Skip to main content

LLM Statistics 2026: Key Numbers, Data & Facts

LLM data on model records, parameter counts, context windows, prices, and benchmarks.

FHFinn Hillebrandt
AI Technology
LLM Statistics 2026: Key Numbers, Data & Facts
Links marked with * are affiliate links. If a purchase is made through such links, we receive a commission.

Which LLM is ahead in 2026?

In the Vals Index, Claude Sonnet 5.5 leads at 70.85%, ahead of Claude Opus 5.5 at 70.18%.

Maybe you aren't hunting for the top scorer anyway. You want to know how many language models exist, what they cost, and how much context they handle.

In this article, you'll find the key LLM statistics for 2026 on providers, prices, parameters, and context windows.

Start with the short version:

TL;DRKey Takeaways
  • Our model catalogue contains 303 records across 39 developer groups. Its newest saved release date is November 10, 2026. The catalogue includes Gemini 4 Argon, which became generally available on October 13.
  • In the Vals Index dated October 27, 2026, Claude Sonnet 5.5 scores 70.85%, Claude Opus 5.5 70.18%, and Gemini 4 Argon 69.24%. Vals Index v2.1 weights tasks in finance, coding, legal work, and tax by their share of the U.S. economy.
  • Selected API base prices are $0.1 / $0.5 for GPT-6 Luna, $0.75 / $3.75 for Gemini 3.8 Flash, and $4 / $20 for Claude Opus 5.5 per 1 million input and output tokens. Context tiers, introductory offers, and regions change the effective price.
  • Ten million tokens do not always mean a 10-million-token context window. Meta says Llama 4 Scout supports that context length; Qwen-Long processes 10-million-token documents through file upload. Gemini 4 Argon has a shared context window of 1,048,576 tokens across input, thinking, and output. Synchronous responses allow up to 65,536 output tokens; Long Decode can use more of the remaining context budget.

1. How Many Large Language Models Are There in 2026?

This catalogue contains 303 records across 39 developer groups. Gemini 4 Argon was announced on September 30 and became generally available on October 13. Qwen 3.8 Flash Next is listed separately in the open-weight directory.

According to the 2026 Stanford AI Index, industry produced more than 90% of notable frontier models in 2025.

2. The Biggest LLM Providers by Model Count

Here, the number of models in the catalogue sits next to each provider's share of token volume at businesses.

2.1 Catalogue Models by Developer

The chart shows the catalogue models by developer:

OpenAI
43 models
Alibaba
43 models
Google
29 models
Anthropic
26 models
Meta
18 models
Mistral AI
18 models
DeepSeek
14 models
Z.ai
12 models
Source: gradually.ai LLM database
gradually.ai

OpenAI has the largest record set at 43, followed by Alibaba at 43 and Google at 29.

2.2 Token Volume Share at Ramp Businesses

The chart shows each provider's share of token volume among businesses that pay through Ramp:

Anthropic
59.8%
OpenAI
37.2%
xAI
1.5%
Google
1.3%
Source: Ramp AI Index
gradually.ai

Anthropic and OpenAI together account for about 97% of token volume in this slice.

3. Parameters and Architecture: What Providers Disclose

Many providers do not publish parameter counts for their hosted frontier models. For open weights, some providers publish both total and per-token active parameters.

3.1 Total Size of Open Models

Moonshot lists Kimi K3 at 2.8 trillion total and 104 billion, plus 16 selected experts per token. DeepSeek-V3 is listed with 671 billion core parameters and 37 billion; its downloadable weights also contain a separate MTP module that is not part of this core count.

Kimi K3 MoE, 104B active
2.8T
Qwen 3.8 2.4T A95B MoE, 95B active
2.4T
DeepSeek-V4-Pro MoE, 49B active
1.6T
Mistral Large 4 MoE, 52B active
1.05T
Kimi K2.6 MoE, 32B active
1T
GLM-5.3 MoE, 40B active
744B
Mistral Large 3 MoE, 41B active
675B
DeepSeek-V3 MoE, 37B active, separate MTP module excluded
671B
Llama 4 Maverick MoE, 17B active
400B

Most examples in the chart use Mixture-of-Experts architectures. Their active parameter count is smaller than the total size. Our catalogue classifies 115 records as MoE.

Mistral Large 4 is included at the centrally recorded total size. Its launch post reports a rounded 1 trillion total parameters and the versioned model page a more precise 1.05 trillion. Both report 52 billion active parameters. The weights have been on Hugging Face since October 29, 2026.

The interactive table uses the central parameter dataset. Rows marked "estimated" are based on analyses or leaks, not provider disclosures.

