US AI models in 2026: GPT, Claude, Gemini and the open ones you can run

Part of our AI models guide: 11 posts on choosing and running AI models.
From India or the Gulf, the newest GPT, Claude and Gemini models mostly process your prompts abroad, and you cannot download them to run closer to home: their makers keep the weights. They also top the independent tests: American labs hold the eleven highest places on Artificial Analysis's leaderboard, led by Anthropic's Claude Fable 5.2, OpenAI's GPT-6.2 Astra and Google's Gemini 4 Neon. You rent them by the token, from the lab or a big cloud. US labs also publish models you can run on your own GPUs, such as OpenAI's gpt-oss, Google's Gemma 4 and Meta's Muse Spark Open, many under the permissive Apache 2.0 licence. The best of these, Thinking Machines' Inkling 2 Small, scores 31 on the same leaderboard, behind China's Kimi K3.1 Pro at 46.
A token is the smallest unit of text an AI model reads and writes, often a short word or part of one. An API is a way for your software to send the model questions over the internet and get answers back, billed per token. A closed-weight model is one whose maker keeps the trained model files (the weights), so you can only use it through the maker's or a cloud's service. An open-weight model is one whose weights you can download and run yourself, under a licence. In restaurant terms, closed is a meal you eat there and open-weight is a dish you take home (see open-weight vs closed AI models, explained). Open source is stricter: the maker also publishes the training code and details of the training data, as Ai2's Olmo models do. Figures are as of 19 November 2026; prices are list prices in US dollars per million tokens, input then output.
US AI models at a glance
Selected closed US models, ranked by Artificial Analysis's Intelligence Index v5 (AA v5), an independent score that combines many tests into one number, like an overall exam grade. All are API only.
Model (2026 release) | Maker | Price, input / output | AA v5 | Where to buy |
|---|---|---|---|---|
Claude Fable 5.2 (9 Nov) | Anthropic | $8 / $40 | 65 | Anthropic API, Amazon Bedrock, Google Cloud, Microsoft Foundry |
GPT-6.2 Astra (17 Nov) | OpenAI | $8 / $40 | 62 | OpenAI API, Amazon Bedrock |
Gemini 4 Neon (18 Nov) | $3 / $15 | 61 | Gemini API (not yet listed on Agent Platform) | |
Gemini 4 Carbon (10 Nov) | $2 / $10; $4 / $20 from 1 Dec | 60 | Gemini API, Agent Platform | |
Anthropic | $4 / $20 | 59 | Anthropic API, Bedrock, Google Cloud, Foundry | |
GPT-6.2 Sol (17 Nov) | OpenAI | $2 / $10 | 57 | OpenAI API, Bedrock |
Anthropic | $2 / $10 | 55 | Anthropic API, Bedrock, Google Cloud, Foundry | |
$1.50 / $7.50 | 54 | Gemini API, Agent Platform | ||
OpenAI | $1 / $5 | 51 | OpenAI API, Bedrock, Foundry | |
$2 / $10; $4 / $20 from 1 Dec | 50 | Gemini API, Agent Platform | ||
Muse Wonder (12 Oct) | Meta | $4 / $20 | 44 | Meta Model API only |
Anthropic | $0.10 / $0.50 (prompts up to 100K tokens) | 36 | Anthropic API, Bedrock, Google Cloud, Foundry | |
Meta | $0.10 / $0.50 (prompts up to 200K tokens) | 36 | Meta Model API only | |
xAI (now SpaceXAI) | $2 / $6 | 35 | xAI API, Foundry, Bedrock | |
OpenAI | $0.10 / $0.50 | 28 | OpenAI API, Bedrock, Foundry, Google Cloud (private preview) | |
Amazon | See AWS pricing | Not scored | Amazon Bedrock |
Scores are for the reasoning setting Artificial Analysis lists, mostly each model's highest. "Agent Platform" is Google's Gemini Enterprise Agent Platform (formerly Vertex AI), which is also what "Google Cloud" means for Claude. Context, how much text a model reads at once, is 1M tokens for Claude, Gemini 4, Muse Wonder, Muse Ember and Nova 2 Lite, 1.05M for GPT and 500K for Grok 4.7. The eleventh US place, Muse Wonder (Full) at 50, is a stronger Muse Wonder offered only in Meta's Muse Code, not the API.
