Stage 8 of 9
Tokens models
Throughput per GPU and price per million tokens decide how the GPU-hour dollar splits between owner and lab.
- Unit
- tokens
- Sells
- model output sold by the million tokens, trained and served on rented or owned GPU-hours
- Buys from
- GPU-hours
- Lead time
- 0.5 years
- Status
- skeleton
Primer
What this stage is
Model output, sold by the million tokens. Labs train models on very large clusters and then serve them, and serving is where the volume is. Its unit is the token, and the two numbers that describe the stage are how many tokens a GPU produces in an hour and what a million of them sells for.
How it works
Throughput per GPU depends on the model's size, the hardware generation and the serving software, and it has risen far faster than hardware alone would explain. Price per million tokens has fallen faster still, driven by competition and by cheaper serving. Training is a one-off cost per model generation, large and lumpy; inference is a running cost that scales with use. Public price pages give the selling price; vendor benchmarks give throughput; the two together give revenue per GPU-hour, which can be set against the cost of that hour from the stages below.
Lab revenue and token volumes are disclosed irregularly and often through the press, so this stage has the widest error bars in the chain.
What to watch
- Tokens per second per GPU for a named model on a named chip.
- Price per million tokens for the leading models, input and output separately.
- Tokens processed per day where a platform discloses it.
- Training run size and cost by generation.
- Lab revenue run-rates, with the sourcing caveat attached.
How it connects
Tokens per GPU-hour times price per million tokens is what the GPU-hour earns in the hands of a lab. Applications above decide how many tokens the world wants and what it will pay.
Figures
| Metric | Value | Unit | As of | Scope | Confidence | Source | Note |
|---|---|---|---|---|---|---|---|
| google_tokens_processed_per_month | 3,200 | trillion tokens per month (3.2 quadrillion) | 2026-05-19 | Google (all surfaces) | high | Google, 'Google I/O 2026: Sundar Pichai's opening keynote' (blog.google, 19 May 2026) | As printed: 3.2e+15 tokens per month (3.2 quadrillion = 3,200 trillion). Monthly token volume across Google's products and APIs as stated by the CEO at I/O 2026; up from 480 trillion a year earlier and 9.7 trillion two years earlier, so about 7x in one year and about 330x in two. |
| google_tokens_processed_per_month_prior_year | 480 | trillion tokens per month | 2025-05 | Google (all surfaces) | high | Google, 'Google I/O 2026: Sundar Pichai's opening keynote' (blog.google, 19 May 2026) | As printed: 4.8e+14 tokens per month (480 trillion). The May 2025 comparison point Google gave at I/O 2026; kept as a separate row so the one-year growth (480 trillion to 3.2 quadrillion, about 6.7x) can be computed from the source's own numbers. |
| google_model_api_tokens_per_minute | 22 | billion tokens per minute through Google's model APIs | 2026-07-22 | Google (Gemini API / Vertex AI) | high | Alphabet, 'Alphabet earnings call Q2 2026: Sundar Pichai remarks' (blog.google, 22 Jul 2026) | As printed: 2.2e+10 tokens per minute through Google's model APIs (22 billion). API-only throughput (excludes Search, YouTube and other first-party surfaces), up from 16 billion per minute one quarter earlier (about 38% quarter-on-quarter). Arithmetic: 22e9 x 60 x 24 x 30 = about 950 trillion tokens per month through the APIs alone. |