Mistral's Small 4 launch post lists 6 billion active parameters per token, or 8 billion including embedding and output layers. Its Hugging Face card says 6.5 billion activated per token, and the table uses that figure. For most current frontier models, the parameter column deliberately reads "unknown":

50 of 303 models
Kimi K3.1 Pro

Moonshot AI

Parameters:
2.8T
Active:
MoE(104B active)
Type:Open Weights
Released:Oct 23, 2026

Moonshot describes the model as 3T-class. The model card on Hugging Face lists 2.8 trillion total and 104 billion active parameters. The weights are available under the own Kimi K3.1 Pro License, which attaches conditions to some commercial use.

Kimi K3

Moonshot AI

Parameters:
2.8T
Active:
MoE(104B active)
Type:Open Weights
Released:Jul 16, 2026

2.8T total, 104B active, 896 experts with 16 active experts per token, KDA hybrid linear attention, native vision, and a 1,048,576-token context window. Model weights are published.

Qwen 3.8 2.4T A95B

Alibaba

Parameters:
2.4T
Active:
MoE(95B active)
Type:Open Weights
Released:Aug 12, 2026
Qwen 3.8 Max 0902

Alibaba

Parameters:
2.4T
MoE:
MoE
Type:Proprietary
Released:Sep 2, 2026

The Qwen3.8-Max-0902 snapshot released September 2, 2026 improves coding, long-horizon autonomous development, agent collaboration, and visual analysis. Alibaba describes Qwen3.8-Max as a 2.4T-parameter MoE model. Its documentation does not publish an active parameter count for the 0902 snapshot and confirms a 1M-token context window.

Claude 3 Opus

Anthropic

Parameters:
2T*
Type:Proprietary
Released:Mar 2024

Analyst estimate: ~2T parameters

Llama 4 Behemoth

Meta

Parameters:
2T
Active:
MoE(288B active)
Type:Open Weights
Released:Apr 2025

Announced but not yet released; 16 experts, ~2T total

GPT-4

OpenAI

Parameters:
1.76T*
Active:
MoE(220B active)
Type:Proprietary
Released:Mar 2023

Leaked: 8x220B experts MoE, ~1.76T total parameters

DeepSeek-V4-Pro

DeepSeek

Parameters:
1.6T
Active:
MoE(49B active)
Type:Open Weights
Released:Apr 24, 2026

1.6T total, 49B active, 1M context, CSA+HCA attention, MIT-licensed

LongCat 2.0

Meituan

Parameters:
1.6T
Active:
MoE(48B active)
Type:Open Weights
Released:Jun 30, 2026

Open-weights MoE with 1.6T total parameters, about 48B active parameters, and a 1M-token context window.

Mistral Large 4

Mistral AI

Parameters:
1.05T
Active:
MoE(52B active)
Type:Open Weights
Released:Oct 6, 2026

API preview. Weights are announced for the end of October 2026. The launch post and the model page both list 52 billion active parameters. Mistral advertises 1M tokens of context; OpenRouter documents 524,288 tokens.

Ring 2.6 1T

inclusionAI

Parameters:
1T
Active:
MoE(63B active)
Type:Open Weights
Released:May 14, 2026
Kimi K2.5

Moonshot AI

Parameters:
1T
Active:
MoE(32B active)
Type:Open Weights
Released:Jan 2026
Kimi K2 Thinking

Moonshot AI

Parameters:
1T
Active:
MoE(32B active)
Type:Open Weights
Released:Nov 4, 2025
Kimi K2 0711

Moonshot AI

Parameters:
1T
Active:
MoE(32B active)
Type:Open Weights
Released:Jul 11, 2025
Kimi K2 0905

Moonshot AI

Parameters:
1T
Active:
MoE(32B active)
Type:Open Weights
Released:Sep 5, 2025
Kimi K2.6

Moonshot AI

Parameters:
1T
Active:
MoE(32B active)
Type:Open Weights
Released:Apr 20, 2026

1T total, 32B active, 384 experts (8 routed + 1 shared), 262K context, native INT4, Modified MIT

Kimi K2.7 Code

Moonshot AI

Parameters:
1T
Active:
MoE(32B active)
Type:Open Weights
Released:Jun 2026

1T total, 32B active, 256K context, thinking always enabled

Qwen 3.6 Max-Preview

Alibaba

Parameters:
1T*
MoE:
MoE
Type:Proprietary
Released:Apr 2026

Sparse MoE (~1T total), 262K context, integrated thinking mode, API-only

Yi-Large

01.AI

Parameters:
1T
MoE:
MoE
Type:Proprietary
Released:May 2024

MoE with ~1T total parameters

MiMo-V2.5-Pro

Xiaomi

Parameters:
1T
Active:
MoE(42B active)
Type:Open Weights
Released:Apr 27, 2026

1T total parameters, 42B active parameters, and a 1M context window.

MiMo-V2.5-Pro-UltraSpeed

Xiaomi

Parameters:
1T
Active:
MoE(42B active)
Type:Open Weights
Released:Jun 2026

Accelerated serving variant of MiMo-V2.5-Pro. Parameter counts refer to the underlying Pro model. Xiaomi's speed figure is a vendor claim, not an independent measurement.

