Popular LLMs

This page now includes both model families and a concrete model catalog with many named models teams actively evaluate in 2026, including newer additions from recent model waves.

Latest refresh (July 23, 2026): Gemini 3.6 Flash (Google, July) is the newest stable — balances speed and intelligence for agentic tasks. Gemini Deep Research enables autonomous multi-step research. GPT-5.6 Sol/Terra/Luna (OpenAI, July) are the frontier — 1.05M context, computer use, web search, and file search. Sol for complex professional work, Terra balances capability with cost, Luna is budget-optimized. Claude Sonnet 5 joins the lineup — 1M context, $3/$15 per MTok (intro: $2/$10 through Aug 31), best speed/intelligence combo. Robostral Navigate (Mistral AI, July 8) for robotics. Leanstral 1.5 (Mistral AI, July 2) open source foundation model. DeepSeek Janus-Pro-7B for multimodal understanding and generation. Qwen3-ASR/TTS for speech recognition and synthesis. DeepSeek V4-Pro-DSpark (889B) and V4-Flash-DSpark (165B) are the newest open flagships.

OpenAI GPT Family

Why it is good: GPT-5.6 Sol/Terra/Luna are the newest frontier with 1.05M context and computer use. GPT-5.5 and GPT-5.5-Pro deliver strong reasoning with 1M context; GPT-5.4 mini (400K context) excels at coding, computer use, and subagent workflows at lower cost. Broad ecosystem and stable operations.

Why it can be bad: still premium in ultra-high-volume deployments; less control than self-hosted alternatives.

Best for: research, real-time analysis, production assistants, complex coding tasks, agentic workflows, and knowledge work requiring current information.

Anthropic Claude Family

Why it is good: Claude Fable 5 (June 2026) is the new frontier — state-of-the-art vision, autonomous long-running tasks with millions of tokens, persistent file-based memory, and $10/$50 per MTok. Mythos 5 is the same model with safeguards lifted for authorized cybersecurity and biomedical researchers. Claude Sonnet 5 (July 2026) offers the best speed/intelligence combination at $3/$15 per MTok. Opus 4.8 remains a strong secondary tier; Haiku 4.5 covers speed and cost.

Why it can be bad: Fable 5 redirects cybersecurity, bio/chemistry, and distillation topics to Opus 4.8 (<5% of sessions); Mythos 5 is controlled access only.

Best for: Frontier software engineering, vision and scientific research, autonomous multi-step coding, long-running agentic workflows, and enterprise use cases requiring careful outputs.

Google Gemini Family

Why it is good: Gemini 3.5 Flash now stable for production; Gemini 3.1 Pro in preview with advanced agentic and coding capabilities; strong multimodal support across text, image, audio, and video. Good fit for Google Cloud users.

Why it can be bad: 3.1 Pro still in preview; consistency can vary across prompt styles and niche coding tasks.

Best for: multimodal apps, agentic coding, search-enriched workflows, Google-native stacks, and enterprise customers.

Google Gemma Family (Open Source)

Why it is good: fully open-weight models under the permissive Gemma License (Apache-2.0 compatible); allows commercial use, modification, and self-hosting without vendor lock-in. Gemma 4 (April 2026) now purpose-built for advanced reasoning and agentic workflows with E2B/E4B tiers for mobile/IoT and 26B/31B for advanced reasoning on personal hardware.

Why it can be bad: requires in-house serving, infrastructure, and compliance review for some regulated industries; smaller fine-tuning community compared to Llama.

Best for: privacy-critical workflows, cost-optimized private inference, mobile/edge AI (E2B/E4B), compliance-sensitive enterprises with IP concerns, and hybrid cloud-on-prem stacks. Ideal for teams wanting full control and no API dependencies.

Meta Llama Family

Why it is good: open-weight flexibility, strong community support, easier self-hosting options, and now includes Llama 4 Maverick for visual reasoning tasks.