| inference_price_decline_per_year_low | 9 | x decline per year in price for a fixed benchmark performance level (low end of range) | 2025-03-12 | global (Epoch AI benchmark sample) | medium | Epoch AI, data insight 'LLM inference prices have fallen rapidly but unequally across tasks' (12 Mar 2025) | Epoch measured the cheapest price to reach a fixed score on six benchmarks over three years; the decline rate depends on the benchmark, from 9x per year to 900x per year, with GPT-4-level science questions falling 40x per year. Fastest drops were in the most recent year and may not persist. |
| inference_price_decline_per_year_high | 900 | x decline per year in price for a fixed benchmark performance level (high end of range) | 2025-03-12 | global (Epoch AI benchmark sample) | medium | Epoch AI, data insight 'LLM inference prices have fallen rapidly but unequally across tasks' (12 Mar 2025) | High end of the same Epoch range; the same page gives 40x per year for GPT-4-level performance on PhD-level science questions as a central example. |
| frontier_training_compute_growth_per_year | 4.1 | x growth per year in training compute of notable models | 2024-05-28 | global (Epoch AI, 333 models 2010-May 2024) | medium | Epoch AI, 'Training compute of frontier AI models grows by 4-5x per year' (28 May 2024) | Trend fit across Epoch's database of compute estimates; the same report puts GPT-4 at 2e25 FLOP and Gemini Ultra at 5e25 FLOP, the two largest runs at the time. A 4x/year trend implies about 16x more compute every two years. |
| frontier_training_cost_growth_per_year | 2.4 | x growth per year in amortized hardware plus energy cost of the final training run | 2024-06-03 | global (Epoch AI frontier-model sample) | medium | Epoch AI, 'How much does it cost to train frontier AI models?' (3 Jun 2024) | Cost of the final training run only (amortized hardware plus electricity), excluding R&D staff and failed runs; Epoch projects the largest runs pass $1 billion by 2027. Hardware is 47-67% of total development cost, staff 29-49%, energy 2-6%. |
| gemini_3_1_pro_input_price_per_mtok | 2 | USD per 1M input tokens (prompts up to 200k tokens, paid tier) | 2026-09-27 | Google Gemini 3.1 Pro Preview | high | Google AI for Developers, 'Gemini Developer API Pricing' (page as fetched 27 Sep 2026) | List price for Google's flagship reasoning model; output is $12.00 per 1M tokens at the same prompt length, and both roughly double for prompts over 200k tokens ($4.00 in, $18.00 out). Gemini 2.5 Pro remains listed at $1.25 in / $10.00 out. |
| gemini_3_1_pro_output_price_per_mtok | 12 | USD per 1M output tokens (prompts up to 200k tokens, paid tier) | 2026-09-27 | Google Gemini 3.1 Pro Preview | high | Google AI for Developers, 'Gemini Developer API Pricing' (page as fetched 27 Sep 2026) | Output tokens (including thinking tokens) cost 6x the input price, the usual ratio for frontier reasoning models. Output is $18.00 per 1M for prompts over 200k tokens. |
| gemini_flash_input_price_per_mtok | 0.75 | USD per 1M input tokens (paid tier, promotional through 31 Dec 2026) | 2026-09-27 | Google Gemini 3.6/3.7/3.8 Flash | high | Google AI for Developers, 'Gemini Developer API Pricing' (page as fetched 27 Sep 2026) | Google's mid-tier Flash models are priced at $0.75 in / $3.75 out through the end of 2026 and are scheduled to double to $1.50 in / $7.50 out from 1 Jan 2027, a rare published price increase; the older Gemini 2.5 Flash is $0.30 in / $2.50 out. |
| grok_4_7_input_price_per_mtok | 2 | USD per 1M input tokens (prompts under 200k tokens) | 2026-09-27 | xAI Grok 4.7 | high | xAI, 'Models and Pricing' developer documentation (docs.x.ai, page as fetched 27 Sep 2026) | xAI's current flagship is $2.00 in / $6.00 out per 1M tokens under 200k context and $4.00 / $12.00 above it; the cheaper Grok 4.3 and 4.20 tiers are $1.25 in / $2.50 out. Output is 3x input, a lower ratio than Google or OpenAI. |
