AI Software Foundation Models
Frontier Foundation Labs
The labs producing frontier or near-frontier AI models, with founding year, latest flagship, last raise, valuation, ARR, and weights status.
- ARR:
- $47B, as of 2026-05
- Last raise:
- $65B (2026-05), lead Altimeter Capital, Dragoneer, Greenoaks and Sequoia Capital (co-leads)
- Latest model:
- Claude Opus 5.5 / Sonnet 5.5 / Haiku 5.5 (2026-09)
- Country:
- US, founded 2021
- Backers:
- Amazon (primary cloud and training partner; up to 5 GW of capacity and a $100B+ AWS commitment), Google and Broadcom (5 GW of TPU capacity), Microsoft and NVIDIA (strategic investors), Altimeter, Dragoneer, Greenoaks, Sequoia, ICONIQ, GIC, Coatue. Claude is available on AWS, Google Cloud and Azure.
- ARR:
- $24B, as of 2026-03
- Last raise:
- $122B (2026-03), lead SoftBank (co-lead; round anchored by Amazon, NVIDIA and SoftBank)
- Latest model:
- GPT-6.1 Sol (2026-09)
- Country:
- US, founded 2015
- Backers:
- OpenAI Foundation (controls OpenAI Group PBC), Microsoft (compute and equity partner), Amazon, NVIDIA, SoftBank, a16z, MGX, TPG, T. Rowe Price. Compute also from Oracle, AWS, CoreWeave and Google Cloud.
- ARR:
- n/d
- Last raise:
- $3.5B (2026-09), lead Samsung Electronics (co-leads EQT Scaleup Europe Fund and PSG Equity)
- Latest model:
- Mistral Large 4 (2026-10)
- Country:
- FR, founded 2023
- Backers:
- Samsung Electronics, EQT (Scaleup Europe Fund), PSG Equity, ASML (Series C lead), Advent, BlackRock funds and accounts, General Catalyst, Andreessen Horowitz, Lightspeed, NVIDIA, Salesforce Ventures, Bpifrance. Distribution via Microsoft Foundry, Amazon Bedrock and Vertex AI.
- ARR:
- n/d
- Last raise:
- $500M (2025-08), lead Radical Ventures and Inovia Capital (co-leads)
- Latest model:
- Command A (2025-03)
- Country:
- Canada, founded 2019
- Backers:
- PSP Investments, Cisco, AMD, Inovia, Index Ventures, Tiger Global. Sales focus on enterprise on-prem and sovereign deployments via Oracle OCI partnership.
- ARR:
- n/d
- Last raise:
- $400M (2024-12), lead Alibaba, Tencent (strategic)
- Latest model:
- GLM-4.5 (2025-07)
- Country:
- CN, founded 2019
- Backers:
- Alibaba, Tencent, Meituan, Xiaomi, Hongshan (formerly Sequoia China), Hillhouse. Tsinghua University KEG lab spinout. Registered with the Cyberspace Administration of China (CAC).
- ARR:
- n/d
- Last raise:
- $208M (2023-11)
- Latest model:
- Jamba Reasoning 3B / Jamba 1.7 (2025-10)
- Country:
- Israel, founded 2017
- Backers:
- Walden Catalyst, NVIDIA, Intel Capital, Google, Pitango, SCB10X. Distribution via AI21 Studio plus Azure, Bedrock, Vertex.
