Gemini 3.8 (Flash & Pro)
Best for native 2M context with sub-125ms live multimodal streaming, top-ranked general reasoning (65.4 on General365), and integrated SynthID Universal provenance watermarks.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Google AI Studio, Vertex AI, Gemini Advanced
Important limitation: Full 2M token context requires enterprise API tier; dynamic thinking mode tokens billed at standard output rates.
Check official model information↗ (opens in a new tab)GPT-6 (Sol & Luna)
Best for dual-engine speculative decoding runtime (Sol 180 tok/s, Luna deep reasoning verifier) with native declarative Intelligent UI component trees.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: ChatGPT Enterprise, Pro, Plus, API
Important limitation: Mandatory C2PA metadata injection enforced on all visual generation; synthetic election asset restrictions active.
Check official model information↗ (opens in a new tab)Claude 3.8 (Workspaces & Motion)
Best for native host POSIX microVM sandboxing, collaborative workspace artifacts GA, and Motion dynamic animated WebGL visualization engine.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Anthropic API, Claude Enterprise, Claude Pro
Important limitation: Higher token price tier; requires hardware virtualization (KVM) support on agent host.
Check official model information↗ (opens in a new tab)Qwen-Image-2.1-Turbo
Best for accelerated 8-step Flow Matching Euler diffusion with 64-channel RGBA VAE for native 2K transparent image synthesis (418 ms on H100).
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Open Weights (Hugging Face / ModelScope), Alibaba Cloud API
Important limitation: Commercial self-hosting requires enterprise commercial license agreement from Alibaba Tongyi Lab.
Check official model information↗ (opens in a new tab)DeepSeek-V3.5
Best for multi-Head Latent Attention (75% KV-cache reduction) and DualPipe scheduling optimized natively across 160,000 Huawei Ascend 910C nodes.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Open Weights (Hugging Face / GitHub), DeepSeek API, Huawei Ascend CANN
Important limitation: Non-CUDA deployment requires Huawei CANN 8.0+ runtime libraries and specialized cluster interconnect switches.
Check official model information↗ (opens in a new tab)OpenAI Operator (Web Agent)
Best for autonomous multi-tab browser execution inside ephemeral MicroVM sandboxes (WebArena 68.2%, OSWorld 44.5%), real-time DOM action verification.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: ChatGPT Pro, ChatGPT Team, OpenAI API (Pilot)
Important limitation: Requires explicit human confirmation for financial transactions, credential entry, and external email dispatch. Rate limited to 10 concurrent browser sessions.
Check official model information↗ (opens in a new tab)Gemini 2.5 Pro (Thinking Mode)
Best for native recursive scratchpad execution with sandboxed Python compiler tool calling; excels at massive document analysis and multi-step math reasoning.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Google Cloud Vertex AI, Google AI Studio, Gemini Advanced
Important limitation: Extended thinking budget up to 32k tokens increases time-to-first-token on complex multi-turn prompts.
Check official model information↗ (opens in a new tab)Llama 4 Scout (Architecture Preview)
Best for sparse Mixture-of-Experts with native draft-model speculative decoding, delivering 2.8x higher throughput on 8x H100 clusters (MATH-500 89.4%, GSM8K 96.2%).
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Hugging Face, Meta AI Research, GitHub
Important limitation: Preview architecture weights for research verification; fine-tuning recipes require FP8-compatible distributed kernels (Megatron-LM or vLLM).
Check official model information↗ (opens in a new tab)Claude 3.7 Sonnet
Best for dual-mode hybrid reasoning with user-defined thinking budgets (SWE-bench Verified 70.3%, Terminal-Bench 58.2%, GPQA Diamond 65.1%) and 60 FPS Computer Use v2.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Claude.ai (Pro/Team/Free), Anthropic API, Amazon Bedrock, Google Cloud Vertex AI
Important limitation: Internal thinking tokens billed at discounted $3.00/1M rate (80% discount vs visible output). Extended thinking mode adds 20–45s latency on deep queries.
