reasoning models (o3-mini / o3-pro) optimizing test-time compute for STEM & coding.
Long-term vision

Altman's June 10, 2025 essay 'The Gentle Singularity': 'We are past the event horizon; the takeoff has started. Humanity is close to building digital superintelligence.' Abundance in compute and fusion energy.

10-year plan

Implement the Stargate supercompute cluster program ($100B-$500B multi-site roadmap) · Roll out personal agent Swarms · Partner with nuclear developers.

“AGI will arrive in 2025... and it is pretty close.” — Sam Altman
Lab 02 of 12
Anthropic

Anthropic

San Francisco, USA · founded 2021

CEO: Dario Amodei

CapEx / Valuations:

Amazon $8B total investment. Series E (Mar 2025): $3.5B at $61.5B; Series F (Sep 2025): $13B at $183B valuation.

Safety-first research lab. Pioneers of Constitutional AI and mechanistic interpretability. Claude Code agent dominates developer commits.

Flagship models · mid-2026
  • Claude Fable 5 & Mythos 5 (June 2026)[74]
    Frontier reasoning engines. Temporary US government access suspension in mid-June; re-deployed globally July 1, 2026, with a new collaborative jailbreak framework. Priced at $10.00 / $50.00.
  • Claude Opus 4.8 (May 2026)[15]
    Safety-first flagship model. 1M context, state-of-the-art SWE-bench Pro planning. Priced at $5.00 / $25.00.
  • Claude Sonnet 5 (June 30, 2026)[72]
    Frontier performance across coding, agents, and professional work at scale. 1M context window. Priced at $3.00 / $15.00.
  • Claude Science (June 30, 2026)[73]
    Customizable AI workbench for scientific research, producing auditable artifacts and computing runtimes.
  • Claude Code[15]
    Agentic terminal CLI tool. Generated $2.5B ARR (est.) by February 2026.
Long-term vision

Amodei's 'Machines of Loving Grace' details how a safe powerful AI could compress 50–100 years of biomedical progress into 5–10 years. Demands verifiable safety standards (ASL-3) globally.

10-year plan

Enforce ASL-3/ASL-4 safety triggers on all training runs · Scale the Claude Code ecosystem globally · Deepen defense and interpretability research.

“A powerful AI — smarter than a Nobel laureate — could arrive as early as 2026 or 2027.” — Dario Amodei
Lab 03 of 12
DeepMind

Google DeepMind

London / Mountain View · founded 2010 · merged 2023

CEO: Demis Hassabis

CapEx / Valuations:

Alphabet internal subsidiary; backed by Google's $65B infrastructure CapEx program in 2025.

The scientific AI powerhouse. Developed AlphaGo, AlphaFold 3, AlphaProof, and the Gemini reasoning ecosystem.

Flagship models · mid-2026
  • Gemini 3.5 Flash & Computer Use (June 2026)[85]
    Flagship agentic workhorse. Integrates native computer-use capabilities to take actions across desktop, mobile, and web sandboxes.
  • Gemini 3.5 Live Translate (June 2026)[85]
    Audio model for real-time speech-to-speech translation supporting over 70 languages with natural intonations.
  • Gemma 4 12B (June 2026)[86]
    Trillion-class reasoning and native voice-vision architecture optimized to run locally on consumer computers (16GB RAM).
  • Nano Banana 2 Lite (June 2026)[85]
    Google's fastest and most cost-effective Gemini Image model to date.
Long-term vision

Hassabis claims we are standing in the 'foothills of the singularity' and that AGI will serve as the ultimate multiplier for physics, chemistry, biology, and math proofs.

10-year plan

Integrate Gemini substrate into Google Search, Workspace, and Android · Deploy π0/Robotics generalist policies · Train next-gen simulators.

“AGI is within reach. Five to ten years out. The next decade will be the most exciting in the history of science.” — Demis Hassabis
Lab 04 of 12
Grok

xAI

Memphis / Bay Area · founded 2023

CEO: Elon Musk

CapEx / Valuations:

$20B Series E in Jan 2026; SpaceX announced full acquisition of xAI in Feb 2026 to align sovereign compute and defense systems.

A division of SpaceX. Built Colossus, the world's largest GPU training cluster in Memphis, integrated with Starlink, Optimus, and SpaceX grids.