Why it can be bad: requires more in-house ML/platform work for top-tier quality and reliability; Scout variants trade some quality for speed.

Best for: cost-aware products, private deployments, custom fine-tuning paths.

Mistral Models

Why it is good: Mistral Medium 3.5 is now frontier-class for agentic and coding tasks; Mistral Large 3 available with open weights; Ministral 3 spans 3B/8B/14B for efficient deployments. Strong European adoption and compliance.

Why it can be bad: ecosystem and tooling footprint can be narrower than hyperscaler platforms.

Best for: latency-sensitive assistants, agentic coding, compact model deployments, regional compliance cases, and European data sovereignty.

Qwen and DeepSeek Families

Why they are good: DeepSeek V4-Pro (862B) and V4-Flash (158B) deliver strong general reasoning and coding; Qwen3.5 series spans 2B to 35B-A3B with hybrid thinking mode, 262K context, multimodal support, and efficient MoE architecture. Both families offer competitive performance at lower cost.

Why they can be bad: deployment/compliance review is required in regulated enterprise environments; thinking mode increases latency for simple tasks.

Best for: agentic coding, private hosting, value-focused inference, and teams needing controllable reasoning depth without paying frontier API prices.

Find the right model in seconds

Search by model name, provider, category, strengths, or limitations. Example searches: "coding", "low cost", "multilingual", "open weight", "enterprise".

Search models

Showing all models.

Concrete Model Catalog (2026)

This list is intentionally broad, so you can shortlist by exact model name before running your own benchmarks.

Direct Model Links

Quick access to the most requested model pages: GPT models, Claude models, Gemini models, Gemma models, Llama 4 Maverick, Llama 4 Scout, Mistral models, Qwen model hub, DeepSeek models

