Token Deals

EvoBioSys systems · LLM hosts by price, capability and jurisdiction · prices: models.dev · intelligence: Artificial Analysis

Token Deals

Name: Token Deals
Creator: EvoBioSys

Which model host to use for which job: an editorial trust ranking of hosts, a price-against-intelligence regression, and the best deal per tier. Hosts outside Europe only ever receive non-confidential data.

Human curation: which hosts, in which order

EvoBioSys editorial ranking and rules, applied before any price comparison. The reasons in the right-hand column are editorial choices, not sourced facts; the privacy line below each host is what its own pages say, with a link to the source. Lower rank wins a price tie.

Rank	Host	Where	May receive	Editorial note	Sourced privacy evidence

Price vs. intelligence

Each dot is one host's listing of a model. x = blended $/1M tokens (log), y = Artificial Analysis Intelligence Index, fetched through their API. The line is a least-squares fit of the score on ln(price) over the paid rows. A dot above the line gives more intelligence than the line predicts at that price. Free rows sit in the strip on the left and stay out of the fit. A model without a score is not plotted, never estimated.

Best deal per tier

Tiers combine capability and price. h is the best price-performance at the lower end, s the best at the higher end, o the best open-weight model unless one nearly as good is much cheaper. Price-performance is how far a listing sits above the regression line at its price. Values within 1.5 points count as equal; prices within 10 percent count as equal, then the higher-trust host wins. The tested fallback is the curated pick that is known to work.

	Tier rule	Tested fallback (curated)	Best deal on a confidential-ok host	Best deal on any host (non-confidential data only)	Ranked, beating the fallback

Exposure: confidential-ok the host may receive confidential data; non-confidential only it never does (non-European hosts, and any host without a data-processing agreement). A leaf marks a host that runs on renewable-powered data centres by its own account.

European model builders and computer use — checked 2026-10-07

Companies that build open or sovereign models in Europe, beyond Mistral. Only facts read on a model card or the company's own page are stated; a blank means not found, not no. Data handling: local-OK open weights you run on your own machine, nothing leaves it; sandbox a hosted API, so non-confidential data only until a no-training agreement is read. European hosting alone does not mean no training.

Builder	Models	Weights and license	Hosted API	Data handling	Status
H Company 🇫🇷 France (Paris)	Holo4-27B, Holo4-35B-A3B: computer-use agents (screenshot, click, type)	27B: CC BY-NC 4.0, non-commercial only (built on Qwen3.8-27B). 35B-A3B: Apache 2.0 (built on Qwen3.6-35B-A3B). Q4_K_M quant of the 27B: 16.9 GB. Real computer use needs H Company's agent harness.	OpenAI-compatible API at api.hcompany.ai. Billed by plan (Free, $20 or $200 a month; top-ups at model cost + 20%). No per-token price published.	local-OK sandbox API	v Specialist for computer use only, and used here for non-commercial work only. OSWorld: 27B 85.2% at $0.08 per task. OSWorld 2.0 (long workflows): 27B 61.7% at $1.22, 35B-A3B 30.9% at $0.61 per task.
Aleph Alpha 🇩🇪 Germany	Kolibri-1 (78B mixture of experts, 3.46B active); Pharia-1-LLM-7B	Kolibri-1: Apache 2.0 (weights only; code, architecture and training method excluded). Pharia-1-LLM-7B: Open Aleph License, non-commercial research and education.	No European token API found for Kolibri; its model card lists no inference provider.	local-OK weights	Not a priced row. Cohere (Canada) and Aleph Alpha signed a definitive business-combination agreement on 2026-09-16; closing awaits regulatory approval. Dual headquarters Berlin and Toronto planned. Data residency after closing: not found.
LightOn 🇫🇷 France	OriOn-Mistral-24B, OriOn-Qwen-32B, Paradigm platform	Not found on lighton.ai; only an aggregator describes the OriOn models as open-weight.	Not found.		Not usable until verified. Listed on Euronext Growth Paris since 2024-11-26.
Multiverse Computing 🇪🇸 Spain	Quasar 438B (CompactifAI compression)	Weights and license not found; one source calls it a proprietary hosted API model. A secondary report says it builds on a Chinese open model.	CompactifAI API; no price on the company site.	sandbox API	Exists (launched 2026-09-02, Artificial Analysis index 43 per press). Not a priced row.
Silo AI 🇫🇮 Finland	Poro 34B	Apache 2.0.			Excluded by default: owned by AMD (United States) since August 2024. Reconsidered only for an offering confirmed EU-hosted.