Inkling

Thinking Machines Lab

Parameters:
975B
Active:
MoE(41B active)
Type:Open Weights
Released:Jul 15, 2026

Open-weights MoE with 975B total parameters, 41B active parameters, and a 1M-token context window.

GLM-5.2

Z.ai

Parameters:
744B
Active:
MoE(40B active)
Type:Open Weights
Released:Jun 16, 2026

Released 2026-06-16, 744B total parameters, 40B active, 1M context, MIT-licensed weights

GLM-5.3

Z.ai

Parameters:
744B
Active:
MoE(40B active)
Type:Open Weights
Released:Aug 18, 2026

Released on August 18, 2026. Same base as GLM-5.2 with 744B total and 40B active parameters. The weights are available from Z.ai on Hugging Face.

GLM-5.1

Z.ai

Parameters:
744B
Active:
MoE(40B active)
Type:Open Weights
Released:Apr 7, 2026

MIT-licensed model with 744B total parameters, 40B active, and a 200K context window.

GLM-5

Z.ai

Parameters:
744B
Active:
MoE(40B active)
Type:Open Weights
Released:Feb 2026

744B total, 40B active per token, with a 200K context window.

DeepSeek-V3 0324

DeepSeek

Parameters:
685B
Active:
MoE(37B active)
Type:Open Weights
Released:Mar 24, 2025
Mistral Large 3

Mistral AI

Parameters:
675B
Active:
MoE(41B active)
Type:Open Weights
Released:Dec 2, 2025

Sparse MoE with 675B total, 41B active

DeepSeek-V3.1 Terminus

DeepSeek

Parameters:
671B
Active:
MoE(37B active)
Type:Open Weights
Released:Sep 22, 2025
DeepSeek-V3.2 Exp

DeepSeek

Parameters:
671B
Active:
MoE(37B active)
Type:Open Weights
Released:Sep 29, 2025

671B serving main model, separate 14B MTP training module (685B shipped weights), 256 experts, 37B active per token, 128K context

DeepSeek-V3.1

DeepSeek

Parameters:
671B
Active:
MoE(37B active)
Type:Open Weights
Released:Aug 21, 2025
DeepSeek-R1 0528

DeepSeek

Parameters:
671B
Active:
MoE(37B active)
Type:Open Weights
Released:May 28, 2025
DeepSeek-V3.2

DeepSeek

Parameters:
671B
Active:
MoE(37B active)
Type:Open Weights
Released:Dec 2025

671B serving main model, separate 14B MTP training module (685B shipped weights), 256 experts, 37B active per token, 128K context

DeepSeek-V3

DeepSeek

Parameters:
671B
Active:
MoE(37B active)
Type:Open Weights
Released:Dec 26, 2024

Trained on 14.8T tokens for ~$5.6M

DeepSeek-R1

DeepSeek

Parameters:
671B
Active:
MoE(37B active)
Type:Open Weights
Released:Jan 2025

Reasoning model based on V3 architecture

DeepSeek-V4.1-Flash

DeepSeek

Parameters:
552B
MoE:
MoE
Type:Open Weights
Released:Sep 10, 2026

552B total, 8B active during input prefill and 16B during output decoding, 1M context, multimodal and MIT-licensed

DeepSeek V4.1 Flash 1017

DeepSeek

Parameters:
552B*
Active:
MoE(16B active)
Type:Open Weights
Released:Oct 17, 2026
Nemotron 3 Ultra

NVIDIA

Parameters:
550B
Active:
MoE(55B active)
Type:Open Weights
Released:Jun 4, 2026
PaLM

Google

Parameters:
540B
Type:Proprietary
Released:Apr 2022
Megatron-Turing NLG

NVIDIA

Parameters:
530B
Type:Proprietary
Released:Jan 2022
Qwen 3 Coder 480B A35B

Alibaba

Parameters:
480B
Active:
MoE(35B active)
Type:Open Weights
Released:Jul 23, 2025
MiniMax M1

MiniMax

Parameters:
456B
Active:
MoE(45.9B active)
Type:Open Weights
Released:Jun 13, 2025
MiniMax-01

MiniMax

Parameters:
456B
Active:
MoE(45.9B active)
Type:Open Weights
Released:Jan 12, 2025
MiniMax M3

MiniMax

Parameters:
428B
Active:
MoE(23B active)
Type:Open Weights
Released:Jun 1, 2026

Native multimodal model with a 1M context and open weights (Hugging Face, MiniMax community license); about 428B parameters, about 23B active (MoE).