Model	Provider	Category	Good For	Watch Out	Download / Access
Gemini 3.6 Flash	Google	Closed stable	Newest stable — balances speed and intelligence for agentic and multimodal tasks	New — limited long-term track record	API access (new)
Gemini Deep Research	Google	Closed agent	Autonomous agent that plans and executes multi-step research across hundreds of sources	Preview status; specialized use case	Preview access
GPT-5.6 Sol	OpenAI	Closed frontier	Newest flagship — complex professional work, 1.05M context, computer use, web/file search	Highest cost tier	API access (new)
GPT-5.6 Terra	OpenAI	Closed balanced	Balances capability with affordability, 1.05M context, computer use	Lower ceiling than Sol on hardest tasks	API access (new)
GPT-5.6 Luna	OpenAI	Closed budget	Budget-optimized deployments, 1.05M context	Lower quality than Terra/Sol	API access (new)
GPT-5.5	OpenAI	Closed frontier	Live web data, 1M context, premium coding and reasoning	Highest cost tier; latency for web calls	API access
GPT-5.5-Pro	OpenAI	Closed highest-intelligence	Hardest reasoning problems, research-grade analysis	Higher latency and cost; overkill for simple tasks	API access (new)
GPT-5.4	OpenAI	Closed balanced	1M context, strong coding at lower cost than 5.5	Not ideal for the hardest reasoning chains	API access
GPT-5.4 mini	OpenAI	Closed efficient	400K context; coding, computer use, subagents at low cost	Smaller context than full 5.4; less depth on hardest tasks	API access (new)
Claude Sonnet 5	Anthropic	Closed balanced (new)	Best combination of speed and intelligence. 1M context, 128K output, $3/$15 per MTok (intro: $2/$10 through Aug 31).	Not as capable as Opus/Fable on hardest tasks	API access (new)
Claude Fable 5	Anthropic	Closed frontier (new)	State-of-the-art vision; autonomous long-running tasks with millions of tokens; persistent file-based memory; $10/$50 per MTok. Redirects cybersecurity/bio/chemistry to Opus 4.8 (<5% of sessions).	Controlled topics redirected; not lowest latency	API access (new)
Claude Mythos 5	Anthropic	Closed frontier — controlled access	Same underlying model as Fable 5; safeguards lifted for authorized cybersecurity professionals and biomedical researchers. $10/$50 per MTok.	Controlled access only — not general availability	Controlled access
Claude Opus 4.8	Anthropic	Closed frontier	Most capable model for complex reasoning and agentic coding; 1M context, adaptive thinking, 128K output	Premium cost ($5/$25 per MTok); moderate latency	API access (new)
Claude Opus 4.7	Anthropic	Closed (legacy)	Strong agentic tasks and reasoning; 1M context with adaptive thinking	Now legacy — consider migrating to Opus 4.8	API access
Claude Opus 4.6	Anthropic	Closed (legacy)	Extended thinking, 1M context, strong reasoning	Legacy; same pricing as 4.8 but less capable	API access
Claude Sonnet 4.6	Anthropic	Closed balanced	Best speed/intelligence combo; 1M context, extended thinking, $3/$15 per MTok	Not as capable as Opus on hardest tasks	API access
Claude Haiku 4.5	Anthropic	Closed small	Fastest model with near-frontier intelligence; 200K context, $1/$5 per MTok	Less robust on deepest tasks; smaller context	API access
Gemini 3.1 Pro	Google	Closed enterprise (preview)	Advanced intelligence, complex problem-solving, agentic coding	Still in preview; higher latency	Preview access
Gemini 3.1 Flash-Lite	Google	Closed efficient	Frontier-class performance at fraction of cost	Lower ceiling than Pro on complex tasks	API access
Gemini 3.5 Live Translate	Google	Closed translation	Real-time speech-to-speech translation in 70+ languages	Preview status	Preview access
Antigravity Agent	Google	Managed agent	Autonomous planning, code execution, web browsing in sandboxed Linux environments	Preview status; requires sandbox setup	Preview access
Gemini 3.5 Flash	Google	Closed fast (stable)	Most intelligent for agentic and coding tasks at speed	Lower quality than Pro tier on complex reasoning	API access (stable)
Gemini 3.1 Pro	Google	Closed enterprise (preview)	Advanced intelligence, complex problem-solving, agentic coding	Still in preview; higher latency	Preview access
Gemini 3 Flash	Google	Closed (preview)	Frontier-class performance at fraction of cost	Preview status; still maturing	Preview access
Gemini 2.5 Pro	Google	Closed	Reasoning and multimodal enterprise apps	Being superseded by 3.x series	API access
Gemma 4 E2B/E4B	Google	Open weight mobile/edge	Optimized for mobile and IoT; compute-efficient inference	Limited ceiling on complex reasoning; requires edge hardware	Model docs · Weights
Gemma 4 26B	Google	Open weight balanced	Advanced reasoning on personal hardware; agentic workflows	Needs good VRAM; careful quantization for consumer GPUs	Model docs · Weights
Gemma 4 31B	Google	Open weight high quality	Most capable Gemma for private reasoning and coding	Hardware intensive; best with 24GB+ VRAM or multi-GPU	Model docs · Weights