Sources: huggingface.co/HCompany (Holo4 cards), hcompany.ai and hcompany.ai/pricing, huggingface.co/Aleph-Alpha/Kolibri-1, Cohere announcement of 2026-09-16, multiversecomputing.com, lighton.ai, Euronext press release of 2024-11-26; AMD acquisition of Silo AI via press.
Not yet checked in a browser: IONOS, STACKIT and OVHcloud price pages. Their listed prices stay as the registry gives them.

Price vs. context

Each dot is one provider's listing of a model — x = blended price per 1M tokens (log), y = context window (log), fill = utility score under your current mix. A ring marks an EU/EEA or Switzerland hosting region.

low score high score EU/EEA or Switzerland hosting

Ranked ledger

Sorted by utility score under your current mix and filters. Click a column to re-sort; hover the ⓘ for jurisdiction edge-case notes.

#	Model / host	Origin	Region	Context	Prompt $/1M	Completion $/1M	Blended $/1M	Score

Token Deals: plain-text summary

What this is

Token Deals compares hosts of large language models by price, capability and jurisdiction. Data as of 2026-10-07. The copy on this page is published by hand and may lag the latest weekly refresh.
Capability is the Artificial Analysis Intelligence Index v4.3.2 from Artificial Analysis (https://artificialanalysis.ai/). Prices come from the models.dev registry. Blended price = 0.25 x input price + 0.75 x output price per 1 million tokens. Value = how far a listing sits above the least-squares line of index score on ln(blended price): score = 28.6 + 6.07 x ln(price), R-squared 0.33, 113 paid listings.

Host ranking (editorial, applied before price)

1. 🇨🇭 Infomaniak (Switzerland): Gold standard overall. Preferred when prices are equal. May receive confidential data.
2. 🇨🇭 Privatemode AI (Switzerland): Best privacy of all hosts: confidential computing. May receive confidential data.
3. 🇮🇹 Regolo AI (Italy): Third in the ranking. In the registry it is often priced above the field. May receive non-confidential data only.
4. 🇸🇪 Berget.AI (Sweden): Stable Swedish option, tested here. May receive confidential data.
6. 🇫🇷 OVHcloud AI Endpoints (France): European (French) host. May receive non-confidential data only.
6.9. 🇫🇷🇫🇮 GreenPT (France and Finland): Slightly above Scaleway in the ranking: renewable-powered data centres on French and Finnish infrastructure. May receive confidential data.
7. 🇫🇷 Mistral AI (direct) (France): French jurisdiction: lowest of the European hosts in this editorial ranking. May receive non-confidential data only.
7. 🇫🇷 Scaleway (France): French jurisdiction: lowest of the European hosts in this editorial ranking. May receive confidential data.
8. 🇸🇪 evroc (Sweden): Swedish sovereign cloud, not yet placed in the ranking. May receive confidential data.
9. 🇩🇪🇫🇮 Hetzner (Germany and Finland): Held to non-confidential data: the free API had no SLA or DPA when last checked. May receive non-confidential data only.
9. 🇩🇪 SAP AI Core (Germany): Acceptable European host. May receive non-confidential data only.
9. 🇩🇪 STACKIT (Germany): European host, treated like the other acceptable hosts. May receive non-confidential data only.
9. 🇧🇪 Umans AI (Belgium): Acceptable European host. May receive non-confidential data only.
9.5. 🇩🇪 Lyceum (Germany): Berlin-based serverless open-model API. Prices from its public pricing page (2026-10-07). It states no context windows, so each row shows the smallest window other hosts list for the same model. No retention or training statement found, so non-confidential data only. May receive non-confidential data only.
20. 🇺🇸 Anthropic (direct) (United States): Non-European: non-confidential data only. May receive non-confidential data only.
20. 🇺🇸 DeepInfra (United States (assumed)): Non-European: non-confidential data only. May receive non-confidential data only.
20. 🇺🇸 OpenAI (direct) (United States): Non-European: non-confidential data only. May receive non-confidential data only.
21. 🇺🇸 NVIDIA build (free endpoints) (United States (assumed)): Free tier, non-European: non-confidential data only. May receive non-confidential data only.
21. 🌐 OpenRouter (global router, region unclear): Router to many providers: non-confidential data only. May receive non-confidential data only.
22. 🇨🇳 DeepSeek (direct) (China): Non-European: non-confidential data only. May receive non-confidential data only.
22. 🌐 Moonshot AI (direct) (global, region unclear): Non-European: non-confidential data only. May receive non-confidential data only.
22. 🌐 Xiaomi (direct) (global, region unclear): Non-European: non-confidential data only. May receive non-confidential data only.
22. 🇨🇳 Z.ai (direct) (China): Non-European: non-confidential data only. May receive non-confidential data only.
At equal price (within 10 percent) the higher-ranked host wins. A better deal on a non-European host is used only for non-confidential data.