| deepseek_v4_pro_output_price_per_mtok | 3.96 | USD per 1M output tokens (peak-hour rate; off-peak is half) | 2026-09-27 | DeepSeek deepseek-v4-pro (DeepSeek-V4-Pro-0813) | high | DeepSeek, 'Models & Pricing' API documentation (api-docs.deepseek.com, page as fetched 27 Sep 2026) | DeepSeek's top open-weight model lists $1.32 per 1M input tokens (cache miss), $0.044 (cache hit) and $3.96 output at peak hours, with all rates halved off-peak ($0.66 in / $1.98 out); this is about one-third of Gemini 3.1 Pro's output price and shows the open-weight price floor. |
| grok_4_training_cost_usd | 490 | USD million, amortized hardware plus energy cost of the training run (Epoch median estimate) | 2025-09-12 | xAI Grok 4 (released Jul 2025) | medium | Epoch AI, data insight 'What did it take to train Grok 4?' (12 Sep 2025, data updated 17 Nov 2025) | As printed: 4.9e+08 USD, amortized hardware plus energy cost of the training run (Epoch median estimate). Epoch's estimate for the most expensive publicly analysed training run: 246 million H100-hours, 310 GWh of electricity and $490 million; Epoch's cost-trend table lists the same run at 5.0e26 FLOP and $500M in 2023 dollars, versus GPT-4 at about $40M and Llama 3.1 405B at about $50M. Epoch says the FLOP and GPU-hour inputs rest on vague xAI statements, so uncertainty is large; 2026 frontier runs (GPT-6, Gemini 3.x, Claude 5.x) have no public estimates yet. |
| openai_revenue_run_rate | 40 | USD bn annualized revenue run rate (more than $40 billion) | 2026-08-14 | OpenAI | low | Semafor, 'OpenAI annual revenue set to top $40 billion' (14 Aug 2026), citing Bloomberg | As printed: 4e+10 USD annualized revenue run rate (more than $40 billion). Bloomberg figure from people familiar with the matter, relayed by Semafor (the Bloomberg original is paywalled); the last company statement is CFO Sarah Friar's blog post that annualized revenue passed $20 billion in 2025, up from $6 billion in 2024, while OpenAI's compute rose from 0.6 GW to 1.9 GW (Reuters, 19 Jan 2026), so the $40 billion implies a doubling in about eight months. |
| anthropic_revenue_run_rate | 65 | USD bn annualized revenue run rate (more than $65 billion) | 2026-07-31 | Anthropic | medium | Quartz via Yahoo Finance, 'Anthropic revenue run rate hits $65 billion ahead of IPO' (18 Aug 2026), citing Bloomberg on an investor update | As printed: 6.5e+10 USD annualized revenue run rate (more than $65 billion). Figure Anthropic gave investors in a regular update, reported by Bloomberg and relayed by Quartz/TechCrunch/CNBC; the same report gives preliminary Q2 2026 revenue above $11.5 billion (from $787 million a year earlier) and positive adjusted operating income. Anthropic's own newsroom stated run-rate revenue 'surpassed $30 billion' on 6 Apr 2026, up from about $9 billion at end-2025, which anchors the trajectory at high confidence. |
| openai_gpt6_astra_output_price_per_mtok | 50 | USD per 1M output tokens (standard tier, prompts under the short-context threshold) | 2026-09-27 | OpenAI gpt-6-astra | high | OpenAI, 'Pricing' developer documentation (developers.openai.com, page as fetched 27 Sep 2026) | OpenAI's top flagship lists $10.00 per 1M input tokens, $1.00 cached input and $50.00 output (long-context prompts: $20.00 in / $75.00 out); the mid-tier gpt-6-sol is $2.00 in / $10.00 out and gpt-6-luna $0.10 in / $0.50 out. Batch and Flex tiers are about half these rates. Quote is the table row: columns are input, cached input, cache writes, output for short then long context. |