GPT vs Claude vs Gemini on price
The top two cost the same, $8 / $40, three points apart. Value sits in the middle: Gemini 4 Hydrogen (54) and Claude Sonnet 5.5 (55) score about ten points below Fable 5.2 for a quarter of its price or less. All three labs take 50% off batch work that can wait.
Price per token is like fuel price per litre: what counts is how far a task gets on it, and models that reason at length burn more tokens. See cost per token, explained, and test models on your own tasks first.
How to use GPT, Claude and Gemini
You cannot install them: OpenAI, Anthropic and Google keep the weights. There are three ways in:
- The lab's own API, paid per token.
- A hyperscaler, such as Amazon Bedrock, Microsoft Foundry or Google Cloud's Agent Platform.
- A maker-supplied system on site. On 14 July 2026 Google announced Gemini on Google Distributed Cloud inside Indian data centres: prompts, weights and outputs stay inside the customer's perimeter, disconnected from the public internet.
Between closed and open there are more shades:
Kind | US example | What you get |
|---|---|---|
Closed, vetted access | Claude Mythos 5.2 | Fable 5.2 with fewer cyber and biology safeguards, only for organisations Anthropic has verified |
Licensed private weights | Microsoft MAI Licensed Weights (announced 17 Nov, "coming soon") | Microsoft models in your own Azure environment, through Microsoft's Frontier Company engagements; not a public download |
Gated open-weight | Meta Muse Spark Open | Download once Meta approves you; display rule and revenue conditions |
Open-weight | gpt-oss, Gemma 4, Nemotron, Inkling 2 Small | Download and run under a licence |
Open source | Ai2 Olmo 3.1 | Weights plus training data, code and evaluations, under Apache 2.0 |
Using them from India, the Gulf or Europe
Anthropic's API processes only in the US or globally, and OpenAI's keeps its newest models in-region only in the EU. Outside Europe, keeping processing local mostly means a hyperscaler, where the newest models arrive last. "Global" means any region with spare capacity; "In-Region" keeps processing where you call from; a Geo profile, Data Zone or multi-region keeps it within an area such as the EU.
Route | India | Gulf (UAE) | Europe and the UK |
|---|---|---|---|
Anthropic API (Claude) | Global or US | Global or US | Global or US (US-only costs 1.1 times) |
Claude Opus 5.5 on Bedrock | Global only, when called from Mumbai or Hyderabad | Global only, when called from the UAE or Bahrain | In-Region in London; EU Geo profile |
Claude on Agent Platform | Not listed for Mumbai | No Gulf region listed | EU multi-region (excludes the UK and Switzerland) |
Claude on Microsoft Foundry | Global only | Global only | Global only (the only data zone is the US) |
OpenAI API | In-country processing for GPT-6 Luna only | Some GPT-5.x models, through OpenAI sales | UK: storage only, not processing; EU: processing, GPT-6.2 included |
GPT-6.1 Terra on Foundry | APAC Data Zone, not India-only (+20%) | No Gulf zone | EU Data Zone (+20%) |
Gemini 4 on Agent Platform | Not in Mumbai (older Flash models are) | No Gulf region listed | EU multi-region for Argon, Carbon and Hydrogen; London: Argon only |
Older models sometimes do better. On Bedrock, Claude Opus 5, Sonnet 5 and GPT-5.6 Terra and Luna have an India Geo profile that keeps processing between Mumbai and Hyderabad, and OpenAI's open-weight gpt-oss-120b runs In-Region in Mumbai. GPT-6.1 Terra is not offered on Bedrock from the UAE or Bahrain. Microsoft says deployments kept inside one country "arrive last, have no guaranteed availability date". These listings change weekly, so check on the day you sign.
