You can combine filters. Only models matching every selection will remain.
Suggested starting points
Three options for different kinds of work. This is not a best-to-worst ranking.
GPT-6 Astra
top pick
OpenAI
Summary
coding pick backed by current AA.
GPT-6 Astra has IQ 56.9 and input $10/M in the sources. Consider it for coding, AI agents; for mass-volume, self-hosted stack, run a second benchmark before rollout.
+Good forcoding · AI agents · vision and multimodal
−Not suited tomass volume · self-hosted stack · strict latency
GLM-5.3 has IQ 31.7 and input $1.4/M in the sources. Consider it for self-hosted stack, sensitive deployments; for premium-agents, top-coding, run a second benchmark before rollout.
+Good forself-hosted stack · sensitive deployments · large batches
Gemini 4 Argon has IQ 50.4 and input $4/M in the sources. Consider it for company knowledge, vision and multimodal; for real-time-latency, self-hosted stack, run a second benchmark before rollout.
+Good forcompany knowledge · vision and multimodal · multilingual content
A concise overview. Detailed numbers, sources and reasoning live on each model page.
No model matches this combination. Try fewer filters.
ModelSummaryGood fit / poor fitData
GPT-6 Astra
OpenAI
Summary
coding pick backed by current AA.
GPT-6 Astra has IQ 56.9 and input $10/M in the sources. Consider it for coding, AI agents; for mass-volume, self-hosted stack, run a second benchmark before rollout.
+Good forcoding · AI agents · vision and multimodal
−Not suited tomass volume · self-hosted stack · strict latency
Gemini 4 Argon has IQ 50.4 and input $4/M in the sources. Consider it for company knowledge, vision and multimodal; for real-time-latency, self-hosted stack, run a second benchmark before rollout.
+Good forcompany knowledge · vision and multimodal · multilingual content
Claude Fable 5.1 has IQ 50 and input $10/M in the sources. Consider it for coding, AI agents; for mass-volume, real-time-latency, run a second benchmark before rollout.
Claude Opus 5.5 has IQ 48.2 and input $4/M in the sources. Consider it for coding, AI agents; for mass-volume, real-time-latency, run a second benchmark before rollout.
Grok 4.7 has IQ 35.2 and input $2/M in the sources. Consider it for coding, fast replies; for self-hosted stack, sensitive deployments, run a second benchmark before rollout.
+Good forcoding · fast replies · document extraction
Qwen3.8 Max has IQ 33.7 and input $2/M in the sources. Consider it for multilingual content, coding; for self-hosted stack, premium-agents, run a second benchmark before rollout.
+Good formultilingual content · coding · large batches
GLM-5.3 has IQ 31.7 and input $1.4/M in the sources. Consider it for self-hosted stack, sensitive deployments; for premium-agents, top-coding, run a second benchmark before rollout.
+Good forself-hosted stack · sensitive deployments · large batches
DeepSeek V4 Pro 0813 has IQ 23.1 and input $1.32/M in the sources. Consider it for large batches, coding; for enterprise-governance, premium-agents, run a second benchmark before rollout.
Claude Sonnet 5.5 has IQ 30.8 and input $2/M in the sources. Consider it for coding, AI agents; for self-hosted stack, mass-volume, run a second benchmark before rollout.
DeepSeek V4.1 Flash has input $0.3/M in the sources. Consider it for large batches, fast replies; for deep-frontier-reasoning, top-coding, run a second benchmark before rollout.
+Good forlarge batches · fast replies · self-hosted stack
−Not suited todeep frontier reasoning · top coding
$0.3input / 1M$1.2output / 1M
open / self-run · specialist · verified by external data
Llama 4 Maverick has input $0.25/M in the sources. Consider it for self-hosted stack, sensitive deployments; for managed-api-comfort, deep-frontier-reasoning, run a second benchmark before rollout.
+Good forself-hosted stack · sensitive deployments · company knowledge
−Not suited tomanaged API comfort · deep frontier reasoning
$0.25input / 1M$0.92output / 1M
open / self-run · specialist · verified by external data
Kimi K3 has IQ 17.4 and input $3/M in the sources. Consider it for coding, large batches; for enterprise-governance, tool-use, run a second benchmark before rollout.
+Good forcoding · large batches · company knowledge
−Not suited tocompany controls and audit · reliable tool use
Command A Plus has IQ 6.2 and input $0/M in the sources. Consider it for company knowledge, document extraction; for top-coding, deep-frontier-reasoning, run a second benchmark before rollout.
GPT-5.6 Sol has input $4/M in the sources. Consider it for coding, AI agents; for self-hosted stack, real-time-latency, run a second benchmark before rollout.
+Good forcoding · AI agents · document extraction
−Not suited toself-hosted stack · strict latency
$4.0input / 1M$20.0output / 1M
mid-budget · specialist · verified by external data
Gemini 3.6 Flash has input $0.75/M in the sources. Consider it for large batches, company knowledge; for deep-coding, real-time-latency, run a second benchmark before rollout.
+Good forlarge batches · company knowledge · fast replies
−Not suited tohard coding work · strict latency
$0.75input / 1M$3.75output / 1M
cheap at volume · specialist · verified by external data
specialistcheap at volumelarge batchescompany knowledge
Mistral Medium 3.5 has input $1.5/M in the sources. Consider it for sensitive deployments, company knowledge; for deep-frontier-reasoning, top-coding, run a second benchmark before rollout.
+Good forsensitive deployments · company knowledge · document extraction
−Not suited todeep frontier reasoning · top coding
$1.5input / 1M$7.5output / 1M
mid-budget · specialist · verified by external data
A curated model selection updated daily. Prices and the Intelligence Index are checked against fresh Artificial Analysis data; additional benchmarks identify their own sources. A change in benchmark methodology can change a score without a change to the model, so older and newer index values may not be directly comparable. If validation fails, the last verified snapshot keeps its original date. When Artificial Analysis has not yet re-scored a model after a methodology change, its index is left blank and it ranks after scored models; I do not reuse an old number from a different scale.