Llama 4 Maverick	Meta	Open MoE multimodal	17B params, 128 experts (402B total); flagship open-weight reasoning and vision	Full MoE serving requires strong infrastructure	Download · Scout
Llama 4 Scout	Meta	Open multimodal efficient	17B params, 16 experts (109B total); edge inference with vision-text support	Lower ceiling than Maverick; best for volume-optimized deployments	Download
Llama 3.1 405B Instruct	Meta	Open weight	Top-end open deployment quality	Heavy infrastructure requirements	Download · 70B · 8B
Llama 3.1 70B Instruct	Meta	Open weight	Strong self-hosted quality/cost balance	Needs good inference stack	Download · 405B · 8B
Llama 3.1 8B Instruct	Meta	Open weight small	Edge and low-cost deployments	Lower performance on complex tasks	Download · 70B · 405B
Llama 3.2 11B Vision	Meta	Open multimodal	Private vision-text pipelines	Requires evals for OCR-heavy cases	Download · 90B
Llama 3.2 90B Vision	Meta	Open multimodal	High-capacity multimodal inference	Infrastructure complexity	Download · 11B
Llama 3.3 70B Instruct	Meta	Open weight	Efficient self-hosted quality, matches 3.1 405B at much lower cost	Needs good inference stack for throughput	Download
Robostral Navigate	Mistral AI	Robotics-specialized	Navigation model for robotics and embodied AI. Launched July 8, 2026.	Specialized — not for general text generation	API access (new)
Leanstral 1.5	Mistral AI	Open source	Lean and accessible foundation model. Open source. Launched July 2, 2026.	Lower capacity than larger models	Download (new)
Pixtral Large	Mistral AI	Multimodal vision	Advanced vision-language model for image understanding and multimodal reasoning	Requires good infrastructure	Download
DeepSeek Janus-Pro-7B	DeepSeek	Open multimodal	Multimodal understanding and generation in one model. Both image-to-text and text-to-image.	7B may be limited for complex tasks	Download (new)
Qwen3-ASR	Alibaba	Speech-specialized	Automatic speech recognition with high accuracy. Available in 0.6B and 1.7B variants.	Specialized for ASR — not general text generation	Download (new)
Qwen3-TTS	Alibaba	Speech-specialized	Text-to-speech with voice design and voice cloning. Multilingual support.	Specialized for TTS — not general text generation	Demo (new)
Mistral Large 3	Mistral AI	Open weight multimodal	Advanced general-purpose with open weights available	Smaller ecosystem vs hyperscalers	API + weights
Mistral Medium 3.5	Mistral AI	Closed frontier	Frontier-class multimodal for agentic and coding tasks	Higher cost than smaller variants	API access
Mistral Small 4	Mistral AI	Closed small	Unified instruction-following, reasoning, and coding	Limited depth on advanced reasoning	API access
Ministral 3 14B	Mistral AI	Open small	Best-in-class text and vision at compact size	Can trail latest closed models	API access
Ministral 3 8B	Mistral AI	Open compact	Efficient text and vision on consumer hardware	Lower ceiling than 14B on complex tasks	API access
Ministral 3 3B	Mistral AI	Open tiny	Ultra-compact with multimodal support for edge/mobile	Very limited on complex reasoning	API access
Devstral 2	Mistral AI	Code-specialized open	Software engineering and code review	Narrower general language strength	API access
Qwen3.5-35B-A3B	Alibaba	Open weight MoE multimodal	MoE (35B total, 3B active); hybrid thinking mode; 262K native context; multimodal (text, image, video); agentic coding. SWE-bench: 73.4	Regional compliance review required; thinking mode adds latency for simple tasks	Download · FP8
Qwen3.5-27B	Alibaba	Open weight balanced	Strong reasoning at moderate size; 80K context	Regional compliance review; needs good VRAM	Model hub
Qwen3.5-9B	Alibaba	Open weight compact	64K context; efficient private inference	Lower ceiling than larger Qwen3.5 variants	Model hub
Qwen3 32B Instruct	Alibaba	Open weight	Strong open-weight multilingual assistant quality	Regional compliance and policy review required	Model hub
DeepSeek V4-Pro-DSpark	DeepSeek	Open/available flagship	Newest and largest — 889B parameters. Enhanced reasoning and agent capabilities.	Massive infrastructure required; governance review in enterprise	Download (new)
DeepSeek V4-Flash-DSpark	DeepSeek	Open/available efficient	165B parameters; faster inference with good quality	Governance review in enterprise	Download (new)
DeepSeek V4-Pro	DeepSeek	Open/available flagship	862B parameters; strong general reasoning and coding	Massive model requires substantial infrastructure	Download (new)
DeepSeek V4-Flash	DeepSeek	Open/available efficient	158B parameters; fast general reasoning at lower cost	Governance review in enterprise; less capable than V4-Pro	Download
DeepSeek V3.2	DeepSeek	Open/available	685B parameters; mature and well-tested	Superseded by V4 series; large infrastructure needed	Download