Best deal per tier

f: Fable through Claude Code on the Max plan (not an open model, not computed).
o (orchestrator): best open-weight model unless a near-equal one costs at most 60 percent as much.
Best on a confidential-ok host: GLM 5.3 🇨🇳 on 🇧🇪 Umans AI: index 44.8, $1.4 / $4.4 per 1M tokens in/out, may receive confidential data.
Best on any host: GLM-5.3 🇨🇳 on 🇺🇸 NVIDIA build (free endpoints): index 44.8, free / free per 1M tokens in/out, may receive non-confidential data only.
s (higher end): best price-performance at index 30 or more.
Best on a confidential-ok host: GLM-5.3-Flash 🇨🇳 on 🇫🇷🇫🇮 GreenPT: index 41.8, $0.127754 / $0.511016 per 1M tokens in/out, may receive confidential data.
Best on any host: GLM-5.3 🇨🇳 on 🇺🇸 NVIDIA build (free endpoints): index 44.8, free / free per 1M tokens in/out, may receive non-confidential data only.
h (lower end): best price-performance at index 8 to below 30.
Best on a confidential-ok host: Qwen3.8 27B 🇨🇳 on 🇮🇹 Regolo AI: index 26.2, $0.58 / $2.42 per 1M tokens in/out, may receive confidential data.
Best on any host: MiniMax-M3 🇨🇳 on 🇺🇸 NVIDIA build (free endpoints): index 29.2, free / free per 1M tokens in/out, may receive non-confidential data only.

Highest-value listings (any host)

GLM-5.3 🇨🇳 on 🇺🇸 NVIDIA build (free endpoints): index 44.8, free / free per 1M tokens in/out, may receive non-confidential data only; +40.0 points above the price line.
Kimi K3 🇨🇳 on 🇺🇸 NVIDIA build (free endpoints): index 43.6, free / free per 1M tokens in/out, may receive non-confidential data only; +38.8 points above the price line.
GLM-5.3-Flash 🇨🇳 on 🇺🇸 NVIDIA build (free endpoints): index 41.8, free / free per 1M tokens in/out, may receive non-confidential data only; +37.0 points above the price line.
DeepSeek V4.1 Flash 🇨🇳 on 🇺🇸 NVIDIA build (free endpoints): index 39.5, free / free per 1M tokens in/out, may receive non-confidential data only; +34.7 points above the price line.
DeepSeek V4 Pro 0813 🇨🇳 on 🇺🇸 NVIDIA build (free endpoints): index 36, free / free per 1M tokens in/out, may receive non-confidential data only; +31.2 points above the price line.
DeepSeek V4 Flash 0731 🇨🇳 on 🇺🇸 NVIDIA build (free endpoints): index 34.3, free / free per 1M tokens in/out, may receive non-confidential data only; +29.5 points above the price line.
GLM-5.2 🇨🇳 on 🇺🇸 NVIDIA build (free endpoints): index 33.7, free / free per 1M tokens in/out, may receive non-confidential data only; +28.9 points above the price line.
MiniMax-M3 🇨🇳 on 🇺🇸 NVIDIA build (free endpoints): index 29.2, free / free per 1M tokens in/out, may receive non-confidential data only; +24.4 points above the price line.
Kimi K2.6 🇨🇳 on 🇺🇸 NVIDIA build (free endpoints): index 27, free / free per 1M tokens in/out, may receive non-confidential data only; +22.2 points above the price line.
Qwen3.8-27B 🇨🇳 on 🇩🇪🇫🇮 Hetzner: index 26.2, free / free per 1M tokens in/out, may receive non-confidential data only; +21.4 points above the price line.

Limits

Scores are Artificial Analysis index values at the thinking level named in each row; where that level is not published, the top published variant is used, which overstates.
Privacy classes come from each host's own public statements where they were read (quotes and links on the page); 'terms not read' means nothing was verified. This is not legal or compliance advice.
A model counts as open-weight only when the weights are published. MiMo V2.6 weights are published under MIT as the MiMo-V2.6-Pro-RL and -Flash-RL checkpoints; the model card does not say the hosted API model is identical.