| # | Lab | Country | Founded | Latest model | Last raise | Valuation | ARR (latest) | Open / Closed | Backers |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Anthropic | US | 2021 | Claude Opus 5.5 / Sonnet 5.5 / Haiku 5.5 2026-09 | $65B (2026-05) | $965B | $47Bas of 2026-05 | closed | Amazon (primary cloud and training partner; up to 5 GW of capacity and a $100B+ AWS commitment), Google and Broadcom (5 GW of TPU capacity), Microsoft and NVIDIA (strategic investors), Altimeter, Dragoneer, Greenoaks, Sequoia, ICONIQ, GIC, Coatue. Claude is available on AWS, Google Cloud and Azure. |
| 2 | OpenAI | US | 2015 | GPT-6.1 Sol 2026-09 | $122B (2026-03) | $852B | $24Bas of 2026-03 | closed | OpenAI Foundation (controls OpenAI Group PBC), Microsoft (compute and equity partner), Amazon, NVIDIA, SoftBank, a16z, MGX, TPG, T. Rowe Price. Compute also from Oracle, AWS, CoreWeave and Google Cloud. |
| 3 | Mistral AI | FR | 2023 | Mistral Large 4 2026-10 | $3.5B (2026-09) | $24B+ | n/d | mixed | Samsung Electronics, EQT (Scaleup Europe Fund), PSG Equity, ASML (Series C lead), Advent, BlackRock funds and accounts, General Catalyst, Andreessen Horowitz, Lightspeed, NVIDIA, Salesforce Ventures, Bpifrance. Distribution via Microsoft Foundry, Amazon Bedrock and Vertex AI. |
| 4 | Cohere | Canada | 2019 | Command A 2025-03 | $500M (2025-08) | $6.8B | n/d | mixed | PSP Investments, Cisco, AMD, Inovia, Index Ventures, Tiger Global. Sales focus on enterprise on-prem and sovereign deployments via Oracle OCI partnership. |
| 5 | 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). |
| 6 | AI21 Labs | Israel | 2017 | Jamba Reasoning 3B / Jamba 1.7 2025-10 | $208M (2023-11) | $1.4B | n/d | mixed | Walden Catalyst, NVIDIA, Intel Capital, Google, Pitango, SCB10X. Distribution via AI21 Studio plus Azure, Bedrock, Vertex. |
| 7 | 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). |
| 8 | Reka AI | US | 2022 | Reka Flash 3 2025-03 | $110M (2025-07) | $1.0B+ | n/d | mixed | DST Global, Radical Ventures, Snowflake Ventures, Nvidia. Snowflake held talks in May 2024 to acquire Reka for more than $1B; the deal did not close. Reka Flash 3 released open-weights under Apache-2.0 Mar 2025. Merged with Moonvalley in Jun 2026. |
| 9 | 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). |
| 10 | 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. |
| 11 | SpaceXAI | US | 2023 | Grok 4.7 2026-09 | $20B (2026-01) | n/a | n/d | mixed | Part of SpaceX (Nasdaq: SPCX) since the 2 Feb 2026 merger. Series E investors included Valor Equity Partners, StepStone Group, Fidelity, Qatar Investment Authority and MGX, with NVIDIA and Cisco as strategic investors. SpaceX reported 1.7 GW of nameplate AI compute at 30 Sep 2026 (Colossus II buildout). Grok models are distributed via X, the Grok apps, an API and cloud marketplaces including Amazon Bedrock and Microsoft Foundry. |
| 12 | 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). |
| 13 | DeepSeek | CN | 2023 | DeepSeek-V4 Preview (V4-Pro and V4-Flash) 2026-04 | n/a | n/a | n/d | open-weights | No funding round or valuation disclosed. DeepSeek-AI's Fire-Flyer 2 paper credits High-Flyer Quant and describes a cluster of 10,000 PCIe A100 GPUs. Models are listed on Baidu AI Cloud, ByteDance Volcano Ark and Huawei Cloud. |
| 14 | 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. |
| 15 | 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. |
Frontier Lab Revenue Trajectory
Quarterly annualized run-rate revenue for the labs that have disclosed numbers. OpenAI and Anthropic anchor the series; xAI, Mistral, and Cohere are sparse. Lines disconnect at null cells rather than drawing through them.