Check official model information↗ (opens in a new tab)OpenAI o3-pro
Best for deep test-time tree search with integrated tool calling (SWE-bench Verified 78.4%, MATH-500 92.1%, FrontierMath 34.2%, GPQA Diamond 87.6%).
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: OpenAI API, ChatGPT Pro, Microsoft Azure OpenAI Service
Important limitation: Thinking tokens billed at standard $60.00/1M output rate. Heavy search budgets can result in $1.50–$3.00 per individual complex reasoning call.
Check official model information↗ (opens in a new tab)Qwen 2.5-Coder-32B-Reasoner
Best for open-weights (Apache 2.0), step-level Process Reward Model scoring, SWE-bench Verified 65.8%, HumanEval+ 81.2%, runs on dual RTX 4090 GPUs in FP8.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Hugging Face, ModelScope, Alibaba Cloud DashScope, vLLM
Important limitation: Requires 4x test-time token generation for beam search; best served on high-throughput vLLM nodes with PagedAttention.
Check official model information↗ (opens in a new tab)Claude 3.7 Sonnet (Hybrid Thinking Mode)
Best for dynamic thinking budgets (1k-64k tokens), SWE-bench Verified 70.3%, HumanEval 96.2%, GPQA Diamond 84.8%, Computer Use 2.0 with 40% lower action latency.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Anthropic API, Claude Enterprise, AWS Bedrock, Google Cloud Vertex AI
Important limitation: Thinking tokens billed at standard output rate ($15.00/1M). Long reasoning traces can substantially increase total request cost and response latency.
Check official model information↗ (opens in a new tab)Claude 3.7 Sonnet (Standard Latency Mode)
Best for instant-generation flagship performance without scratchpad latency, state-of-the-art general instruction following and natural voice prose.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Anthropic API, Claude Free/Pro/Team/Enterprise
Important limitation: Complex competitive mathematics and multi-step bug localization require enabling thinking mode for maximum verification accuracy.
Check official model information↗ (opens in a new tab)o3-mini (High Effort)
Best for controllable test-time search (MATH-500 88.2%, SWE-bench Verified 59.4%, AIME 2025 79.2%), eBPF-isolated microVM execution.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: OpenAI Developer API, ChatGPT Plus/Team/Enterprise
Important limitation: High reasoning effort incurs variable latency (up to 30s) per query. Hidden reasoning tokens billed at regular output rate ($1.60/1M).
Check official model information↗ (opens in a new tab)o3-mini (Medium Effort)
Best for balanced latency/reasoning trade-off (MATH-500 84.8%, SWE-bench Verified 51.2%), fast structured tool calling.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: OpenAI Developer API, ChatGPT Plus/Team/Enterprise
Important limitation: Slightly lower competitive math depth than full o1, optimized for high-throughput code editing pipelines.
Check official model information↗ (opens in a new tab)Gemini 2.5 Pro Reasoner
Best for test-time tree search (MATH-500 86.4%, SWE-bench Verified 61.2%, FrontierMath 28.4%), integrated robotics spatial grounding.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Google AI Studio, Vertex AI, Gemini API
Important limitation: Thought/scratchpad tokens billed at $5.00/1M output rate. Robotic kinematic planning introduces 800ms-1.4s inference overhead per motion primitive.
Check official model information↗ (opens in a new tab)Mistral Large 3
Best for hybrid 3:1 Mamba-3 SSM / Transformer architecture, 78.4% reduction in KV-cache VRAM footprint, RULER 256k 88.6%.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Open weights (Mistral Research License), Mistral La Plateforme
Important limitation: Dense 145B parameters. Requires 8x H200 or B200 accelerators for full 1M-context serving.
Check official model information↗ (opens in a new tab)Llama 4 Frontier Preview
Best for sparse MoE reasoning (MMLU-Pro 78.4%), automated software engineering (SWE-bench Verified 52.6%), sub-byte serving efficiency.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Open weights (research license), Hugging Face, Meta AI
Important limitation: 400B total parameters across 128 routed experts with top-16 activation (35B active parameters per token). Trained natively in Microscaling FP4 (MXFP4) on 64,000 Blackwell B200 GPUs. Peak throughput requires OCP MXFP4-compliant silicon.