ERNIE 4.5 VL 424B A47B

Baidu

Parameters:
424B
Active:
MoE(47B active)
Type:Open Weights
Released:Jun 30, 2025
Hermes 4 405B

Nous Research

Parameters:
405B
Type:Open Weights
Released:Aug 6, 2025
Llama 3.1 405B

Meta

Parameters:
405B
Type:Open Weights
Released:Jul 2024
Llama 4 Maverick

Meta

Parameters:
400B
Active:
MoE(17B active)
Type:Open Weights
Released:Apr 2025

128 experts, 17B active per token

Trinity Large Thinking

Arcee AI

Parameters:
398B
Active:
MoE(13B active)
Type:Open Weights
Released:Apr 1, 2026
Jamba Large 1.7

AI21 Labs

Parameters:
398B
Active:
MoE(94B active)
Type:Open Weights
Released:Jul 2025

* Estimate, not confirmed by the developer

Source: developer figures, linked per model
gradually.ai

3.2 Memory Needs for Local Models

For local models, memory matters most. All weights have to sit in RAM, even for Mixture-of-Experts models where only part of the network is active per token. The chart converts total parameters into gigabytes at 4.89 bits per weight (Q4_K_M quantization). Context cache and runtime add a few more gigabytes in practice:

Qwen 3.5 122B A10B 122 B, MoE
74.6 GB
GPT OSS 120B 117 B, MoE
71.6 GB
Gemma 4 31B 30.7 B
18.8 GB
Qwen 3.8 27B 27 B
16.5 GB
Gemma 4 26B A4B 25.2 B, MoE
15.4 GB
Mistral Small 3.2 24 B
14.7 GB
GPT OSS 20B 21 B, MoE
12.8 GB
Qwen 3.5 9B 9 B
5.5 GB
Granite 4.1 8B 8 B
4.9 GB
Source: Model cards, own calculation
gradually.ai

The vendor already ships GPT OSS in its own 4-bit format, so its file size deviates from this calculation.

4. Context Windows: Comparing Provider-Reported Limits

The chart shows the documented context window of eight current models. Llama 4 Scout leads by a wide margin at just over 10 million tokens, while most other top models sit at around one million. Grok 4.7 offers half of that:

Llama 4 Scout
10,485,760 tokens
GPT-6.1 Sol
1,050,000 tokens
Gemini 4 Argon
1,048,576 tokens
Kimi K3
1,048,576 tokens
Claude Opus 5.5
1,000,000 tokens
Claude Sonnet 5.5
1,000,000 tokens
DeepSeek-V4-Pro
1,000,000 tokens
Grok 4.7
500,000 tokens
Sources: Provider documentation
gradually.ai

The interactive table combines 344 records from the central dataset. Context limits vary by model, provider, and API endpoint:

50 of 344 models
Llama 4 Scout
10.5M
Qwen-Long
10M
Gemini 2.0 Pro
2M
Gemini 1.5 Pro
2M
Grok 4.1 Fast
2M
Grok 4 Fast
2M
GPT-6.1 Sol
1.1M
GPT-6.1 Terra
1.1M
GPT-6 Astra
1.1M
GPT-6 Sol
1.1M
GPT-6 Luna
1.1M
GPT-5.6 Sol
1.1M
GPT-5.6 Terra
1.1M
GPT-5.6 Luna
1.1M
GPT-5.5
1.1M
GPT-5.5 Pro
1.1M
GPT-5.4
1.1M
GPT-5.4 Pro
1.1M
Gemini 4 Carbon
1M
Kimi K3.1 Pro
1M
Llama 4 Maverick
1M
Gemini 4 Argon
1M
Gemini 3.8 Flash
1M
Gemini 3.7 Flash
1M
Gemini 3.6 Flash
1M
Gemini 3.5 Flash
1M
Gemini 3.1 Pro Preview
1M
Gemini 3.5 Flash-Lite
1M
Gemini 3.1 Flash-Lite
1M
Gemini 3 Flash Preview
1M
Gemini 2.5 Pro
1M
Gemini 2.5 Flash
1M
Gemini 2.5 Flash-Lite
1M
DeepSeek V4.1 Flash 1017
1M
Mistral Large 4
1M
Kimi K3
1M
Muse Wonder
1M
Muse Spark 1.3
1M
Muse Spark 1.2
1M
Laguna S 2.1
1M
Kolibri 1
1M
GPT-4.1
1M
GPT-4.1 mini
1M
GPT-4.1 nano
1M
Claude Fable 5.2
1M
Gemini 3 Pro
1M
Gemini 2.0 Flash
1M
Gemini 1.5 Flash
1M
Grok 4.3
1M
Grok 4.20 Reasoning
1M
Logarithmic axis: each gridline step is ten times the previous one.
Source: provider docs, linked per model
gradually.ai

Meta lists a supported 10.5M-token context window for Llama 4 Scout and says the model was trained with 256,000 tokens. Qwen-Long handles documents up to 10 million tokens through file upload and references. Alibaba's instructions apply only in mainland China's Beijing region.

OpenAI lists a 1,050,000-token context for GPT-6.1 Sol, while Anthropic lists 1M for Claude Opus 5.5 and 1M for Sonnet 5.5.