Milestone Disclosures
- 2025-01Anthropic. Run-rate revenue of approximately $1B at the beginning of 2025. (Anthropic (Series F announcement))
- 2025-08Anthropic. Run-rate revenue over $5B. (Anthropic (Series F announcement))
- 2026-02Anthropic. Run-rate revenue of $14B. (Anthropic (Series G announcement))
- 2026-03OpenAI. Generating $2B in revenue per month. (OpenAI ($122B funding announcement))
- 2026-04Anthropic. Run-rate revenue surpassed $30B. (Anthropic (Amazon compute post))
- 2026-05Anthropic. Run-rate revenue crossed $47B. (Anthropic (Series H announcement))
Model Capability Benchmarks
Frontier models scored across 6-10 benchmarks (MMLU-Pro, GPQA, SWE-Bench, AIME, HLE, Arena Elo, more), color-graded by band. Self-reported scores are flagged.
| Model | MMLU-Pro% | GPQA Diamond% | SWE-Bench Verified% | AIME 2025% | HLE% | Arena Elo | MMMU% | ARC-AGI v2% |
|---|---|---|---|---|---|---|---|---|
GPT-5 OpenAI extended-thinking variant (high) | 87.1 | 85.7* | 74.9* | 94.6* | 26.3 | 1,437 | 84.2* | 9.9 |
Claude Opus 4.5 Anthropic extended-thinking enabled | 89.5 | 87.0* | 80.9* | n/a | 26.3 | 1,460 | 80.7* | 37.6 |
Claude Sonnet 4.5 Anthropic extended-thinking optional | 87.5 | 83.4* | 77.2* | 87.0* | 14.1 | 1,448 | 77.8* | 13.6 |
Gemini 2.5 Pro Google DeepMind thinking on (default) | 86.2 | 86.4* | 67.2* | 88.0* | 22.1 | 1,452 | 82.0* | 4.9 |
Grok 4 SpaceXAI Grok 4 (reasoning), no tools | 86.6 | 87.5* | n/a | 91.7* | n/a | 1,409 | n/a | 16.0 |
Llama 4 Maverick Meta no extended-thinking variant | 80.5* | 69.8* | n/a | 19.3 | 5.3 | 1,327 | 73.4* | 0.0 |
DeepSeek-R1 DeepSeek reasoning model (think-aloud) | 84.0* | 71.5* | 49.2* | n/a | 8.5 | 1,395 | n/a | 1.3 |
Qwen3 235B Alibaba hybrid thinking-mode toggle | n/a | 71.1* | n/a | 81.5* | 11.8 | 1,374 | n/a | n/a |
OpenAI o3 OpenAI reasoning model (high effort) | 85.3 | 83.3* | 69.1* | 88.9* | 20.6 | 1,434 | 82.9* | 6.5 |
GLM-4.5 Zhipu (Z.ai) thinking-mode toggle | 84.6* | 79.1* | 64.2* | n/a | 9.6 | 1,408 | n/a | n/a |
API Token Pricing
Current pricing per million tokens for input, output, and cached input across the major frontier APIs, sorted by input ascending. Includes context window and pricing-as-of date. Useful for comparing per-token economics across labs at a fixed quality tier.
| Model | Provider | Input$/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 V4.1-Flash | DeepSeek | $0.30 | $1.20 | $0.006 | 1.0M | 2026-11-03 |
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 | |
DeepSeek V4-Pro | DeepSeek | $1.32 | $3.96 | $0.044 | 1.0M | 2026-11-03 |
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 |
GPT-5.6 Sol | OpenAI | $5.00 | $30.00 | $0.50 | 1.1M | 2026-07-09 |
GPT-6 Astra | OpenAI | $10.00 | $50.00 | $1.00 | 1.1M | 2026-09-03 |
Claude Opus 4.5 | Anthropic | $15.00 | $75.00 | $1.50 | 200K | 2025-09-25 |
Token Price Decline Over Time
Log-scale view of dollars per million tokens by capability tier, 2023 to late 2025, showing roughly 10x annual deflation in frontier API pricing.