Check official model information↗ (opens in a new tab)V4 Pro
Best for long-context reasoning, coding and agent workflows.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: DeepSeek API, app and web
Important limitation: Official August 13 production revision DeepSeek-V4-Pro-0813. Prices shown are the current pre-August 17 tariff; announced peak/off-peak rates are future-dated. No independent 0813 benchmark package was verified by the edition cutoff.
Check official model information↗ (opens in a new tab)Claude Opus 5
Best for long-horizon reasoning, interactive agent work and complex professional tasks.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Anthropic API, Amazon Bedrock, Google Cloud and Microsoft Foundry
Important limitation: ARC-AGI-3 performance is benchmark-specific. Anthropic effort labels are provider-specific and should not be treated as standardized units of compute across vendors.
Check official model information↗ (opens in a new tab)GPT-5.6 Sol
Best for difficult planning, research and multi-step professional work.
Value for moneyProvisional premium-priced18/100 value indexCapability may be strong, but the comparable API price is high relative to other plotted models. The performance side uses a recent same-family proxy rather than the exact current model.
Where it is available: API and eligible ChatGPT plans
Important limitation: Launch comparisons are provider-published; confirm task-specific performance independently.
Check official model information↗ (opens in a new tab)GPT-5.6 Terra
Best for everyday writing, email drafting, analysis and balanced cost.
Value for moneyPremium-priced25/100 value indexCapability may be strong, but the comparable API price is high relative to other plotted models.
Where it is available: API and eligible ChatGPT plans
Important limitation: Use current API documentation for limits and regional availability.
Check official model information↗ (opens in a new tab)GPT-5.6 Luna
Best for quick drafts, summaries and routine high-volume tasks.
Value for moneyProvisional balanced41/100 value indexReasonable capability for the price, with tradeoffs against cheaper or stronger alternatives. The performance side uses a recent same-family proxy rather than the exact current model.
Where it is available: API and eligible ChatGPT plans
Important limitation: Lower price does not imply best performance for every workload.
Check official model information↗ (opens in a new tab)Claude Fable 5
Best for long documents, careful editing and sustained project work.
Value for moneyPremium-priced17/100 value indexCapability may be strong, but the comparable API price is high relative to other plotted models.
Where it is available: Claude and API where available
Important limitation: Anthropic documents additional safeguards and retention behavior for Mythos-class traffic.
Check official model information↗ (opens in a new tab)Grok 4.5
Best for coding, agent-style tasks and workflows that use current information.
Value for moneyBalanced43/100 value indexReasonable capability for the price, with tradeoffs against cheaper or stronger alternatives.
Where it is available: xAI API and Grok products
Important limitation: Requests exceeding 200K context use higher prices: $4 input and $12 output per 1M tokens.
Check official model information↗ (opens in a new tab)Gemini 3.5 Flash
Best for fast multimodal work, Google users and quick everyday assistance.
Value for moneyPrice verification neededNo stable comparable input-and-output API price is available in the current registry.
Where it is available: Gemini ecosystem and developer services
Important limitation: This edition did not capture a stable region-neutral official API price; verify before purchase.
Check official model information↗ (opens in a new tab)V4 Flash
Best for lower-cost long-document and technical workflows where regional access works.
Value for moneyProvisional outstanding value100/100 value indexHigh evidence-backed capability relative to the blended USD API price. The performance side uses a recent same-family proxy rather than the exact current model.
Where it is available: DeepSeek API through the explicit deepseek-v4-flash and deepseek-v4-pro model names; legacy aliases are past their retirement deadline.
Important limitation: DeepSeek primary documentation identifies April 24 as the V4 Preview API availability date. Later third-party score snapshots vary and should be read with their evaluation date and methodology.
Check official model information↗ (opens in a new tab)Qwen3.7-Max
Best for complex reasoning, coding and work inside the Alibaba Cloud ecosystem.