Flagship models · mid-2026
  • Grok 4.3 & Voice Think Fast (April 2026)[75]
    Flagship model with native computer-use API, 2M context window, real-time X sentiment lookup, and expressive low-latency voice APIs.
  • Grok Build (0.1) & /goal (June 2026)[76]
    Agentic coding agent with terminal CLI, supporting autonomous task completion checklists via `/goal` and multi-session developer dashboards.
  • Voice Agent Builder (July 1, 2026)[77]
    No-code tool to create and deploy personalized voice agents in under 2 minutes, powered by Grok Speech APIs.
  • Grok 5 (In Training)
    Estimated 6T parameter MoE training on the expanded Colossus 2 cluster (targeting true AGI).
Long-term vision

A maximally truth-seeking AI designed to understand the universe. Musk positions xAI as the compute-heavy sovereign alternative, tightly integrated with Tesla robotics and SpaceX launch systems.

10-year plan

Expand Memphis cluster from 100K H100s to a 1-million GPU roadmap · Integrate Grok as the primary brain of Tesla's Optimus Gen 3 humanoid · Deploy edge inference.

“Digital superintelligence will arrive by 2026. Grok 5 will be the most advanced training project in history.” — Elon Musk
Lab 05 of 12

Meta AI · MSL

Menlo Park, USA · founded 2013 · Restructured 2025

CEO: Mark Zuckerberg

CapEx / Valuations:

$65B infrastructure investment in 2025; meta capex projected up to $100B in 2026; owns 49% of Scale AI.

Pivoted from pure open-weights Llama to closed-weights flagship models under Meta Superintelligence Labs.

Flagship models · mid-2026
  • Muse Spark (April 2026)[84]
    First closed-weights model. Re-architected in June 2026 to serve as the unified core substrate powering all Meta AI services.
  • Llama 4 Maverick (Muse Spark)[9]
    400B total / 17B active MoE. Shipped under community weights license with 700M MAU clause.
  • Llama 4 Scout[9]
    109B total / 17B active MoE. Natively multimodal with a 10M token context window.
Long-term vision

Deploying 'personal superintelligence' across Meta Smart Glasses and consumer apps. Releasing open-weights for mid-scale models, but keeping frontier reasoning proprietary to protect value.

10-year plan

Deploy 57 million sq ft of datacenter capacity globally · Expand Ray-Ban Meta glasses AI features · Build 1-million GPU clusters.

“Muse Spark represents a transition to personal superintelligence. It is closed-weights to ensure safety.” — Mark Zuckerberg
Lab 06 of 12
DeepSeek

DeepSeek

Hangzhou, China · founded 2023

CEO: Liang Wenfeng

CapEx / Valuations:

Self-funded. Raised a historic $7B at a $45B valuation in June 2026 from national and private entities.

The cost-efficiency disruptor. Shook the industry by matching closed models at 90% lower training and inference costs.

Flagship models · mid-2026
  • DeepSeek V4 Pro (July 2026)[87]
    1.6T parameter MoE. Official pricing: $0.435 input / $0.87 output per 1M. Supports peak / off-peak pricing multipliers during Beijing business hours.
  • DeepSeek V4 Flash (July 2026)[87]
    Highly efficient budget reasoning model. $0.14 input / $0.28 output per 1M, cached input at $0.0028/M. Legacy endpoints map here.
  • DeepSeek-R1 (Legacy)
    Reasoning model trained with pure RL, scaling test-time compute. AIME math score: 86.7%.
Long-term vision

AGI via algorithmic ingenuity rather than brute-force hardware scaling. Providing high-fidelity open weights to democratize AGI and counter hardware export restrictions.

10-year plan

Optimize models for Huawei Ascend hardware · Keep all models under open-weights MIT license · Accelerate synthetic data verification.

“We are not here to win a price war. We are here to push the boundary of intelligence.” — Liang Wenfeng
Lab 07 of 12
Qwen

Alibaba Qwen

Hangzhou, China · founded 2023

CEO: Eddie Wu

CapEx / Valuations:

Backed by Alibaba Group's cloud capEx; part of China's $28.3B National Semiconductor Big Fund.

The backbone of Alibaba Cloud. Qwen3.7-Max is one of the highest-rated models on Artificial Analysis.

Flagship models · mid-2026
  • Qwen3.7-Max (May 2026)[9]
    Reasoning flagship, 256K context, optimized for 35-hour agent loops. Priced at $2.50 / $7.50.
  • Qwen3.7-Plus[9]
    Multimodal agent foundation model unifying vision, language, and action planning.
  • Qwen3-Coder-Next
    State-of-the-art coding specialist with advanced tool use and repository-level comprehension.
Long-term vision

Eddie Wu's vision is to make Alibaba Cloud the primary runtime for global AI agents. Open weights for smaller models, closed weights for the absolute frontier.

