AI Software Foundation Models
Frontier Foundation Labs
| # | Lab | Country | Founded | Latest model | Last raise | Valuation | ARR (latest) | Open / Closed | Backers |
|---|---|---|---|---|---|---|---|---|---|
| 1 | OpenAI | US | 2015 | GPT-5 2025-08 | $6.6B (2025-10) | $500B | $13Bas of 2025-09 | closed | Microsoft (preferred compute partner, equity stake), Thrive Capital, SoftBank, Nvidia, Khosla Ventures, Tiger Global |
| 2 | Anthropic | US | 2021 | Claude Opus 4.5 / Sonnet 4.5 2025-09 | $5.0B (2025-08) | $170B | $5.0Bas of 2025-07 | closed | Amazon (multi-billion strategic, Bedrock distribution, Trainium/Inferentia compute), Google (multi-billion, GCP TPU compute), Lightspeed, Spark Capital, Fidelity |
| 3 | Google DeepMind | US/UK | 2010 | Gemini 2.5 Pro 2025-03 | n/a | n/a | n/d | closed | Wholly owned by Alphabet (GOOGL). DeepMind acquired by Google 2014; merged with Google Brain into Google DeepMind in April 2023. |
| 4 | SpaceXAI | US | 2023 | Grok 4 2025-07 | $10B (2025-07) | $80B | n/d | mixed | Valor Equity Partners, Andreessen Horowitz, Sequoia, Vy Capital, Fidelity, Saudi Kingdom Holding. Compute partnerships with Oracle (OCI) and on-prem Colossus cluster (Memphis). Grok models distributed via X (formerly Twitter) consumer app. |
| 5 | Meta AI / FAIR | US | 2013 | Llama 4 2025-04 | n/a | n/a | n/d | open-weights | Wholly inside Meta Platforms (META). FAIR founded 2013; Llama line distributed under Llama Community License (open-weights with commercial use clause for entities under 700M MAU). |
| 6 | Mistral AI | FR | 2023 | Mistral Large 2 2024-07 | $640M (2024-06) | $6.2B | n/d | mixed | General Catalyst, Andreessen Horowitz, Lightspeed, Microsoft (commercial partnership), NVIDIA, Salesforce, BNP Paribas, Bpifrance (France sovereign). Distribution via Azure AI Foundry, Bedrock, Vertex AI. |
| 7 | Cohere | Canada | 2019 | Command A 2025-03 | $500M (2024-07) | $5.5B | $100Mas of 2024-12 | mixed | PSP Investments, Cisco, Fujitsu, AMD, Inovia, Index Ventures, Tiger Global. Sales focus on enterprise on-prem and sovereign deployments via Oracle OCI partnership. |
| 8 | AI21 Labs | Israel | 2017 | Jamba 1.6 2025-03 | $208M (2023-11) | $1.4B | n/d | mixed | Walden Catalyst, NVIDIA, Intel Capital, Google, Pitango, SCB10X. Distribution via AI21 Studio plus Azure, Bedrock, Vertex. |
| 9 | DeepSeek | CN | 2023 | DeepSeek-R1 2025-01 | n/a | n/a | n/d | open-weights | Wholly funded by High-Flyer Quant (parent quantitative-trading firm); no external venture capital reported as of the as-of date. Compute on a self-operated A100 plus H800 cluster in China. |
| 10 | Zhipu AI (Z.ai) | CN | 2019 | GLM-4.5 2025-07 | $400M (2024-12) | $3.0B | n/d | mixed | Alibaba, Tencent, Meituan, Xiaomi, Hongshan (formerly Sequoia China), Hillhouse. Tsinghua University KEG lab spinout. Registered with the Cyberspace Administration of China (CAC). |
| 11 | Alibaba Qwen | CN | 2023 | Qwen3 / Qwen 2.5 Max 2025-04 | n/a | n/a | n/d | mixed | Wholly inside Alibaba Cloud (BABA). Qwen 2.5, Qwen 2.5 Coder, Qwen3 weights released under Apache-2.0 in most sizes. Flagship Qwen 2.5 Max remains API-only on Alibaba Cloud Model Studio. |