Value for moneyCompare in source currencyThe tracker does not convert CNY pricing into USD because exchange rates and regional billing change.
Where it is available: Alibaba Cloud Model Studio; regional endpoints vary
Important limitation: Pricing, endpoint names and availability differ between China and international Model Studio regions.
Check official model information↗ (opens in a new tab)Kimi K3
Best for very long documents, software work and extended agent-style tasks.
Value for moneyCompare in source currencyThe tracker does not convert CNY pricing into USD because exchange rates and regional billing change.
Where it is available: Kimi API platform; region and account eligibility vary
Important limitation: Prices are official CNY rates per million tokens and must not be displayed as USD.
Check official model information↗ (opens in a new tab)GLM-5.2
Best for long-running coding and autonomous workflow experiments.
Value for moneyCompare in source currencyThe tracker does not convert CNY pricing into USD because exchange rates and regional billing change.
Where it is available: Zhipu BigModel platform; regional access varies
Important limitation: Provider capability statements require independent task-specific reproduction.
Check official model information↗ (opens in a new tab)ERNIE 5.0
Best for Chinese-language and multimodal tasks in Baidu services.
Value for moneyPerformance evidence pendingThe model has a price, but no exact or recent same-family result in the six-category benchmark registry.
Where it is available: Baidu Qianfan and ERNIE services; international catalog availability varies
Important limitation: The displayed API price is from Baidu’s international Qianfan catalog; China-region billing differs.
Check official model information↗ (opens in a new tab)Doubao Seed 2.1
Best for production-oriented coding, agent and multimodal tasks in Volcengine.
Value for moneyCompare in source currencyThe tracker does not convert CNY pricing into USD because exchange rates and regional billing change.
Where it is available: Volcengine Ark; regional availability varies
Important limitation: Volcengine pricing uses tiered regional tables; verify the exact model ID and input-length tier.
Check official model information↗ (opens in a new tab)MiniMax-M3
Best for coding, reasoning and long-context work through compatible endpoints.
Value for moneyPrice verification neededNo stable comparable input-and-output API price is available in the current registry.
Where it is available: MiniMax API platform and compatible endpoints
Important limitation: Use the current regional pricing table and distinguish API pay-as-you-go from Token Plan subscriptions.
Check official model information↗ (opens in a new tab)Step 3.7 Flash
Best for lower-cost multimodal reasoning and computer-use-style tasks.
Value for moneyCompare in source currencyThe tracker does not convert CNY pricing into USD because exchange rates and regional billing change.
Where it is available: StepFun open platform; regional access varies
Important limitation: Official prices are CNY per million tokens. Confirm current quotas and model lifecycle notices.
Check official model information↗ (opens in a new tab)Hunyuan A13B
Best for efficient reasoning in Tencent Cloud and Chinese enterprise workflows.
Value for moneyCompare in source currencyThe tracker does not convert CNY pricing into USD because exchange rates and regional billing change.
Where it is available: Tencent Cloud Hunyuan / migration path to TokenHub
Important limitation: Tencent is migrating newer model access toward TokenHub; verify the active platform before integration.
Check official model information↗ (opens in a new tab)Qwen-Image-3.0
Best for images containing text, interfaces, diagrams and dense layouts.
Value for moneyPrice verification neededNo stable comparable input-and-output API price is available in the current registry.
Where it is available: Qwen and Alibaba Cloud ecosystem; verify regional endpoint availability
Important limitation: Capabilities and examples are provider-published; independent benchmark reproduction is pending.
Check official model information↗ (opens in a new tab)Qwen-Audio-3.0-TTS Flash / Plus
Best for real-time or higher-fidelity text-to-speech generation.
Value for moneyPrice verification neededNo stable comparable input-and-output API price is available in the current registry.
Where it is available: Alibaba Cloud Model Studio; rollout and regional availability may vary
Important limitation: The 300ms-level first-packet latency and multilingual results are provider-reported; verify independently for production workloads.
Check official model information↗ (opens in a new tab)Llama 4 Scout
Best for very long context and multimodal local or private-cloud deployment.
