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 (April 12, 2026): GPT-5 Turbo now includes video understanding alongside live web retrieval; Claude 4.5 Sonnet extends to 2M tokens with improved code generation; o3 now natively multimodal with spatial reasoning; Llama 4.1 Scout 25% faster inference, Llama 4.2 Adventurer for visual tasks; Grok-3 emerging as strong analytical alternative; DeepSeek R1.5 improves reasoning on complex problems; Mistral Ultra competitive for enterprise; Gemini 2.5 Ultra now available for enterprise customers; Gemma 4 now covered as concrete 4B/12B/27B variants for private and hybrid deployment paths; Qwen3.6-35B-A3B now available — MoE model (35B total, 3B active) with 262K native context, multimodal input, agentic coding, and hybrid thinking mode. AIME 2026: 92.7, SWE-bench Verified: 73.4. Pricing cuts across model families enable new cost-optimization strategies. Video generation now integral to major platforms; copyright licensing accelerating; context windows stabilizing at 200K+ as baseline.
Why it is good: GPT-5 Turbo now includes live web data; exceptional reasoning, coding quality, broad ecosystem, and stable operations. Recently reduced pricing on GPT-4 routes by 40%.
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, and knowledge work requiring current information.
Why it is good: Claude 4.5 Sonnet now supports 2M token context; excellent writing quality, careful reasoning, and conservative style. Strong for compliance-heavy workflows.
Why it can be bad: may be more cautious than desired for creative tasks; slower on high-volume inference than newer alternatives.
Best for: legal/compliance review, long-form analysis, multi-document workflows, sensitive enterprise use cases, and policy-heavy content generation.
Why it is good: strong multimodal support, very capable for vision + text pipelines, good fit for Google Cloud users.
Why it can be bad: consistency can vary across prompt styles and niche coding tasks.
Best for: multimodal apps, search-enriched workflows, Google-native stacks, and enterprise customers.
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 now spans practical 4B, 12B, and 27B tracks that cover edge to high-quality private inference. No licensing restrictions on derived works or private outputs.
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, compliance-sensitive enterprises with IP concerns, and hybrid cloud-on-prem stacks. Ideal for teams wanting full control and no API dependencies.
Why it is good: open-weight flexibility, strong community support, easier self-hosting options, and now includes Llama 4.2 Adventurer 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.
Why it is good: efficient models with strong speed/quality balance and excellent European adoption.
Why it can be bad: ecosystem and tooling footprint can be narrower than hyperscaler platforms.
Best for: latency-sensitive assistants, compact model deployments, regional compliance cases.
Why they are good: often strong coding and reasoning performance for cost, popular in open model benchmarking. Qwen3.6-35B-A3B is a MoE model (35B total, 3B active) with hybrid thinking mode, 262K token native context, multimodal support (text, image, video), and agentic coding with repo-level reasoning. AIME 2026: 92.7, GPQA Diamond: 86.0, SWE-bench Verified: 73.4.
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.
Search by model name, provider, category, strengths, or limitations. Example searches: "coding", "low cost", "multilingual", "open weight", "enterprise".
Showing all models.
This list is intentionally broad, so you can shortlist by exact model name before running your own benchmarks.