Note: closed models usually provide API access rather than direct weight downloads.

Longlist: Best Candidates by Workload

Use this longer shortlist before final benchmarks. Pick 3-5 from each category, run your evals, and track quality, latency, and cost per successful task.

Reasoning and Decision Support

GPT-5.6 Sol
Claude Fable 5
Claude Opus 4.8
GPT-5.5-Pro
GPT-5.5
Claude Sonnet 5
Gemini 3.1 Pro
Gemini 2.5 Pro
Gemma 4 31B
Llama 4 Maverick
DeepSeek V4-Pro-DSpark
Qwen3.5-35B-A3B

Coding and Developer Assistant Workloads

GPT-5.6 Sol
Claude Fable 5
Claude Opus 4.8
Claude Sonnet 5
GPT-5.5
GPT-5.4 mini
DeepSeek V4-Pro-DSpark
Mistral Medium 3.5
Devstral 2
Llama 4 Maverick
Qwen3.5-35B-A3B
Qwen3 32B Instruct
Gemini 3.5 Flash

High-Volume, Cost-Efficient Automation

GPT-5.6 Luna
GPT-5.4
GPT-5.4 mini
Claude Haiku 4.5
Gemini 3.5 Flash
Gemini 3.1 Flash-Lite
Gemma 4 E2B/E4B
Gemma 4 26B
Llama 4 Scout
Mistral Small 4
Ministral 3 8B

Private or Self-Hosted Enterprise Paths

Llama 4 Maverick
Llama 4 Scout
Llama 3.3 70B Instruct
Gemma 4 31B
Gemma 4 26B
Mistral Large 3
Qwen3.5-35B-A3B
Qwen3 32B Instruct
DeepSeek V4-Flash-DSpark
DeepSeek V4-Pro-DSpark

New Model Radar for July 2026

Gemini 3.6 Flash (July) is Google's newest stable model — balances speed and intelligence for agentic and multimodal tasks.
Gemini Deep Research (July) enables autonomous multi-step research across hundreds of sources.
GPT-5.6 Sol (July) is OpenAI's newest frontier — 1.05M context, computer use, web search, and file search. Built for complex professional work.
GPT-5.6 Terra balances capability with cost at 1.05M context. Good choice for most use cases.
GPT-5.6 Luna is budget-optimized with the same 1.05M context window.
Claude Sonnet 5 (July) joins the lineup — 1M context, $3/$15 per MTok (intro: $2/$10 through Aug 31). Best speed/intelligence combination.
Claude Fable 5 (June 9) remains the frontier for long-running autonomous tasks — state-of-the-art vision, persistent file-based memory at $10/$50 per MTok.
Gemini 3.1 Pro is now in preview with advanced reasoning and agentic coding capabilities.
Gemini 3.5 Live Translate enables real-time speech-to-speech translation in 70+ languages.
Antigravity Agent is Google's managed agent for autonomous planning and code execution in sandboxed environments.
DeepSeek V4-Pro-DSpark (889B) and V4-Flash-DSpark (165B) are the newest open flagships from DeepSeek.
Robostral Navigate (July 8) is Mistral's new robotics navigation model for embodied AI.
Leanstral 1.5 (July 2) is an open source foundation model from Mistral AI.
DeepSeek Janus-Pro-7B offers multimodal understanding and generation in a single compact model.
Qwen3-ASR/TTS adds speech recognition and synthesis capabilities to the Qwen family.
For agentic workflows, test Claude Fable 5 against GPT-5.6 Sol to find optimal quality/cost balance for your use case.

How to Pick from This List

Do not pick only by benchmark rank. Validate with your own workload: prompt complexity, response latency, failure tolerance, and monthly token budget.

Continue with the comparison matrix and then read clear recommendations.