AI Agent Store/Model intelligenceBuild an AI worker ↗THE MODEL SELECTION SERIES
Open weights let you obtain the trained parameters, subject to the model license. That can create more deployment options, but it does not automatically make a model free to run, fully open source or unrestricted for commercial use.
Snapshot · LiveBench question set: 25 June 2026 · View every source ↗
The comparison uses hosted API prices and advertised endpoint context limits. It does not estimate the GPU memory, utilization, networking or operations cost of self-hosting. The same weights can behave differently under quantization or a different serving stack.
Review redistribution, commercial-use and acceptable-use conditions on the model’s own release page. A family name is not enough: parameter count, version, quantization and thinking settings can all change the comparison.
A provider advertising tool support does not guarantee your local server implements the same protocol. Validate schema adherence, parallel tool calls, error handling and context limits with the deployment you plan to use.
Hosted prices are a useful starting point. For self-hosting, calculate utilization and total monthly infrastructure cost, including idle capacity and operational time. Benchmark throughput on your actual hardware before treating a theoretical token price as achievable.
The comparison starts with open-weight models. Prices represent hosted API endpoints; check the model license before self-hosting.
YOUR WORKLOAD. YOUR PRIORITIES.
Standard text API estimate. Include reasoning tokens in output.
A scenario estimate, not a measured task bill. Excludes cache discounts, tool fees, retries and provider premiums. Workloads exceeding a catalog context limit receive no cost or Value rating. Advertised context is not a guarantee of retrieval quality.
Practical Value Our blend of Agent Fit and affordability: 70% capability + 30% cost percentile by default. Cost uses your token workload. A relative score within this 56-model snapshot, not a claim of dollars saved. · General agent
70% Agent Fit + 30% affordability. Rankings use your workload; filters do not change the reference cohort.
THE EVIDENCE, SIDE BY SIDE
| Compare | Our task-weighted mix of LiveBench category scores. A useful shortlist signal, not a measured agent success rate. Change the task profile to change the weights. | Our blend of Agent Fit and affordability: 70% capability + 30% cost percentile by default. Cost uses your token workload. A relative score within this 56-model snapshot, not a claim of dollars saved. | Equal-weight average of LiveBench and Arena percentile ranks within the same nine matched configurations. Models without both results receive no score. This small cohort does not rank the whole market. | Automatically checked tasks in seven areas: reasoning, coding, agentic coding, mathematics, data analysis, language and instruction following. Overall is the equal-weight category mean. All rows use the 2026-06-25 question set. | People compare answers without knowing which model wrote them. A higher rating means more preferred answers, not a percent correct. Published ± ranges can overlap; small score gaps may not mean a real difference. | Estimated text API bill: requests × (input tokens × input price + output tokens × output price) ÷ 1,000,000. Include billed reasoning in output tokens. Excludes caching, tools, retries, taxes, hosting and long-context premiums. | The advertised token capacity for the prompt, conversation and response. This is a size limit, not proof the model can reliably use every detail. Catalog endpoints may have different limits. | Input / outputUSD / 1M tokens | |
|---|---|---|---|---|---|---|---|---|---|
01 DeepSeek V4.1 Flash ↗DeepSeek · maxOpen weights | 78.9 | 83.5 | — | 81.1 | — | $2.14 | 1.05M | $0.094 / $0.6OpenRouter catalog | |
02 Kimi K3 ↗Moonshot AI · source defaultOpen weights | 77.4 | 60.7 | — | 79.2 | — | $60 | 1.05M | $3 / $15OpenRouter catalog | |
03 Qwen3.8 Flash Next ↗Alibaba · source defaultOpen weights | 76.2 | — | — | 76.2 | — | — | Not verified | Not verified | |
04 DeepSeek V4 Flash Vision Exp ↗DeepSeek · experimental · source defaultOpen weightsPreview | 75.1 | 79.8 | — | 76.8 | — | $3.52 | 1.05M | $0.22 / $0.66OpenRouter catalog | |