| claude_fable_5_1_output_price_per_mtok | 50 | USD per 1M output tokens | 2026-09-27 | Anthropic Claude Fable 5.1 | high | Anthropic, 'Pricing' page (claude.com/pricing, as fetched 27 Sep 2026) | Anthropic's top tier lists $10 per 1M input tokens and $50 output (cache read $0.25, cache write $12.50), identical to OpenAI's gpt-6-astra; the mainstream Opus 5.5 is $4 in / $20 out, Sonnet 5 $2 in / $10 out and Haiku 4.5 $1 in / $5 out, with batch processing discounted 50%. Quote is the contiguous page text of the Fable 5.1 pricing card. |
| grok_4_7_output_price_per_mtok | 6 | USD per 1M output tokens (prompts under 200k tokens) | 2026-09-27 | xAI Grok 4.7 | high | xAI, 'Models and Pricing' developer documentation (docs.x.ai, page as fetched 27 Sep 2026) | xAI's current flagship is $2.00 in / $0.50 cached / $6.00 out per 1M tokens under 200k prompt tokens and $4.00 / $1.00 / $12.00 at or above it; output is 3x input, a lower ratio than OpenAI's or Google's 5-6x. Quote is the table row (model, rate limit, input, cached input, output). |
| open_weight_share_of_openrouter_tokens | 33 | percent of OpenRouter token volume served by open-source/open-weight models (approximately one-third) | 2025-11 | OpenRouter (100 trillion tokens, Nov 2024-Nov 2025) | medium | OpenRouter and a16z, 'State of AI: An Empirical 100 Trillion Token Study' (Dec 2025) | OpenRouter routes developer API traffic across hundreds of models, so it over-weights price-sensitive developers relative to enterprise contracts; Chinese open models averaged about 13% of weekly tokens over the study year, concentrated in late 2025. A Bloomberg chart of OpenRouter data reported by OfficeChai (26 Jun 2026) says Google, OpenAI and Anthropic together fell from about 70% of OpenRouter tokens in June 2025 to about 30% in June 2026 (low confidence, secondary). No lab publishes an open-versus-closed share of total industry tokens. |
Conversions
| From → To | Factor | Unit | As of | Confidence | Source | Note |
|---|---|---|---|---|---|---|
| Applications and end demand → Tokens | 13 | USD of API tokens per developer per active day, Claude Code enterprise average | 2026-09-27 | high | Anthropic, Claude Code documentation, 'Manage costs effectively' (accessed 27 Sep 2026) | Anthropic's own average for enterprise deployments billed per token, with 90 percent of users under $30 per active day and $150-250 per developer per month. Illustrative tokens (model mix is not published): at Sonnet 5 list prices of $2 per million input and $10 per million output (claude.com/pricing), $13 buys 1.3 million output tokens or 6.5 million uncached input tokens, so a typical coding seat is in the low millions of tokens per active day, the heavy-tier band in consumption_tiers.csv. |
| Applications and end demand → Tokens | 0.1 | USD per Agentforce agent action (list, via Flex Credits) | 2026-09-27 | high | Salesforce, 'Agentforce Pricing' page (accessed 27 Sep 2026) | Arithmetic: Flex Credits cost '$500 USD/Per 100k Credits' = $0.005 per credit; an Agentforce action consumes 20 credits, so 20 x $0.005 = $0.10 per action (voice actions are 30 credits = $0.15). Together with $2 per conversation this is the public price of a unit of enterprise agent work, which the 3.2 billion Agentic Work Units per quarter can be multiplied against. |
| Applications and end demand → Tokens | 0.23 | USD per second of generated video, Runway Gen-4.5 on the Standard plan | 2026-09-27 | high | Runway, 'Pricing' page (accessed 27 Sep 2026) | Arithmetic: Standard plan $12 per month for 625 credits = $0.0192 per credit; Gen-4.5 burns 12 credits per second, so 12 x $0.0192 = $0.2304 per second (about 52 seconds of video per month). On the Max plan ($76 for 9,500 credits, $0.008 per credit) the same second costs $0.096. This is what a consumer pays; the compute cost behind it is not disclosed. |