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
- 2024-06-21Claude 3.5 Sonnet undercuts Opus at 1/5 price
- 2024-07-18GPT-4o mini: tier-A capability at $0.15/MTok input
- 2025-01-20DeepSeek-R1: open-weights reasoning model at $0.55 input. China selloff.
- 2025-02-08DeepSeek-V3 API at sub-$0.30 input ($0.27)
- 2025-08-07GPT-5 flagship at $1.25 input: a 24x cut from GPT-4 original in 29 months
- 2026-07-09GPT-5.6 Sol launch list $5 / $30: first OpenAI flagship price-up versus GPT-5
- 2026-08-21Sol promo $4 / $20 through at least Nov 21, 2026
- 2026-09-03GPT-6 Astra $10 / $50: price-up versus GPT-5 and versus Sol promo
Training Compute and Cost
Estimated training FLOPs (log scale) and dollar cost for frontier models from GPT-3 onward, anchored on Epoch AI data. Toggle between FLOPs and cost.
Not Disclosed (FLOPs)
Open Weights vs Closed Labs
Chatbot Arena Elo rankings of the top closed-API models versus the top open-weights models, tracking the gap that narrowed after DeepSeek-R1.
Top Closed vs Top Open Gap (Chatbot Arena Elo)
39 Elo
39 points between the top closed model (Gemini 4 Carbon, 1,526) and the top open-weights model (Kimi K3, 1,487) on the 11 November 2026 leaderboard, about a 56 percent expected win rate for the leader. All 19 entries above Kimi K3 are closed models from Google, Meta and Anthropic.
Closed-API
(top 6)Gemini 4 Carbon
Google DeepMind (2026-11)
1526
Elo
Gemini 4 Argon
Google DeepMind (2026-10)
1522
Elo
Muse Wonder
Meta (2026-10)
1511
Elo
Claude Opus 5.5
Anthropic (2026-09)
1507
Elo
Claude Fable 5
Anthropic (2026-06)
1504
Elo
Claude Opus 4.6
Anthropic (2026-02)
1503
Elo
Open-Weights
(top 6)Kimi K3
Moonshot AI (2026-07)
1487
Elo
Kimi K3 License (modified MIT): commercial use allowed; Model-as-a-Service firms above $20M revenue need a separate agreement
Kimi K3.1 Pro
Moonshot AI (2026-11)
1485
Elo
Kimi K3.1 Pro License (modified MIT): same terms as Kimi K3
GLM-5.3
Zhipu (Z.ai) (2026-08)
1482
Elo
GLM-5.3 License (MIT-style): commercial use allowed; Model-as-a-Service firms above $10B revenue need a Z.ai security review
MiMo-V2.6-Pro
Xiaomi (2026-09)
1480
Elo
MIT
GLM-5.3-Flash
Zhipu (Z.ai) (2026-08)
1475
Elo
MIT
DeepSeek-V4.1-Flash
DeepSeek (2026-09)
1475
Elo
MIT
Inflection
DeepSeek-R1 released with open weights under the MIT licence; DeepSeek described it as on par with OpenAI o1 and free to commercialize. (DeepSeek release post)
Closing-Gap Thesis
- The best open-weights model, Kimi K3, scores 39 points below the leader, about a 56 percent expected win rate for the leader in a head-to-head vote.
- Chinese labs hold all six open-weights places: Moonshot AI (Kimi K3, Kimi K3.1 Pro), Zhipu (GLM-5.3, GLM-5.3-Flash), Xiaomi (MiMo-V2.6-Pro) and DeepSeek (V4.1 Flash).
- Meta's highest-scoring model on the board, Muse Wonder (#3), is closed-API.
- Implication: for many chat workloads an open model within about 40 points of the leader is a workable substitute that can run on owned infrastructure or through neoclouds (Together, Fireworks), while the closed labs hold all of the top 19 places.
- 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.