10-year plan

Integrate Qwen into Alibaba Cloud swarms · Deploy autonomous coding models · Target APAC enterprise market share.

“Open weights set the baseline, but agent swarms in the cloud represent the monetization.” — Qwen Team
Lab 08 of 12
Mistral

Mistral AI

Paris, France · founded 2023

CEO: Arthur Mensch

CapEx / Valuations:

Series C (June 2024): €600M at €5.8B valuation. Strategic debt and cloud partnerships from Microsoft, NVIDIA, and BNP Paribas.

The sovereign European champion. Scaled highly efficient models via algorithmic optimizations, supported by French government and corporate partnerships.

Flagship models · mid-2026
  • Mistral Large 3 & Le Chat Work Mode (June 2026)[17]
    Flagship multilingual model integrated with a dedicated 'Work' reasoning mode in Le Chat for complex task orchestration.
  • Mistral OCR 4 (June 2026)[88]
    State-of-the-art document intelligence model supporting 170 languages and layout bounding box extractions.
  • Mistral Vibe 2.0 (June 2026)[17]
    Terminally online autonomous developer coding agents integrated into corporate repositories.
Long-term vision

Mensch positions Mistral AI as Europe's open and independent alternative, refusing to engage in extreme CapEx brute-force training while focusing on efficiency and local sovereignty.

10-year plan

Partner with European enterprises and governments · Scale the Mistral Large API · Optimize models for on-premise hardware deployments.

“We want to make open weights models that can compete with the best proprietary APIs, giving Europe true digital sovereignty.” — Arthur Mensch
Lab 09 of 12
Cohere

Cohere

Toronto, Canada / San Francisco, USA · founded 2019

CEO: Aidan Gomez

CapEx / Valuations:

Valued at $5.5B. Raised $450M Series D (June 2024) and $1B Series E (2025). Backed by NVIDIA, Oracle, and Salesforce.

The enterprise AI leader. Focusing on sovereign, multi-cloud, and privately deployable models tailored for high-volume RAG and business automation workflows.

Flagship models · mid-2026
  • Command A+ (May 2026)[86]
    218B MoE model under Apache 2.0. Built for enterprise RAG and tool use; runs locally on two H100 GPUs.
  • Command A Vision
    Multimodal document understanding model for enterprise analytics.
Long-term vision

AI independence for enterprises. Enabling companies to run frontier-grade reasoning systems entirely within secure, private clouds.

10-year plan

Expand the North workspace client · Accelerate low-bit W4A4 quantization execution · Deepen RAG and multilingual search capabilities.

“Sovereign AI is about empowering organizations to run and control their own models, rather than outsourcing their brains.” — Aidan Gomez
Lab 10 of 12

Perplexity AI

San Francisco, USA · founded 2022

CEO: Aravind Srinivas

CapEx / Valuations:

Valued at $20B as of May 2026 Series E-6. Over $450M ARR. Targeting a 2028 IPO.

The conversational search disruptor. Reimagined search engines as agentic reasoning boxes, undercutting Google's traditional ad-supported model.

Flagship models · mid-2026
  • Sonar Pro (June 2026)
    Advanced web-reasoning search engine combining real-time index retrieval and test-time thought sequences.
  • Sonar Large
    Moderate latency search backend for high-volume consumers.
Long-term vision

Relentless curiosity through answer engines. Replacing the blue-link index with direct, synthesized answers and auditable sources.

10-year plan

Expand revenue sharing partnerships with publishers · Scale enterprise search swarms · Integrate deeper reasoning compute into search feeds.

“We want to be the default answer box for the world, satisfying curiosity in seconds without ad clickbait.” — Aravind Srinivas
Lab 11 of 12

ElevenLabs

New York, USA / London, UK · founded 2022

CEO: Mati Staniszewski

CapEx / Valuations:

Valued at $11B (Feb 2026). Raised a $500M Series D led by Sequoia Capital.

The voice synthesis champion. Developed state-of-the-art text-to-speech, real-time voice translation, and AI audio sound effect generators.

Flagship models · mid-2026
  • Eleven Multilingual v3 (Feb 2026)
    SOTA real-time voice cloning and speech generation across 32+ languages with emotional nuance.
  • Reader App & Voice Isolator
    Consumer audio reading utilities and real-time audio isolation filters.
Long-term vision

Making content universally accessible in any language and voice, with human-like emotional nuance.