| Applications and end demand → Tokens | 0.4 | USD per second of generated video with audio, Google Veo 3.1 Standard (720p and 1080p), API list price | 2026-09-27 | high | Google AI for Developers, 'Gemini Developer API Pricing' page, Veo 3.1 section (accessed 27 Sep 2026) | Per-second API price for the standard Veo 3.1 video model; the same table lists Veo 3.1 Fast at $0.10 per second (720p) and $0.60 per second for 4K standard output. Arithmetic: an 8-second clip costs 8 x $0.40 = $3.20 at standard quality, versus about 8 x $0.2304 = $1.84 on Runway's Standard plan. The quote is the table cell for the standard 720p/1080p rate; the page's row label is 'Veo 3.1 Standard video with audio price (default)'. |
| Applications and end demand → Tokens | 1,000,000,000,000 | tokens per year that define a 'trillion-token' Azure Foundry enterprise customer | 2026-07-29 | medium | Microsoft FY26 Q4 earnings call, 29 Jul 2026, Satya Nadella remarks (transcript hosted by StockAnalysis) | A threshold Microsoft uses to count its largest API customers, not an average; arithmetic: 1e12 / 365 = about 2.7 billion tokens per day per such customer, or roughly 2,700 heavy-tier developers at 1 million tokens a day each. On the FY26 Q3 call Microsoft put the count at over 300 customers on track for the year. Third-party transcript, so quote accuracy depends on the transcriber. |
| GPU-hours → Tokens | 10,000 | output tokens per second per GPU (Llama 3.3 70B, NVIDIA B200, NVFP4, at 50 tokens/s per user, InferenceMAX 1K/1K) | 2025-10-13 | medium | NVIDIA Technical Blog, 'NVIDIA Blackwell Leads on New SemiAnalysis InferenceMAX Benchmarks' (13 Oct 2025) | Vendor-run result on SemiAnalysis's open InferenceMAX suite for a 70B dense model at moderate interactivity: 10,000 tokens/s per GPU = 36 million tokens per GPU-hour (10,000 x 3,600); NVIDIA says this is more than 4x the H200, implying roughly 2,500 tokens/s per H200. Throughput falls steeply as per-user speed rises (for DeepSeek-R1 on B200 SemiAnalysis shows about 4,800 tokens/s/GPU at 42 tokens/s per user but about 500 at 100), so any single figure must carry its interactivity target. |
| GPU-hours → Tokens | 3,750 | output tokens per second per GPU (DeepSeek-R1 671B, FP4, 8x Blackwell B200 DGX, max-throughput mode) | 2025-03-18 | medium | NVIDIA Technical Blog, 'NVIDIA Blackwell Delivers World-Record DeepSeek-R1 Inference Performance' (18 Mar 2025) | Arithmetic: 30,000 tokens/s per 8-GPU DGX B200 / 8 GPUs = 3,750 tokens/s per GPU, or 13.5 million tokens per GPU-hour, for a 671B-parameter mixture-of-experts reasoning model at low interactivity; the same system serves over 250 tokens/s to one user when tuned for latency instead. Vendor benchmark at FP4, so an upper bound for production serving. |
| GPU-hours → Tokens | 0.19 | USD of rented GPU time per 1M output tokens (B200 at Lambda on-demand $6.69/GPU-hr, Llama 3.3 70B at 10,000 tokens/s per GPU) | 2026-09-27 | medium | Lambda, 'GPU cloud pricing: rent NVIDIA H100, H200, and B200' (on-demand instances table, as fetched 27 Sep 2026) | Arithmetic: 10,000 tokens/s x 3,600 s = 36,000,000 tokens per GPU-hour; $6.69 / 36 = $0.186 per 1M output tokens for GPU rental alone, before margin, idle time, input-token processing and the lower utilisation of real traffic. Against list prices of $3.96 (DeepSeek), $6 (Grok 4.7), $12 (Gemini 3.1 Pro) and $50 (gpt-6-astra, Claude Fable 5.1) per 1M output tokens, the raw GPU cost is a small fraction of the price of a frontier closed model but a large fraction of an open-weight one. Quote is the table text; the header column is PRICE/GPU/HR. |