What happens to your data
Route | Trains on your API data? | Keeps prompts and outputs? |
|---|---|---|
OpenAI API | No, unless you opt in | Abuse-monitoring logs for up to 30 days; zero retention needs OpenAI's approval |
Anthropic API | Not without your express permission | Not retained by default; Fable and Mythos 5.x require 30 days |
Amazon Bedrock | Model makers have no access to your prompts or outputs | Claude Fable 5.x: kept inside AWS for up to 30 days for AWS review |
Google Agent Platform | Not without your prior permission | Cache deleted within 24 hours; OpenAI models logged up to 30 days in the US by default |
Gemini API, paid tier | No; free-tier data is used to improve Google's products | Logged for a limited period to detect misuse |
One catch: with Global routing on Bedrock, retained data stays where the request was processed, possibly outside India.
US models you can download
Parameters are the numbers a model learned in training (B means billion). Most of these models use only a small active share per token, which makes them faster, though all must fit in GPU memory; dense models use them all. 4-bit (NVFP4) is a compressed format needing about a quarter of the full 16-bit version's memory.
Model (maker) | Released | Total / active parameters | Licence | Smallest setup (from the card unless marked) | AA v5 |
|---|---|---|---|---|---|
Inkling 2 Small (Thinking Machines) | Weights about 12 Nov 2026 (reported) | 594B / 17B | Apache 2.0 (card tag; no LICENSE file) | 4x H200 or 2x B300 at 4-bit (NVFP4) | 31 |
Muse Spark Open (Meta) | 15 Oct 2026 | About 816B / 42B | Muse Model License, gated | At least 700 GB of GPU memory at 4-bit | 25 |
Beam 1.0 (Reflection AI) | Oct 2026 | 501B / 23B | Apache 2.0 | No GPU count given | 18 |
Nemotron 3.5 Super VL (NVIDIA) | 15 Oct 2026 | 120B / 12B | OpenMDW-1.1 | One H200 at 4-bit | 14 |
Nemotron 3 Ultra (NVIDIA) | 4 Jun 2026 | 550B / 55B | OpenMDW-1.1 | One 8x B200 in BF16; the card also lists 8x H200 or 16x H100 (TensorRT-LLM: Blackwell only) | 12 |
Granite 4.2 30B (IBM) | Aug 2026 | 30B, dense | Apache 2.0 | One 80 GB GPU (our arithmetic) | 6 |
Muse Glimmer-30B (Meta) | Aug 2026 | 29.6B, dense | Apache 2.0 | A 24 to 32 GB GPU | 6 |
gpt-oss-120b (OpenAI) | Aug 2025 | 117B / 5.1B | Apache 2.0 | One 80 GB GPU | 4 |
Gemma 4 31B (Google) | 2 Apr 2026 | About 31B, dense | Apache 2.0 | One 80 GB GPU (our arithmetic) | 4 |
Olmo 3.1 32B (Ai2) | Dec 2025 | 32B, dense | Apache 2.0, open source | One 80 GB GPU (our arithmetic) | Not scored |
Version 5 rebuilt the scale, so small and older models score low; never compare these with earlier versions or with figures some cards quote. GPU counts and licences for more models are in our GPU sizing guide.
Most of these are Apache 2.0, which allows commercial use if you keep the licence notice. NVIDIA's newer Nemotron models use OpenMDW-1.1, also permissive, with notice and litigation clauses. Muse Spark Open is the restrictive one. Downloads need Meta's manual approval, and products must display "Built with Muse Spark Open". A business selling models as a service needs a separate agreement with Meta once its group revenue tops $20M in any 12 months. Internal use and features built into your own product are exempt. Not legal advice.
