AI Model Rankings
Models ranked on Safeguard's own data first, then market usage from OpenRouter. Every board names its source and date.
- Safeguard model registry: data through
- Safeguard model registry and Safeguard Gold model directory: data through
Safeguard rankings
From the Safeguard model registry and Safeguard Gold. Pick a board; each model links to its page in the catalog.
Input price
Provider list price per 1M input tokens, from each provider's own pricing page.
| Rank | Model | Input / 1M tokens |
|---|---|---|
| 1 | text-embedding-3-smallby OpenAI | $0.02Checked Oct 8, 2026 |
| 2 | Command R7Bby Cohere | $0.0375Checked Oct 8, 2026 |
| 3 | Text Embedding v4by Alibaba Cloud (Qwen) | $0.07Checked Oct 8, 2026 |
| 4 | Embed 5 Fastby Cohere | $0.08Checked Oct 8, 2026 |
| 5 | Claude Haiku 5.5by Anthropic | $0.10Checked Oct 8, 2026 |
| 6 | GPT-6 Lunaby OpenAI | $0.10Checked Oct 8, 2026 |
| 7 | Embed 5 Proby Cohere | $0.12Checked Oct 8, 2026 |
| 8 | text-embedding-3-largeby OpenAI | $0.13Checked Oct 8, 2026 |
| 9 | Codestral Embedby Mistral AI | $0.15Checked Oct 8, 2026 |
| 10 | Command R (08-2024)by Cohere | $0.15Checked Oct 8, 2026 |
Each provider's own list price per 1M input tokens, read on the provider's pricing page on the date shown in each row. Safeguard adds no markup. Models with no published per-token price (priced on request, per minute, per image or per search) are not listed, and models that use no tokens are on the token-free board.
Source: Safeguard model registry (safeguard.sh/models), as of 2026-10-08. All rights reserved.
Output price
Provider list price per 1M output tokens, from each provider's own pricing page.
| Rank | Model | Output / 1M tokens |
|---|---|---|
| 1 | Command R7Bby Cohere | $0.15Checked Oct 8, 2026 |
| 2 | Ministral 3 14Bby Mistral AI | $0.20Checked Oct 8, 2026 |
| 3 | Qwen3.8 Flashby Alibaba Cloud (Qwen) | $0.47Checked Oct 8, 2026 |
| 4 | Claude Haiku 5.5by Anthropic | $0.50Checked Oct 8, 2026 |
| 5 | GPT-6 Lunaby OpenAI | $0.50Checked Oct 8, 2026 |
| 6 | Command R (08-2024)by Cohere | $0.60Checked Oct 8, 2026 |
| 7 | Mistral Small 4by Mistral AI | $0.60Checked Oct 8, 2026 |
| 8 | Codestralby Mistral AI | $0.90Checked Oct 8, 2026 |
| 9 | DeepSeek V4.1 Flashby DeepSeek | $1.20Checked Oct 8, 2026 |
| 10 | Mistral Large 3by Mistral AI | $1.50Checked Oct 8, 2026 |
Each provider's own list price per 1M output tokens, read on the provider's pricing page on the date shown in each row. Safeguard adds no markup. Models with no published per-token price (priced on request, per minute, per image or per search) are not listed, and models that use no tokens are on the token-free board.
Source: Safeguard model registry (safeguard.sh/models), as of 2026-10-08. All rights reserved.
Context length
The context window each model accepts, as its provider publishes it.
| Rank | Model | Context |
|---|---|---|
| 1 | Llama 4 Scoutby Meta | 10M |
| 2 | GPT-6 Astraby OpenAI | 1.05M |
| 3 | GPT-6 Lunaby OpenAI | 1.05M |
| 4 | GPT-6 Solby OpenAI | 1.05M |
| 5 | GPT-6.1 Solby OpenAI | 1.05M |
| 6 | Gemini 3.1 Pro (Preview)by Google | 1.05M |
| 7 | Gemini 3.5 Flashby Google | 1.05M |
| 8 | Gemini 3.5 Flash-Liteby Google | 1.05M |
| 9 | Gemini 3.8 Flashby Google | 1.05M |
| 10 | Muse Spark 1.3by Meta | 1.05M |
The context window in tokens of each published model, as recorded in the model registry.
Source: Safeguard model registry (safeguard.sh/models), as of 2026-10-08. All rights reserved.