Mistral documents a 1-million-token limit for Large 4. The central table separately records 1,048,576 tokens for the OpenRouter endpoint, a different serving configuration. Mistral Small 4 supports 256K tokens, matching its launch announcement and model card.

According to Google's model page, Gemini 4 Argon has a shared context window of 1,048,576 tokens across input, thinking, and output. Synchronous responses allow up to 65,536 output tokens. Long Decode can use more of the remaining context budget, but input and thinking still count toward the shared limit.

For more on the model families, see our overviews of the Claude models and Gemini models.

5. What Does an LLM Cost? Prices per 1 Million Tokens

The chart compares base input prices for selected API models. Values are per 1 million tokens.

GPT-6 Luna Input up to 272,000 tokens; output price $0.50
$0.10 input
Gemini 3.8 Flash output price $3.75
$0.75 input
GPT-6.1 Sol Input up to 272,000 tokens; output price $10.00
$2.00 input
Claude Opus 5.5 output price $20.00
$4.00 input

GPT-6 Luna and GPT-6.1 Sol have higher price tiers above 272,000 input tokens. Google lists Gemini 3.8 Flash's introductory rates of $0.75 input and $3.75 output through December 31, 2026, followed by scheduled standard rates of $1.50 and $7.50 from January 1, 2027.

Alibaba lists Qwen 3.8 Flash in Frankfurt at a regional yuan rate of CNY 0.8 input and CNY 2.7 output per million tokens. For an estimate for your use, see the API cost calculator.

6. LLM Performance Head to Head

Three leaderboards measure different strengths. Vals Index scores agent tasks, the Artificial Analysis Intelligence Index rates general model performance, and Vectara's leaderboard tracks the hallucination rate when summarizing.

6.1 Vals Index for Agent Tasks

Vals Index v2.1 combines agent tasks in finance, coding, legal work, and tax, weighting each area by its share of U.S. GDP. The chart shows the snapshot from October 27, 2026. Other models sit between ranks 3 and 7 and are not shown.

Claude Sonnet 5.5 Rank 1
70.85%
Claude Opus 5.5 Rank 2
70.18%
Gemini 4 Argon Rank 3, limited access
69.24%
GPT-6.1 Sol Rank 7
63.94%
Source: Vals Index leaderboard and methodology
gradually.ai

Claude Sonnet 5.5 leads the index and Gemini 4 Argon ranks third. Google initially rolled it out only to selected cyber defense teams. Google's current model card lists the Gemini app, API, AI Studio, business offerings, Antigravity, and AI Mode as distribution channels; access depends on the specific product.

The gap between Claude Sonnet 5.5 and Claude Opus 5.5 is 0.67 points, which Vals says is within the usual measurement error of about ±0.9 points.

6.2 Artificial Analysis Intelligence Index

The Artificial Analysis Intelligence Index v5.0 combines several benchmarks into one overall score. The chart shows each model's best listed configuration, retrieved on October 29, 2026. The leaderboard lists several reasoning settings of the same model separately, so they are not repeated here.

Gemini 4 Carbon setting high
59 points
Claude Opus 5.5 setting max
58 points
GPT-6 Astra setting max
57 points
GPT-6.1 Sol setting max
55 points
Claude Sonnet 5.5 setting max
54 points
GPT-6.1 Terra setting max
51 points
Gemini 4 Argon setting high
50 points
Source: Artificial Analysis
gradually.ai

Gemini 4 Carbon leads with 59 points, and Claude Opus 5.5 follows one point behind. Vals Index and the Artificial Analysis index order the models differently because they weight different tasks.

6.3 Hallucination Rate When Summarizing

Vectara's Hallucination Leaderboard measures how often a model invents content when summarizing a document. Lower is better:

GPT-6 Sol
6.5%
DeepSeek V4 Pro
8.6%
GPT-6 Astra
8.7%
GPT-5.5
9.3%
Claude Opus 5.5
9.9%
Gemini 3.1 Pro (Preview)
10.4%
Kimi K2.6
10.8%
Gemini 4 Argon
12.1%
Source: Vectara Hallucination Leaderboard
gradually.ai

The selection shows eight models from the leaderboard, including older generations such as GPT-5.5 and Gemini 3.1 Pro. The lowest rates there belong to small models built for summarization, so the list mixes model classes. GPT-6 Sol and Gemini 4 Argon are 5.6 percentage points apart.

7. Open Source vs. Proprietary

Our catalogue classifies 135 records as proprietary and 168 as openly available. Of those, 162 are listed as open weights and 6 as open source.

For March 2026, the Stanford AI Index reports a 3.3% lead (1,503 versus 1,454 rating points) by the top closed model over the top open model on Chatbot Arena.

Six of the ten highest-ranked Arena models were closed in that snapshot. The report compares this with a 0.5% gap in August 2024. Browse open models in our open-source LLM directory.

