Blog
Observatory

AI Model Observatory

Structured data on significant AI models, maintained from official announcements and documentation. Each entry tracks capability, access and limitations - not marketing claims.

15 models·8 developers
ModelDeveloperReleasedTypeContextModalitiesReasoningToolsWeights
Claude Opus 4Anthropic2025-05-22Large language model200Ktext, imageYesYesclosed
Claude Sonnet 4Anthropic2025-05-22Large language model200Ktext, imageYesYesclosed
Qwen3Alibaba Cloud2025-04-29Large language model131KtextYesYesopen
Gemini 2.5 FlashGoogle DeepMind2025-04-17Large language model1.0Mtext, image, audio, videoYesYesclosed
o3OpenAI2025-04-16Reasoning model200Ktext, imageYesYesclosed
o4-miniOpenAI2025-04-16Reasoning model200Ktext, imageYesYesclosed
GPT-4.1OpenAI2025-04-14Large language model1.0Mtext, imageNoYesclosed
Llama 4 MaverickMeta2025-04-05Large language model (MoE)1.0Mtext, imageNoYesopen
Llama 4 ScoutMeta2025-04-05Large language model (MoE)10.5Mtext, imageNoYesopen
Gemini 2.5 ProGoogle DeepMind2025-03-25Large language model1.0Mtext, image, audio, videoYesYesclosed
Command ACohere2025-03-13Large language model256KtextNoYesopen
DeepSeek-R1DeepSeek2025-01-20Reasoning model128KtextYesNoopen
DeepSeek-V3DeepSeek2024-12-26Large language model (MoE)128KtextNoYesopen
Mistral Large 2Mistral AI2024-07-24Large language model128KtextNoYesopen
GPT-4oOpenAI2024-05-13Large language model128Ktext, image, audioNoYesclosed

Model details

Claude Opus 4Anthropic · 2025-05-22
Type
Large language model
Modalities
text, image
Context window
200K tokens
Max output
32K tokens
Reasoning
Yes
Tool use
Yes
Agent capable
Yes
Weights
closed
Licence
Proprietary
Known limitations
  • Extended thinking uses additional tokens
  • Image understanding but no image generation
Source
Claude Sonnet 4Anthropic · 2025-05-22
Type
Large language model
Modalities
text, image
Context window
200K tokens
Max output
16K tokens
Reasoning
Yes
Tool use
Yes
Agent capable
Yes
Weights
closed
Licence
Proprietary
Known limitations
  • Lower capability ceiling than Opus on complex tasks
  • Extended thinking available but uses more tokens
Source
Qwen3Alibaba Cloud · 2025-04-29
Type
Large language model
Modalities
text
Context window
131K tokens
Reasoning
Yes
Tool use
Yes
Agent capable
Yes
Weights
open
Licence
Apache 2.0
Known limitations
  • Hybrid thinking mode switches between fast and reasoning
  • Largest dense variant is 32B parameters
Source
Gemini 2.5 FlashGoogle DeepMind · 2025-04-17
Type
Large language model
Modalities
text, image, audio, video
Context window
1.0M tokens
Max output
66K tokens
Reasoning
Yes
Tool use
Yes
Agent capable
Yes
Weights
closed
Licence
Proprietary
Known limitations
  • Optimised for speed over peak capability
  • Lower ceiling on hard reasoning tasks than Pro
Source
o3OpenAI · 2025-04-16
Type
Reasoning model
Modalities
text, image
Context window
200K tokens
Max output
100K tokens
Reasoning
Yes
Tool use
Yes
Agent capable
Yes
Weights
closed
Licence
Proprietary
Known limitations
  • Higher latency and cost due to chain-of-thought
  • Internal reasoning tokens are not shown
Source
o4-miniOpenAI · 2025-04-16
Type
Reasoning model
Modalities
text, image
Context window
200K tokens
Max output
100K tokens
Reasoning
Yes
Tool use
Yes
Agent capable
Yes
Weights
closed
Licence
Proprietary
Known limitations
  • Smaller model with less world knowledge than o3
  • Optimised for cost over peak capability
Source
GPT-4.1OpenAI · 2025-04-14
Type
Large language model
Modalities
text, image
Context window
1.0M tokens
Max output
33K tokens
Reasoning
No
Tool use
Yes
Agent capable
Yes
Weights
closed
Licence
Proprietary
Known limitations
  • Not a reasoning model - use o-series for complex multi-step logic
Source
Llama 4 MaverickMeta · 2025-04-05
Type
Large language model (MoE)
Modalities
text, image
Context window
1.0M tokens
Reasoning
No
Tool use
Yes
Agent capable
Yes
Weights
open
Licence
Llama 4 Community License
Known limitations
  • MoE architecture - active parameters are a fraction of total
  • Licence has commercial use restrictions above 700M monthly active users
Source
Llama 4 ScoutMeta · 2025-04-05
Type
Large language model (MoE)
Modalities
text, image
Context window
10.5M tokens
Reasoning
No
Tool use
Yes
Agent capable
No
Weights
open
Licence
Llama 4 Community License
Known limitations
  • 10M token context is experimental
  • Smaller active parameter count than Maverick
Source
Gemini 2.5 ProGoogle DeepMind · 2025-03-25
Type
Large language model
Modalities
text, image, audio, video
Context window
1.0M tokens
Max output
66K tokens
Reasoning
Yes
Tool use
Yes
Agent capable
Yes
Weights
closed
Licence
Proprietary
Known limitations
  • Very large context window but retrieval quality degrades with volume
  • Thinking tokens billed separately
Source
Command ACohere · 2025-03-13
Type
Large language model
Modalities
text
Context window
256K tokens
Reasoning
No
Tool use
Yes
Agent capable
Yes
Weights
open
Licence
CC-BY-NC 4.0
Known limitations
  • Non-commercial licence for open weights
  • Text only - no image input
Source
DeepSeek-R1DeepSeek · 2025-01-20
Type
Reasoning model
Modalities
text
Context window
128K tokens
Max output
33K tokens
Reasoning
Yes
Tool use
No
Agent capable
No
Weights
open
Licence
MIT
Known limitations
  • Text only - no image input
  • Reasoning traces can be very long, increasing cost
  • No native tool calling
Source
DeepSeek-V3DeepSeek · 2024-12-26
Type
Large language model (MoE)
Modalities
text
Context window
128K tokens
Max output
8K tokens
Reasoning
No
Tool use
Yes
Agent capable
No
Weights
open
Licence
MIT (model), DeepSeek License (some outputs)
Known limitations
  • Text only
  • MoE with 37B active of 671B total parameters
Source
Mistral Large 2Mistral AI · 2024-07-24
Type
Large language model
Modalities
text
Context window
128K tokens
Reasoning
No
Tool use
Yes
Agent capable
Yes
Weights
open
Licence
Mistral Research License (non-commercial)
Known limitations
  • Research licence restricts commercial use without agreement
  • 123B parameters requires substantial hardware
Source
GPT-4oOpenAI · 2024-05-13
Type
Large language model
Modalities
text, image, audio
Context window
128K tokens
Max output
16K tokens
Reasoning
No
Tool use
Yes
Agent capable
Yes
Weights
closed
Licence
Proprietary
Known limitations
  • Hallucination on niche domains
  • No persistent memory across sessions without external tooling
Source

Search the site