| 12 | 01.AI | CN | 2023 | Yi-Lightning 2024-10 | $200M (2024-05) | $1.0B | n/d | mixed | Alibaba Cloud, Hongshan, Sinovation Ventures. Founded by Kai-Fu Lee. Earlier Yi-34B and Yi-9B weights open under Yi Series License (commercial use allowed with registration). |
| 13 | Tencent Hunyuan | CN | 2023 | Hunyuan-Turbo S 2025-02 | n/a | n/a | n/d | mixed | Wholly inside Tencent (TCEHY / 0700.HK). Hunyuan-Large MoE released open-weights Nov 2024 under Tencent Hunyuan Community License. |
| 14 | Reka AI | US | 2022 | Reka Flash 3 2025-03 | $60M (2024-09) | $300M | n/d | mixed | DST Global, Radical Ventures, Snowflake Ventures, Nvidia. Snowflake reportedly explored acquisition mid-2024 ($1B range) but deal did not close. Reka Flash 3 released open-weights under Apache-2.0 Mar 2025. |
| 15 | Black Forest Labs | Germany | 2024 | FLUX.1.1 Pro Ultra 2024-11 | $31M (2024-08) | $1.0B | n/d | mixed | Andreessen Horowitz, General Catalyst, MätchVC. Founded by ex-Stability AI researchers (creators of Stable Diffusion). FLUX models distributed via API plus partner platforms (xAI Grok uses FLUX for image generation in X). |
Frontier Lab Revenue Trajectory
Milestone Disclosures
- 2023-08OpenAI. Crossed $1B annualized revenue. (The Information)
- 2024-02OpenAI. Crossed $2B annualized revenue. (Financial Times)
- 2024-08OpenAI. Crossed $4B annualized revenue. (The Information)
- 2024-12Anthropic. Crossed $1B annualized revenue. (The Information)
- 2025-06OpenAI. Crossed $10B annualized revenue. (Reuters)
- 2025-07Anthropic. Crossed $5B annualized revenue. (Bloomberg)
Model Capability Benchmarks
| Model | MMLU-Pro% | GPQA Diamond% | SWE-Bench Verified% | AIME 2025% | HLE% | Arena Elo | MMMU-Pro% | ARC-AGI v2% |
|---|---|---|---|---|---|---|---|---|
GPT-5 OpenAI extended-thinking variant (high) | 87.5 | 89.4 | 74.9* | 94.6* | 26.5 | 1,410 | 79.4* | 6.4 |
Claude Opus 4.5 Anthropic extended-thinking enabled | 86.9 | 87.9* | 77.2* | 90.0* | 24.5 | 1,388 | 76.5* | 4.6 |
Claude Sonnet 4.5 Anthropic extended-thinking optional | 83.4 | 83.4* | 77.2* | 87.0* | 19.8 | 1,370 | 70.2* | n/a |
Gemini 2.5 Pro Google DeepMind deep-think mode enabled | 86.7 | 86.4* | 67.2* | 86.7* | 21.6 | 1,397 | 81.7* | 4.9 |
Grok 4 SpaceXAI Heavy variant (parallel agents) | 86.6 | 87.5* | 72.0 | 95.0* | 25.4 | 1,380 | n/a | 15.9 |
Llama 4 Maverick Meta no extended-thinking variant | 80.5* | 69.8* | 43.4 | 63.0 | 8.4 | 1,271 | 73.4* | n/a |
DeepSeek-R1 DeepSeek reasoning model (think-aloud) | 84.0 | 71.5* | 49.2* | 87.5* | 8.6 | 1,357 | n/a | 1.3 |
Qwen3 235B Alibaba hybrid thinking-mode toggle | 81.8 | 71.1* | 47.0 | 81.5* | 11.8 | 1,331 | n/a | n/a |
OpenAI o3 OpenAI reasoning model (high effort) | 85.0 | 87.7* | 71.7* | 88.9* | 20.3 | 1,374 | 82.9* | 3.5 |
GLM-4.5 Zhipu (Z.ai) thinking-mode toggle | 84.6* | 79.1* | 64.2* | 91.0* | n/a | 1,325 | n/a | n/a |
API Token Pricing
| Model | Provider | Input$/MTok | Output$/MTok | Cached input$/MTok | Contexttokens | As of |
|---|---|---|---|---|---|---|
Llama 4 Maverick | Meta (via partners) | $0.27 | $0.85 | n/a | 1.0M | 2025-04-08 |
DeepSeek-R1 | DeepSeek | $0.55 | $2.19 | $0.14 | 64K | 2025-02-08 |
Qwen3 235B | Alibaba Cloud | $0.60 | $1.80 | n/a | 128K | 2025-04-29 |
GLM-4.5 | Zhipu (Z.ai) | $0.60 | $2.20 | n/a | 128K | 2025-07-28 |