| GPU-hours → Tokens | 0.1 | USD per 1M output tokens (DeepSeek-R1 on GB200 NVL72 at about 75 tokens/s per user; NVIDIA's own cost model) | 2025-10-13 | medium | NVIDIA Technical Blog, 'NVIDIA Blackwell Leads on New SemiAnalysis InferenceMAX Benchmarks' (13 Oct 2025) | NVIDIA's stated cost per million tokens using InferenceMAX throughput and its own GPU-hour cost assumptions (not disclosed on the page): $1.56 on H200 versus just over $0.10 on GB200 NVL72 for DeepSeek-R1 at about 75 tokens/s per user, a 15x gap; for gpt-oss-120b at 100 tokens/s per user it fell from $0.11 to $0.02 in two months of software updates. Shows how fast serving cost moves with hardware generation and software, independent of list prices. |
| Tokens → GPU-hours | 6 | FLOP of training compute per parameter per training token (the 6ND rule) | 2026-09-27 | high | Epoch AI, 'Data on AI models: documentation, estimation methods' (page as fetched 27 Sep 2026) | Standard heuristic for dense transformers: training FLOP = 6 x parameters x tokens (2 FLOP per parameter per token for the forward pass, about twice that for the backward pass). Example: Llama 3 405B x 15.6 trillion tokens x 6 = 3.79e25 FLOP, matching Meta's reported 3.8e25. For mixture-of-experts models use active rather than total parameters (Qwen3.8-Max: 95B active of 2.4T total). |
| GPU-hours → Tokens | 30,840,000 | H100 GPU-hours to train one frontier open model (Llama 3.1 405B, pretraining plus fine-tuning; Meta disclosure) | 2024-07-23 | high | Meta, 'Llama 3.1 Model Card' (llama-models repository, July 2024), training-time table | Meta's table gives 30.84M GPU-hours for the 405B model, 7.0M for 70B and 1.46M for 8B, 39.3M in total, on H100-80GB at 700 W TDP; at Lambda's on-demand $3.99 per H100-hour that is about $123 million of rented GPU time for the 405B run (30.84e6 x 3.99), versus Epoch's $50 million amortized-hardware estimate for the same run. Epoch estimates Grok 4 (2025) at 246 million H100-hours, eight times more. |
| GPU-hours → Tokens | 1,230,000,000,000,000,000 | FLOP of training compute realised per H100 GPU-hour (Llama 3 405B: 3.8e25 FLOP over 30.84M GPU-hours) | 2024-07-31 | high | Meta, 'The Llama 3 Herd of Models' (arXiv 2407.21783, 31 Jul 2024), section 1 and section 3.3 | Arithmetic: 3.8e25 FLOP / 30.84e6 GPU-hours = 1.23e18 FLOP per GPU-hour = 3.4e14 FLOP/s sustained, about 35% of an H100's 9.89e14 FLOP/s dense BF16 peak (model FLOP utilisation), on up to 16,000 H100s. Epoch's Grok 4 inputs (5.0e26 FLOP over 246M H100-hours) imply 2.0e18 per GPU-hour. Use this to turn a GPU-hour budget into a training-compute budget or vice versa. |
Players
| Company | Ticker | Role | Listed | Note |
|---|---|---|---|---|
| Alphabet (Google DeepMind) | NASDAQ:GOOGL | frontier lab with in-house TPUs and the largest disclosed token volume (Gemini) | yes | Processed over 3.2 quadrillion tokens per month across its surfaces in May 2026 (7x in one year) and about 22 billion tokens per minute through its model APIs by July 2026, served on its own TPUs; Gemini 3.1 Pro lists at $2 in / $12 out per 1M tokens, the cheapest of the four Western flagships. |
| OpenAI | private | frontier lab (GPT-6 family, ChatGPT); largest buyer of rented compute | no | CFO said annualized revenue passed $20 billion in 2025 (from $6 billion in 2024) as its compute rose from 0.6 GW to 1.9 GW; Bloomberg reported a run rate above $40 billion in Aug 2026, and OpenAI said in Apr 2026 it had secured 10 GW of US capacity three years ahead of its 2029 Stargate target. |