The chart shows the license classification for the six developer groups with the most records in our catalogue. Proprietary records extend to the left, open ones to the right.

More Proprietary
More open licenses
Mistral AIhighest share of open licenses: 94%
Anthropichighest share of Proprietary: 100%
Proprietary
Open-weights
Open-source
Source: gradually.ai LLM database
gradually.ai

Anthropic has no bar on the right because all Anthropic entries in the catalogue are proprietary.

8. Knowledge Cutoff: How Current Are the Models?

A knowledge cutoff describes the training data date a provider reports for a model. This selection shows provider-reported training cutoffs for six models:

Claude Opus 5.5
June 2026
Claude Sonnet 5.5
June 2026
GPT-6 Luna
May 18, 2026
GPT-6.1 Sol
April 30, 2026
GPT-6 Astra
April 30, 2026
GPT-6 Sol
April 20, 2026

Anthropic lists June 2026 for Opus 5.5 and Sonnet 5.5. OpenAI lists May 18, 2026 for GPT-6 Luna, April 30, 2026 for GPT-6.1 Sol, April 30, 2026 for GPT-6 Astra, and April 20, 2026 for GPT-6 Sol.

9. Selected Releases and the Catalogue Timeline

The timeline shows selected releases with provider announcements. It runs from GPT-4o, DeepSeek-R1, and GPT-5 through Claude Opus 5.5 and Claude Sonnet 5.5 to Gemini 4 Argon and Mistral Large 4. OpenAI records further launches in its API changelog:

May 2024
GPT-4o
OpenAI introduces a model for real-time audio, image, and text processing.
Jan. 2025
DeepSeek-R1
DeepSeek releases model weights and a technical report under the MIT license.
Aug. 2025
GPT-5
OpenAI introduces GPT-5 as a system with fast and extended reasoning modes.
Sep. 22, 2026
Claude Opus 5.5 and GPT-6 Sol/Luna
Anthropic introduces Opus 5.5; OpenAI releases GPT-6 Sol and GPT-6 Luna.
Sep. 28, 2026
Claude Sonnet 5.5
Anthropic releases the second model in the Claude 5.5 family.
Sep. 29, 2026
GPT-6.1 Sol
OpenAI releases the API model for complex coding and professional work.
Sep. 30, 2026
Gemini 4 Argon
Google announces Gemini 4 Argon and begins a phased rollout to selected cyber defense teams. Its shared context window is 1,048,576 tokens across input, thinking, and output. Synchronous responses allow up to 65,536 output tokens. Long Decode can use more of the context budget left after input and thinking.
Oct. 6, 2026
Mistral Large 4
Mistral starts a public API preview. The weights follow on Hugging Face on October 29, 2026 under the Mistral Ecosystem Open-Weight License, a modified MIT license.

Gemini 4 Argon's saved release date is October 13, the date it became generally available according to the Gemini API release notes.

Spreading every model recorded in the gradually.ai catalog across its release month shows how releases have clustered. The darker a cell, the more models came out that month.

JanFebMarAprMayJun201912020120212022211202312202411553420258661431120264181115816JulAugSepOctNovDecTotal2019120201202102022116202312331220247183194820251411118513110202615131492125
New models in the catalog:1-34-78-1011-1415-18
Source: gradually.ai LLM database
gradually.ai

10. Model Status: Active, Deprecated, Legacy

This chart shows our status classification for 103 model records from Anthropic, Google, and OpenAI:

103models
Active3534.0%
Deprecated5048.5%
Legacy65.8%
Limited access43.9%
Pro-exclusive21.9%
API only11.0%
Preview32.9%
Open source21.9%
Source: gradually.ai LLM database
gradually.ai

Across these 103 records, the dataset classifies 35 active, 50 deprecated, 6 legacy, 4 with limited access, 2 Pro-exclusive, 1 available through API only, 3 preview, 2 open source.

11. Which LLM providers are the most popular?

The model counts above show supply. The popularity index shows which providers people actually search for and use.

In the gradually.ai AI vendor popularity index, OpenAI (97 points), Google (91 points) and Anthropic (88 points) lead. The index combines search demand for the vendor and its AI brands, usage via OpenRouter and Hugging Face and mentions in tech media.

Use the switch above the chart to change between world, EU and Germany.

Vendor popularity index

0 to 100, higher is more popular

OpenAI97
Google91
Anthropic88
DeepSeek85
Meta79
Alibaba78
Perplexity76
ElevenLabs73
MiniMax73
Z.ai73
Moonshot AI72
Microsoft70
Mistral AI70
xAI70
Stability AI69
ByteDance68
NVIDIA67
Tencent65
Black Forest Labs64
Nace.AI62
Cohere61
Skywork AI61
Xiaomi60
Amazon60
Runway60
StepFun60
Kuaishou60
Liquid AI59
IBM58
ShengShu Technology57
Mapika57
Baidu56
inclusionAI52
Upstage50
Arcee AI48
Bilibili Index Team47
AI2147
Institute of Foundation Models45
Nex AGI41

As of Nov 17, 2026 · Source: gradually.ai popularity index

12. Which LLMs are the most popular?

With language models too, a wide choice doesn't mean wide use. Here you can see which models are searched for, used and mentioned the most.