10-year plan

Launch conversational interactive APIs for gaming · Expand licensing options for professional voice actors · Support real-time call dubbing.

“Our goal is to break down language barriers through emotional, expressive voice synthesis.” — Mati Staniszewski
Lab 12 of 12

Runway

New York, USA · founded 2018

CEO: Cristóbal Valenzuela

CapEx / Valuations:

Valued at $5.3B (Feb 2026). Raised $315M Series E led by General Catalyst.

The generative video pioneer. Co-creators of Stable Diffusion, building general world models to simulate physical reality.

Flagship models · mid-2026
  • Gen-3 Alpha
    High-fidelity text-to-video and image-to-video foundation models with advanced camera control and temporal consistency.
  • Gen-4 World Model (In Preview)
    Simulating interactive 3D physics environments and temporal world simulations.
Long-term vision

Developing general world models to simulate physical reality, transforming creative storytelling and game engines.

10-year plan

Scale Hollywood production partnerships · Train models on interactive 3D and physics simulators · Integrate real-time editing controls.

“Generative AI is not just a tool; it is a new creative medium that will redefine how we build worlds.” — Cristóbal Valenzuela
Ecosystem Leaderboard

Frontier Model Registry & Pricing

Frontier Labs Comparison and Leaderboard Stats
LabFlagship ModelParameter ScaleContext WindowAPI Pricing (In/Out per 1M)Licensing Model
OpenAIGPT-5.6 Sol (June 2026)Undisclosed MoE1,050,000+ tokensVaries / Partner RolloutProprietary API
OpenAIgpt-oss-120b117B total / 5.1B active128,000 tokensFree (limited)Community Open-Access
AnthropicClaude Fable 5Undisclosed1,000,000 tokens$10.00 / $50.00Proprietary API
AnthropicClaude Opus 4.8Undisclosed1,000,000 tokens$5.00 / $25.00Proprietary API
Google DeepMindGemini 3.5 FlashUndisclosed1,000,000 tokens$0.075 / $0.30Proprietary API
Google DeepMindGemini 3.1 Pro / Flash LiteUndisclosed2,000,000 tokens$2.00 / $12.00Proprietary Preview
xAIGrok 4.3Undisclosed2,000,000 tokens$5.00 / $25.00Proprietary API
Meta AILlama 4 Maverick (Muse Spark)400B total / 17B active1,000,000 tokensFree (inference cost only)Llama Community License
Meta AILlama 4 Scout109B total / 17B active10,000,000 tokensFree (inference cost only)Llama Community License
DeepSeekDeepSeek V4 Pro1.6T parameter MoE1,000,000 tokens$0.435 / $0.87Permissive MIT License
DeepSeekDeepSeek V4 FlashUndisclosed MoE1,000,000 tokens$0.14 / $0.28Permissive MIT License
Alibaba QwenQwen3.7-MaxUndisclosed256,000+ tokens$2.50 / $7.50Proprietary API
Moonshot AIKimi K2.7-Code1T parameters MoE262,000 tokens$0.38 / $1.72Open Weights
Zhipu AIGLM-5.2Undisclosed200,000 tokens$0.80 / $2.56Open Weights (MIT)
XiaomiMiMo-V2-ProUndisclosed1,040,000 tokens$1.00 / $3.00Proprietary API
MiniMaxMiniMax M2.7Undisclosed205,000 tokens$0.30 / $1.20Proprietary API
ByteDanceSeed 2.0 (Doubao)Undisclosed262,000 tokens$0.25 / $2.00Proprietary API
Mistral AIMistral Large 3Undisclosed256,000 tokens$2.00 / $6.00Proprietary API / Commercial
Mistral AIMistral Small 4Undisclosed128,000 tokens$0.20 / $0.60Community weights (non-commercial)
CohereCommand A+ (May 2026)218B sparse MoE128,000 tokensFree (Weights)Apache 2.0 Open Weights
Perplexity AISonar ProSearch Agent MoE32,000 tokensSubscription Pro ($20/mo)Proprietary Answer Engine
ElevenLabsMultilingual v3Audio foundationN/A (Audio) tokens$15.00 / hour generatedProprietary API
RunwayGen-3 AlphaWorld physical simulatorN/A (Video) tokens$0.05 / video secondProprietary Platform
← Previous: Chapter 02 · How It Works
Table of Chapters
Next: Chapter 04 · Use Cases →