US vs Chinese open models
Model | Country | Weights | AA v5 score |
|---|---|---|---|
Claude Fable 5.2 (closed leader) | US | Closed | 65 |
Kimi K3.1 Pro | China | Open | 46 |
DeepSeek V4.1 Flash (1017) | China | Open | 33 |
GLM-5.3 | China | Open | 32 |
Inkling 2 Small | US | Open | 31 |
Muse Spark Open | US | Open | 25 |
Beam 1.0 | US | Open | 18 |
Inkling 2 Small ranks fourth on Artificial Analysis's open-weights list, behind three Chinese models. In OpenRouter's own top 10 by tokens for the week to 18 November, the only US models were GPT-6 Luna (6th) and Claude Opus 5.5 (7th). That data leans to developers and is not market share; see the most popular AI models for the wider picture, and Chinese AI models for their safety and licences.
What Washington is doing
- Open models. America's AI Action Plan (23 July 2025) has a section titled "Encourage Open-Source and Open-Weight AI". It leaves the release decision to the developer.
- Government-backed models. The Department of Energy launched its Genesis Open Models Initiative on 7 August 2026, and Ai2's $152M OMAI project (NSF and NVIDIA, August 2025) builds models for science that Ai2 describes as fully open.
- Closed frontier models. Executive Order 14409 (2 June 2026) sets up voluntary government access to covered frontier models for up to 30 days before release, and says it creates no mandatory licensing.
- Export controls. A Commerce Department (BIS) rule released on 16 November 2026 formally rescinds the 2025 AI Diffusion Rule, including its control on closed model weights; open weights were never covered. It took effect on publication in the Federal Register on 19 November 2026. It adds chip licence requirements for Indonesia, Malaysia, Singapore, Thailand and Brazil, and keeps UAE shipments licensed unless all parties are approved UAE entities. Shipping advanced chips to a data centre abroad also needs a licence if it will give more than 12 million TPP (a chip-performance measure) of computing power to companies headquartered in China, other Country Group D:5 countries or Macau, and suppliers must check the data centre's customers.
This is general information, not legal advice.
A server for gpt-oss, Nemotron or Muse
GPT, Claude and Gemini stay on their makers' servers, so we can help only with the American open models. One 8-GPU B200 server holds Nemotron 3 Ultra in BF16 or Muse Spark Open at 4-bit, or runs several copies of gpt-oss-120b. We find that server, or an H100, H200 or B300 one, or a GB300 NVL72 rack on request, through verified operators in six regions: North America, the UK, Europe, India, Asia Pacific and the Gulf. Contracts run a year or more, paid upfront. Name the US open model and your traffic, and we will match it to a server and get quotes, or see our GPU servers.
Where these figures come from
Dates below are 2026 unless shown.
- Labs (model pages, cards and licences are linked in the tables): Anthropic pricing, residency, retention, in Foundry. OpenAI pricing, GPT-6.2 release (17 Nov), residency, your data. Google pricing, Gemini API terms, Gemini in India (14 Jul). Meta pricing, Muse Wonder launch (12 Oct). xAI SpaceXAI rebrand (6 Jul).
- Hyperscalers: AWS GPT-6.2 on Bedrock (17 Nov), Opus 5.5 card, regions, data protection, retention. Microsoft GPT-6.1 Terra (26 Oct), Grok 4.7 (6 Oct), deployment types. Google Cloud Agent Platform models, residency, data governance, OpenAI models (22 Oct).
- Scores: Artificial Analysis leaderboard, open-weights ranking (Intelligence Index v5.0.1) and v5 launch note (27 Oct).
- Usage: OpenRouter rankings, week to 18 Nov (OpenRouter's own data).
- Policy: AI Action Plan (23 Jul 2025), DOE Genesis (7 Aug), Ai2 OMAI (Aug 2025), EO 14409 (2 Jun), Diffusion Rule summary (Morrison Foerster, Feb 2025), BIS press release (16 Nov), summaries by Baker McKenzie and Fenwick.