Newest models
Models by release date, newest first.
| Rank | Model | Released |
|---|---|---|
| 1 | Claude Haiku 5.5by Anthropic | Oct 7, 2026 |
| 2 | Mistral Large 4 (Public Preview)by Mistral AI | Oct 6, 2026 |
| 3 | Nano Banana 2.1by Google | Oct 6, 2026 |
| 4 | GPT-6.1 Solby OpenAI | Sep 29, 2026 |
| 5 | Claude Sonnet 5.5by Anthropic | Sep 28, 2026 |
| 6 | Claude Opus 5.5by Anthropic | Sep 22, 2026 |
| 7 | GPT-6 Lunaby OpenAI | Sep 22, 2026 |
| 8 | GPT-6 Solby OpenAI | Sep 22, 2026 |
| 9 | DeepSeek V4.1 Flashby DeepSeek | Sep 10, 2026 |
| 10 | GPT Image 2.5 Sunburstby OpenAI | Sep 8, 2026 |
Published models by the release date the provider announced, newest first. Models without a published release date are not listed.
Source: Safeguard model registry (safeguard.sh/models), as of 2026-10-08. All rights reserved.
Token-free models
Models that use no model tokens, so there is no per-token charge.
| Rank | Model | Per 1M tokens |
|---|---|---|
| 1 | Zeroby Safeguard | $0 |
Models that consume no model tokens (for example deterministic analysis), so they cost nothing per token.
Source: Safeguard model registry (safeguard.sh/models), as of 2026-10-08. All rights reserved.
Licence terms from Safeguard Gold
How freely each model's licence allows commercial use, from its profile on Safeguard Gold.
| Rank | Model | Commercial use |
|---|---|---|
| 1 | DeepSeek V4 Proby DeepSeek | YesMIT |
| 2 | DeepSeek V4.1 Flashby DeepSeek | YesMIT |
| 3 | Gemma 4 31Bby Google | YesApache 2.0 |
| 4 | Muse Glimmerby Meta | YesApache 2.0 |
| 5 | Qwen3.8 27Bby Alibaba Cloud (Qwen) | YesApache 2.0 |
| 6 | gpt-oss-120bby OpenAI | YesApache 2.0 |
| 7 | Llama 3.3 70B Instructby Meta | ConditionalLlama 3.3 Community License |
| 8 | Llama 4 Scoutby Meta | Unknownother |
Method: each published model linked to its Safeguard Gold profile is classed by the licence Gold records for it, with the same classes as Gold's model pages: commercial use OK (for example MIT, Apache 2.0, CC BY), commercial use with conditions (community licences such as Llama and Gemma, responsible-AI licences, copyleft, and proprietary terms), no commercial use (CC BY-NC), or unclear (no or a custom licence). Gold records no known issues or vulnerabilities for models, so none are counted: this is licence guidance, not a vulnerability score, and not legal advice. Models without a Gold profile are not listed.
Source: Safeguard Gold model directory (open-model licences from the Hugging Face Hub) and Safeguard model registry (safeguard.sh/models), as of 2026-10-08. All rights reserved.
Deployment options
How many published models can run each way: through the Safeguard API, in public or private cloud, on-prem or air-gapped.
| Rank | Deployment option | Models |
|---|---|---|
| 1 | Private cloud | 61 |
| 2 | Public cloud | 44 |
| 3 | On-prem | 21 |
| 4 | Air-gapped | 21 |
| 5 | Safeguard API | 4 |
How many of the 61 published models with deployment options set can run each way through Safeguard. Set per model by Safeguard in the model registry. Closed third-party models are never offered on-prem or air-gapped.
Source: Safeguard model registry (safeguard.sh/models), as of 2026-10-08. All rights reserved.
How these rankings are measured
Safeguard rankings
These boards come from the Safeguard model registry, the data behind the model catalog, and from Safeguard Gold. Nothing is estimated: a model without the data a board needs is left off that board.
Prices are each provider's list price per 1M tokens, taken from the provider's own pricing page, with the date it was checked. No markup on model usage: you pay the provider's list price. Models without a published per-token price, and token-free models, are not on the price boards.
Context length and release dates are as each provider publishes them for the model.
Token-free models use no model tokens. Safeguard Zero, for example, runs deterministic analysis with no LLM.
Licence terms come from each model's profile on Safeguard Gold, grouped by how freely the licence allows commercial use: yes, conditional, no or unknown. Gold holds no known-issues data for models yet, so this board reflects licence terms only. It is not a security assessment.
Deployment and country counts are the number of published models the registry lists for each deployment option and each inference country or region. Regions for closed third-party models are recorded only from the cloud provider's own availability pages.
Self-healing security runs on Safeguard.
Your first fix PR is minutes away.
No sales call required, even your agent can complete the purchase over MCP.