The most popular right now are Gemini 3.8 Flash (88 points), GLM-5.3 (87 points) and DeepSeek-V4.1-Flash (87 points). 448 language models have a value in the gradually.ai popularity index, which combines search demand, usage and mentions in tech media.

Use the switch above the chart to change between world, EU and Germany.

The most popular AI language models

Popularity index from 0 to 100, higher is more popular.

20 of 100 models

Gemini 3.8 Flash88
GLM-5.387
DeepSeek-V4.1-Flash87
Claude Opus 5.587
Kimi K385
GLM-5.3-Flash85
DeepSeek-V4-Flash85
GPT-6 Luna85
Grok 4.784
GPT-6 Astra84
GPT-5.6 Luna83
GPT-6.1 Sol83
Qwen 3.8 27B82
GLM-5.282
Claude Opus 581
Claude Haiku 5.581
Claude Sonnet 580
Claude Sonnet 5.580
Gemini 4 Carbon79
GPT-6.1 Terra78

Source: gradually.ai popularity index

13. How strong are the leading LLMs in the overall index?

The benchmark figures in this article come from individual tests. The overall index combines them into one value per model.

The gradually.ai overall index combines independent benchmarks into one value. Claude Fable 5.2 (130.9 points), Gemini 4 Carbon (127.0 points) and Claude Opus 5.5 (125.7 points) lead. 95 language models are ranked. 100 points is the average of a fixed reference group; 15 points more means a model wins about three of four benchmark comparisons.

Overall index (Gradually Meta-Score v6)

20 of 95 models

Claude Fable 5.2max130.9
Gemini 4 Carbonhigh127.0
Claude Opus 5.5max125.7
Claude Fable 5.1max124.1
Claude Sonnet 5.5max121.9
Gemini 4 Argonhigh121.2
GPT-6 Astramax119.6
Muse Wonder Fullmax118.4
GPT-6.1 Solmax117.5
Claude Opus 5max117.3
Muse Wondermax113.7
Claude Fable 5max112.8
Kimi K3.1 Promax112.6
Muse Ember 1.0max110.5
GPT-5.6 Solmax109.3
GPT-6 Solmax108.4
Gemini 3.7 Flashhigh107.7
DeepSeek V4.1 Flash 1017max107.1
GPT-6.1 Terramax106.5
Gemini 3.8 Flashhigh106.4

14. How often do LLMs make up answers?

More parameters and larger context windows say nothing about whether a model gets facts right. This measurement does.

Artificial Analysis uses AA-Omniscience to measure how often a model guesses wrong on knowledge questions instead of admitting it does not know. Gemini 4 Carbon (4%), Gemini 4 Argon (4%) and Command A Plus (14%) hallucinate least. The median of the 66 measured models is 39%. Each model counts with its best measured setting.

Hallucination rate

Share of wrong answers among all answers that are not correct, lower is better. Best measured setting per model.

20 of 66 models

Gemini 4 Carbon4%
Gemini 4 Argon4%
Command A Plus14%
Muse Ember 1.017%
Muse Wonder Full18%
Qwen 3.8 27B18%
Grok 4.720%
Muse Wonder21%
Solar Pro 421%
Muse Spark 1.322%
Nemotron 3 Ultra23%
Claude Haiku 4.523%
Mistral Large 425%
DeepSeek V4.1 Flash 101725%
GLM-5.226%
Granite 4.2 30B27%
Gemini 3.5 Flash-Lite27%
Qwen 3.8 Max 090227%
Muse Spark Open28%
K2 Horizon 375B-A23B29%

Measured by Artificial Analysis (AA-Omniscience)

15. How fast do LLMs answer?

Besides price and quality, speed decides whether a model suits chats and agents.

Nemotron 3.5 Lightning (207 tokens/s), Inkling Small (166 tokens/s) and Claude Haiku 5.5 (154 tokens/s) are the fastest. The median of the 99 models OpenRouter measured is 68 tokens/s of output.