| Anthropic | private | frontier lab (Claude); enterprise and coding API leader | no | Told customers on 6 Apr 2026 that run-rate revenue had surpassed $30 billion (from about $9 billion at end-2025) with over 1,000 businesses each spending more than $1 million a year; Bloomberg reported the run rate above $65 billion at end-July 2026. Top model Claude Fable 5.1 lists at $10 in / $50 out per 1M tokens. |
| Meta Platforms | NASDAQ:META | largest open-weight model publisher (Llama) with the biggest self-funded compute budget | yes | Guided 2026 capital expenditures to $130-145 billion (Q2 2026 alone $31.08 billion), the largest AI build of any lab that gives models away; its Llama 3.1 405B disclosed 30.84 million H100 GPU-hours of training, the most detailed public frontier-run cost breakdown. |
| xAI (SpaceXAI, part of SpaceX) | private | frontier lab (Grok) with the largest self-built GPU clusters (Colossus) | no | SpaceX acquired xAI on 2 Feb 2026 in an all-stock deal valuing xAI at $250 billion; Epoch estimates Grok 4's training at 246 million H100-hours and $490 million, Grok 4.7 lists at $2 in / $6 out per 1M tokens, and in May 2026 it leased the whole Colossus 1 site (over 220,000 NVIDIA GPUs, 300 MW) to Anthropic. |
| DeepSeek | private | Chinese open-weight lab setting the price floor for frontier-class inference | no | deepseek-v4-pro lists $1.32 per 1M input tokens (cache miss) and $3.96 output at peak hours, halved off-peak, about one-third of Gemini 3.1 Pro's output price; owned by the hedge fund High-Flyer and a main driver of Chinese models' rising share of OpenRouter tokens. |
| Alibaba Group (Qwen) | NYSE:BABA / HKEX:9988 | largest open-weight model family (Qwen) and Chinese cloud incumbent | yes | Qwen3.8-Max (generally available 3 Aug 2026) is a 2.4-trillion-parameter mixture-of-experts model activating 95 billion parameters with a 1-million-token context, with weights scheduled for open release; Qwen3.8 Max ranked among the top five models by usage on OpenRouter in the week to 26 Sep 2026. |
| Mistral AI | private | European open-weight frontier lab | no | Raised EUR 3 billion in a Series D led by Samsung Electronics at a post-money valuation above EUR 21 billion (8 Sep 2026) to scale compute for training, positioning itself as the 'sovereign, open-weight' supplier for 125+ enterprises including Airbus, ASML and HSBC. |
| Moonshot AI (Kimi) | private | Chinese frontier lab (Kimi K3) preparing a Hong Kong IPO | no | SCMP reported (22 Jul 2026) the company was closing a round at a US$30 billion valuation, had sought up to US$50 billion, and told investors it may list in Hong Kong within six months; Kimi K3, released July 2026, is one of the open-weight models displacing US closed models on OpenRouter. |
| Cohere | private | enterprise and sovereign-deployment model lab (Canada) | no | Told investors in a February 2026 memo (obtained by CNBC) that it reached about $240 million USD of annual recurring revenue in 2025 against a $200 million target, with about 70% gross margins and a $7 billion valuation; a niche player by token volume but a benchmark for private-deployment economics. |
| Reflection AI | private | US open-weight frontier lab (pre-release) | no | Raised $2 billion at an $8 billion valuation in Oct 2025 (investors include Nvidia, Sequoia, Lightspeed, DST, Citi) to train an open-weight frontier model on tens of trillions of tokens as a US answer to DeepSeek; press reports put a later round at a $25 billion valuation (Apr 2026), and no frontier checkpoint had shipped by Sep 2026. |