GPT-5 | OpenAI | $1.25 | $10.00 | $0.13 | 400K | 2025-08-07 |
Gemini 2.5 Pro | $1.25 | $10.00 | $0.31 | 1.0M | 2025-06-17 | |
OpenAI o3 | OpenAI | $2.00 | $8.00 | $0.50 | 200K | 2025-06-10 |
Claude Sonnet 4.5 | Anthropic | $3.00 | $15.00 | $0.30 | 200K | 2025-09-25 |
Grok 4 | SpaceXAI | $3.00 | $15.00 | $0.75 | 256K | 2025-07-09 |
Claude Opus 4.5 | Anthropic | $15.00 | $75.00 | $1.50 | 200K | 2025-09-25 |
Token Price Decline Over Time
Annotated Events
- 2023-03-14GPT-4 launches at $30 / $60 ($/MTok)
- 2023-11-06GPT-4 Turbo: 3x input cut
- 2024-05-13GPT-4o multimodal: another 2x cut
- 2024-06-20Claude 3.5 Sonnet undercuts Opus at 1/5 price
- 2024-07-18GPT-4o mini: tier-A capability at $0.15/MTok input
- 2024-12-26DeepSeek-V3 open-weights at sub-$0.30 input
- 2025-01-20DeepSeek-R1: reasoning at GPT-4 Turbo class for $0.55 input. China selloff.
- 2025-08-07GPT-5 flagship at $1.25 input: a 24x cut from GPT-4 original in 29 months
Training Compute and Cost
Not Disclosed (FLOPs)
Open Weights vs Closed Labs
Top Closed vs Top Open Gap (Chatbot Arena Elo)
53 Elo
53 Elo between #1 closed (GPT-5) and #1 open (DeepSeek-R1) at the October 2025 snapshot. Twelve months earlier (October 2024), the equivalent gap was approximately 100 Elo (GPT-4o vs Llama 3.1 405B). Two years earlier (October 2023), the gap was greater than 200 Elo.
Closed-API
(top 6)GPT-5
OpenAI (2025-08)
1410
Elo
Gemini 2.5 Pro
Google DeepMind (2025-03)
1397
Elo
Claude Opus 4.5
Anthropic (2025-09)
1388
Elo
Grok 4
SpaceXAI (2025-07)
1380
Elo
OpenAI o3
OpenAI (2025-04)
1374
Elo
Claude Sonnet 4.5
Anthropic (2025-09)
1370
Elo
Open-Weights
(top 6)DeepSeek-R1
DeepSeek (2025-01)
1357
Elo
MIT (model weights); DeepSeek License (commercial use allowed)
Qwen3 235B
Alibaba (2025-04)
1331
Elo
Apache-2.0
GLM-4.5
Zhipu (Z.ai) (2025-07)
1325
Elo
MIT (open-weights tier)
Llama 4 Maverick
Meta (2025-04)
1271
Elo
Llama 4 Community License (commercial use under 700M MAU)
Qwen 2.5 Max
Alibaba (2025-01)
1253
Elo
API-only for Qwen 2.5 Max; Qwen 2.5 72B is Apache-2.0
DeepSeek-V3
DeepSeek (2024-12)
1247
Elo
MIT
Inflection
DeepSeek-R1 release. First open-weights reasoning model to enter the top 10 of the Arena leaderboard. Triggered an approximate $600B selloff in NVIDIA on 2025-01-27 ('China AI shock'). (Bloomberg 2025-01-27)
Closing-Gap Thesis
- Open-weights models reached the top tier of the Arena leaderboard in early 2025 (DeepSeek-R1) and held position through October 2025.
- The capability gap shrank from approximately 200 Elo (early 2024) to approximately 50 Elo (October 2025), an approximately 75 percent reduction in 18 months.
- Pricing pressure: open-weights models are typically priced 5x to 20x cheaper than closed flagships at the same Arena Elo (DeepSeek-R1 at $0.55 / $2.19 vs GPT-5 at $1.25 / $10.00).
- Implication: substitutability is rising. Enterprises increasingly deploy open weights on owned infrastructure or via neoclouds (Together, Fireworks) for cost reasons. Closed-lab pricing power remains in extended-thinking, multimodal, and agent-tool surfaces.
- Cross-reference: this dynamic is one driver of the M&A and consolidation thesis covered in Tab 6 (Capital and M&A) of the AI Software sector.