Speed

20 of 99 models

Nemotron 3.5 Lightning207tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
Inkling Small166tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
Claude Haiku 5.5154tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
DeepSeek-V4.1-Flash134tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
Gemini 3.5 Flash-Lite131tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
Gemini 3.1 Flash-Lite129tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
DeepSeek V4.1 Flash 1017128tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
GPT-4.1 nano127tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
Gemini 3.5 Flash119tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
Gemini 3.6 Flash118tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
GPT OSS 120B117tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
Muse Spark 1.3115tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
Gemini 3.1 Pro Preview114tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
Mistral Small 4104tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
Ling 3.0 Flash102tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
GPT-5.4 mini100tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
GPT-5 nano98tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
Claude Sonnet 5.597tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
Grok 4.20 Reasoning96tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17
GPT-5 mini95tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-17

Output tokens per second, measured by OpenRouter

16. What do LLMs cost via API?

The per-million-token prices in this article apply per model. Here you can see how far apart they are across the market.

Via the API, Ling 3.0 Flash ($0.03), GPT OSS 20B ($0.06) and Nemotron 3.5 Lightning ($0.08) per 1M tokens are the cheapest. The median of the 102 priced language models is $1.50; the most expensive costs $20.00. All values are blended prices of three parts input to one part output, as in the LLM price index.

API prices

Blended price in US dollars per 1M tokens, 3 parts input to 1 part output

20 of 102 models

Ling 3.0 Flash0.03  $
GPT OSS 20B0.06  $
Nemotron 3.5 Lightning0.08  $
GPT-5 nano0.14  $
Llama 4 Scout0.15  $
MiMo V2.6 Flash0.18  $
MiMo-V2.50.18  $
Claude Haiku 5.50.20  $
GPT OSS 120B0.20  $
GPT-6 Luna0.20  $
Muse Ember 1.00.20  $
GLM-5.3-Flash0.24  $
Nemotron 3.5 Super VL 120B A12B0.24  $
DeepSeek V4.1 Flash 10170.26  $
DeepSeek-V4.1-Flash0.26  $
GPT-4o mini0.26  $
MiniMax M3.1 Flash0.26  $
Mistral Small 40.26  $
DeepSeek-V3.20.32  $
MiniMax M2.70.37  $

Source: gradually.ai LLM price index

17. Market Position and Conclusion

Bottom line:

For a specific use case, compare the relevant task evaluation, actual input and output rates, context limits, and access together. We also publish separate statistics on OpenAI, Anthropic, Google Gemini, Grok, and DeepSeek.

Frequently Asked Questions

Our catalogue contains 303 model records across 39 developer groups. The catalogue includes Gemini 4 Argon, which became generally available on October 13.

It depends on the ranking. Claude Sonnet 5.5 leads the Vals Index dated October 27, 2026, at 70.85%, ahead of Claude Opus 5.5 (70.18%) and Gemini 4 Argon (69.24%). Vals Index v2.1 combines tasks in finance, coding, legal work, and tax, weighted by their share of the U.S. economy. In the Artificial Analysis Intelligence Index v5.0, Gemini 4 Carbon leads with 59 points, followed by Claude Opus 5.5 with 58.

Among the selected standard prices, GPT-6 Luna costs $0.1 per million input and $0.5 per million output tokens up to 272,000 input tokens; higher tiers apply above that threshold.

Alibaba’s Frankfurt pricing table for Global deployment lists CNY 0.8 per million input and CNY 2.7 per million output tokens for Qwen3.8-Flash. Its separate Model Studio pricing table lists $0.113 and $0.382 for Global deployment. For more providers and tiers, see the LLM price index.

Our catalogue classifies 135 records as proprietary, 162 as open weights, and 6 as open source. In its 2026 report, the Stanford AI Index compares the top closed and open-weight models on Chatbot Arena. For March 2026, the top closed model led by 3.3%, with Arena ratings of 1,503 versus 1,454, a 49-point difference.

You can find more on open models and their terms of use in our article on open-source LLMs.

Up to 10 million tokens: that is what Meta says Llama 4 Scout supports, though it was trained with 256,000 tokens. Qwen-Long processes documents up to 10 million tokens through file uploads and references, not in a single prompt. Gemini 4 Argon has a shared context window of 1,048,576 tokens across input, thinking, and output. Synchronous responses allow up to 65,536 output tokens. Long Decode can use more of the remaining context budget, but input and thinking still count toward the shared limit.

Because many providers do not publish them for their hosted frontier models. For some open-weight models, vendors publish both total and active parameters. Moonshot lists 2.8 trillion total parameters and 104 billion active parameters for Kimi K3. For DeepSeek V4 Pro, the vendor lists 1.6 trillion total parameters and 49 billion active parameters.
FH

Finn Hillebrandt

AI Expert & Blogger

Finn Hillebrandt is the founder of Gradually AI, an SEO and AI expert. He helps online entrepreneurs simplify and automate their processes and marketing with AI. Finn shares his knowledge here on the blog in 50+ articles as well as through the AI Business Club.

Learn more about Finn and the team, follow Finn on LinkedIn, join his Facebook group for ChatGPT, OpenAI & AI Tools or do like 17,500+ others and subscribe to his AI Newsletter with tips, news and offers about AI tools and online business. Also visit his other blog, Blogmojo, which is about WordPress, blogging and SEO.