Quick access to the most requested model pages: GPT models, Claude models, Gemini models, Gemma models, Llama 4 Maverick, Llama 4 Scout, Llama 3.3 70B, Llama 3.1 8B, Llama 3.1 70B, Mixtral 8x7B, Qwen3 model hub, Qwen3 32B, DeepSeek V3, DeepSeek R1
| Model | Provider | Category | Good For | Watch Out | Download / Access |
|---|---|---|---|---|---|
| GPT-5 Turbo | OpenAI | Closed frontier | Live web data, premium coding and reasoning | Highest cost tier; latency for web calls | API access (new) |
| GPT-5 | OpenAI | Closed | General reasoning and coding | Premium cost (high), though cheaper than Turbo | API access |
| GPT-5 mini | OpenAI | Closed small | Balanced quality/latency for production APIs | Not ideal for the hardest reasoning chains | API access |
| GPT-4o | OpenAI | Closed multimodal | Fast assistant UX and multimodal tasks | Cost at high scale | API access |
| GPT-4o mini | OpenAI | Closed small | Cost-sensitive high-volume automation (40% price cut as of March 31) | Lower ceiling on hard reasoning; now very economical | API access |
| o3 | OpenAI | Reasoning-first multimodal | Complex multi-step logic, now with native image/document understanding | Latency and cost per hard query; best for genuinely difficult problems | API access (multimodal) |
| o4-mini | OpenAI | Reasoning-efficient | Technical Q&A and coding workflows | Can require prompt tuning | API access |
| Claude 4.5 Sonnet | Anthropic | Closed | Long-context writing and analysis (now 2M tokens) | Conservative tone in some flows; still slower than newer alternatives | API access |
| Claude 4 Haiku | Anthropic | Closed small | Fast responses and triage (35% price cut as of March 31) | Less robust on deepest tasks; now very economical | API access |
| Claude 4 Opus | Anthropic | Closed flagship | High-stakes synthesis | Throughput economics | API access |
| Gemini 2.5 Ultra | Closed enterprise | Enterprise multimodal, advanced reasoning and analysis | Higher latency; requires enterprise contract | Enterprise access (new) | |
| Gemini 2.5 Pro | Closed | Reasoning and multimodal enterprise apps | Task variance across prompt styles | API access | |
| Gemini 2.5 Flash | Closed fast | Low-latency assistant endpoints | Lower quality than premium tier | API access | |
| Gemma 4 4B Instruct | Open weight compact | Low-VRAM local assistants, lightweight RAG, and fast edge inference | Lower ceiling on difficult coding/reasoning than larger variants | Model docs · Weights | |
| Gemma 4 12B Instruct | Open weight balanced | Balanced private inference for coding, support, and multi-document workflows | Needs careful quantization/runtime setup for 12-16GB VRAM systems | Model docs · Weights | |
| Gemma 4 27B Instruct | Open weight high quality | High-quality private reasoning/coding tiers without API lock-in | Hardware intensive; best with larger VRAM or multi-GPU setups | Model docs · Weights | |
| Llama 4 Maverick | Meta | Open MoE multimodal | Flagship open-weight reasoning, vision + text pipelines | Full MoE serving requires strong infrastructure | Download · Scout |
| Llama 4.1 Scout | Meta | Open multimodal efficient | Edge inference with 25% faster inference than prior versions | Lower ceiling than Maverick; best for volume-optimized deployments | Download (new) |
| Llama 4 Scout | Meta | Open multimodal efficient | Edge inference, vision-text tasks, low-cost deployments | Lower ceiling than Maverick on complex reasoning | 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 |
| Mistral Large 2 | Mistral AI | Closed | High-quality enterprise assistants | Smaller ecosystem vs hyperscalers | API access |
| Mistral Medium | Mistral AI | Closed | Balanced production usage | Benchmark carefully vs peers | API access |
| Mistral Small | Mistral AI | Closed small | Fast cost-efficient chat | Limited depth on advanced reasoning | API access |
| Mixtral 8x22B | Mistral AI | Open MoE | Strong open-weight generation quality | Operational complexity | Download · 8x7B |
| Mixtral 8x7B | Mistral AI | Open MoE | Efficient self-hosting | Can trail latest closed models | Download · 8x22B |
| Codestral | Mistral AI | Code-specialized | Code generation and completion | Narrower general language strength | Download |
| Qwen3 32B Instruct | Alibaba | Open weight | Strong open-weight multilingual assistant quality | Regional compliance and policy review required | Model hub |
| Qwen3.6-35B-A3B | Alibaba | Open weight MoE multimodal | MoE (35B total, 3B active); hybrid thinking mode; 262K native context (up to ~1M with YaRN); multimodal (text, image, video); agentic coding with repo-level reasoning. AIME 2026: 92.7, GPQA Diamond: 86.0, SWE-bench: 73.4 | Regional compliance review required; thinking mode adds latency for simple tasks | Download · FP8 |
| QwQ-32B | Alibaba | Reasoning open | Reasoning-focused private usage | Evals needed for stability | Download |
| DeepSeek V3 | DeepSeek | Open/available | General reasoning and coding value | Governance review in enterprise | Download |
| DeepSeek R1.5 | DeepSeek | Reasoning-focused | Improved analytical reasoning and problem-solving (March 2026 release) | Latency on complex outputs; governance review required | Download (new) |
| DeepSeek R1 | DeepSeek | Reasoning-focused | Difficult multi-step reasoning tasks | Latency on complex outputs | Download |
| DeepSeek Coder V3 | DeepSeek | Code-specialized | Developer assistants and code review | General writing less strong | Model hub |
Note: closed models usually provide API access rather than direct weight downloads.
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.
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.