Value for moneyLocal test requiredThe weights are free to obtain, but hardware, power, hosting and operations determine total cost.
Local deployment: Meta says an Int4 version fits on one NVIDIA H100 GPU.
Where it is available: Downloadable open weights through Meta and ecosystem partners
Important limitation: The weights have no per-token license fee, but a practical deployment still needs GPU capacity, storage, power, security and operations.
Check official model information↗ (opens in a new tab)Llama 4 Maverick
Best for higher-capability multimodal workloads with open-weight control.
Value for moneyLocal test requiredThe weights are free to obtain, but hardware, power, hosting and operations determine total cost.
Local deployment: Meta describes deployment on a single H100 DGX host or distributed infrastructure.
Where it is available: Downloadable open weights through Meta and ecosystem partners
Important limitation: Maverick stores 400B total parameters even though 17B are active per token, so hosting is materially heavier than the active-parameter figure alone suggests.
Check official model information↗ (opens in a new tab)gpt-oss-20b
Best for local reasoning on constrained hardware and rapid private experimentation.
Value for moneyLocal test requiredThe weights are free to obtain, but hardware, power, hosting and operations determine total cost.
Local deployment: OpenAI says the model can run with about 16 GB of memory.
Where it is available: Free downloadable weights; self-hosted or third-party hosted
Important limitation: OpenAI does not serve this model through ChatGPT or the OpenAI API. The operator pays for compute, storage, monitoring and maintenance.
Check official model information↗ (opens in a new tab)gpt-oss-120b
Best for high-capacity private reasoning and agent workflows.
Value for moneyLocal test requiredThe weights are free to obtain, but hardware, power, hosting and operations determine total cost.
Local deployment: OpenAI says the model fits on one 80 GB GPU.
Where it is available: Free downloadable weights; self-hosted or third-party hosted
Important limitation: The weights are free, but OpenAI says the model needs about 80 GB of memory; total deployment cost depends on utilization and operations.
Check official model information↗ (opens in a new tab)Ministral 3 14B
Best for smaller local enterprise assistant with multimodal support.
Value for moneyLocal test requiredThe weights are free to obtain, but hardware, power, hosting and operations determine total cost.
Local deployment: Dense 14B model; actual memory depends on precision, quantization and runtime.
Where it is available: Downloadable weights and hosted ecosystem endpoints
Important limitation: A free model license does not make production inference free. Quantization, hardware, throughput and support requirements determine cost.
Check official model information↗ (opens in a new tab)Mistral Large 3
Best for frontier-class open-weight enterprise customization.
Value for moneyLocal test requiredThe weights are free to obtain, but hardware, power, hosting and operations determine total cost.
Local deployment: Data-center-class sparse model; deployment normally requires substantial accelerator infrastructure.
Where it is available: Downloadable weights and hosted ecosystem endpoints
Important limitation: The model has 675B total parameters and 41B active parameters; storage and serving requirements remain substantial despite sparse activation.
Check official model information↗ (opens in a new tab)Gemma 3 27B
Best for multilingual local assistant, document and image workflows.
Value for moneyLocal test requiredThe weights are free to obtain, but hardware, power, hosting and operations determine total cost.
Local deployment: Dense 27B model; actual memory depends on precision, quantization and runtime.
Where it is available: Downloadable open model through Google and ecosystem partners
Important limitation: The model is free to obtain under Google’s terms, but businesses must still budget for hardware, inference software, monitoring and updates.
Check official model information↗ (opens in a new tab)Nemotron 3 Super 120B-A12B
Best for efficient long-running multi-agent and code workflows.
Value for moneyPotentially strong locallyThe weights are free to obtain, but hardware, power, hosting and operations determine total cost.
Local deployment: 120B total / 12B active sparse model; optimized for NVIDIA Blackwell and data-center deployment.
Where it is available: Open weights, datasets and recipes; self-hosting and NVIDIA ecosystem deployment
Important limitation: NVIDIA optimizes the model for its accelerator stack. Cost and throughput can differ substantially on other hardware and runtimes.
Check official model information↗ (opens in a new tab)