05 GLM-5.3 ↗Z.ai · source defaultOpen weights | 73.7 | 71.4 | — | 76.1 | — | $13.68 | 1.31M | $0.84 / $2.64OpenRouter catalog | |
06 Qwen3.8 27B ↗Alibaba · source defaultOpen weights | 73.5 | 74.7 | — | 75.3 | — | $10.2 | 1.00M | $0.42 / $3OpenRouter catalog | |
07 DeepSeek V4 Pro 0813 ↗DeepSeek · source defaultOpen weights | 73.3 | 73.9 | — | 77.4 | — | $10.56 | 1.05M | $0.66 / $1.98OpenRouter catalog | |
08 DeepSeek V4 Flash 0731 ↗DeepSeek · source defaultOpen weights | 71.1 | 79.2 | — | 74.2 | — | $1.68 | 1.31M | $0.04 / $0.64OpenRouter catalog | |
09 Inkling ↗Thinking Machines · xhighOpen weights | 68.9 | 64.1 | — | 71.9 | — | $18.1 | 1.05M | $1 / $4.05OpenRouter catalog | |
10 GLM-5.2 ↗Z.ai · source defaultOpen weights | 68.5 | 70.1 | — | 73.2 | — | $10.58 | 1.05M | $0.6496 / $2.04OpenRouter catalog | |
11 GLM-5.3 Flash ↗Z.ai · source defaultOpen weights | 67.2 | 74.8 | — | 71.6 | — | $2.5 | 1.31M | $0.15 / $0.5OpenRouter catalog | |
12 DeepSeek V4 Pro ↗DeepSeek · source defaultOpen weights | 67.1 | 67.9 | — | 71.6 | — | $13.37 | 1.05M | $0.9553 / $1.91OpenRouter catalog | |
13 Kimi K2.6 ↗Moonshot AI · thinkingOpen weights | 66.9 | 63.5 | — | 70.5 | — | $17.5 | 262K | $0.95 / $4OpenRouter catalog | |
14 Kimi K2.7 Code ↗Moonshot AI · source defaultOpen weights | 64.8 | 65.8 | — | 68.4 | — | $13.66 | 262K | $0.7062 / $3.3OpenRouter catalog | |
15 Nemotron 3 Ultra ↗NVIDIA · 550B A55B · source defaultOpen weights | 63.8 | 66.1 | — | 67.4 | — | $10.8 | 262K | $0.6 / $2.4OpenRouter catalog | |
16 DeepSeek V4 Flash ↗DeepSeek · source defaultOpen weights | 61.6 | 73.1 | — | 65.5 | — | $1.24 | 1.05M | $0.0886 / $0.1772OpenRouter catalog | |
17 Qwen3.6 27B ↗Alibaba · source defaultOpen weights | 60.0 | 65.8 | — | 64.0 | — | $8.6 | 262K | $0.32 / $2.7OpenRouter catalog |
— means missing comparable evidence, never zero. Consensus covers only nine matched configurations. Model IDs, reasoning settings, source links and pricing differences are available on every model page. Open weights does not imply unrestricted commercial use.
LOOK BEYOND ONE NUMBER
| Measure | Claude Opus 5.5max | GPT-6 Solmax | DeepSeek V4.1 Flashmax |
|---|---|---|---|
| Agent FitOur task-weighted mix of LiveBench category scores. A useful shortlist signal, not a measured agent success rate. Change the task profile to change the weights. | 79.4 | 74.7 | 78.9 |
| Practical ValueOur blend of Agent Fit and affordability: 70% capability + 30% cost percentile by default. Cost uses your token workload. A relative score within this 56-model snapshot, not a claim of dollars saved. | 61.0 | 62.7 | 83.5 |
| Evidence ConsensusEqual-weight average of LiveBench and Arena percentile ranks within the same nine matched configurations. Models without both results receive no score. This small cohort does not rank the whole market. | Insufficient evidence | Insufficient evidence | Insufficient evidence |
| Workload costEstimated text API bill: requests × (input tokens × input price + output tokens × output price) ÷ 1,000,000. Include billed reasoning in output tokens. Excludes caching, tools, retries, taxes, hosting and long-context premiums. | $80 | $40 | $2.14 |
| Context windowThe advertised token capacity for the prompt, conversation and response. This is a size limit, not proof the model can reliably use every detail. Catalog endpoints may have different limits. | 1.00M | 1.05M | 1.05M |
| Tool callingThe catalog declares support for returning structured tool calls. Support does not measure whether the model picks the right tool or uses it correctly. | Supported | Supported | Supported |
| ReasoningCan it work through constraints and connect clues? LiveBench uses spatial, navigation, perspective-taking and logic-puzzle tasks. Useful for planning, but not a direct test of long-running agents. | 92.2 | 88.7 | 86.7 |
| CodingCan it write or complete code that passes tests? These are contained programming problems, not entire software projects. | 89.3 | 81.8 | 80.0 |
| Agentic codingCan it edit code in a tool-using workflow? LiveBench tests JavaScript, TypeScript and Python tasks. Results depend on the benchmark harness as well as the model. | 71.7 | 52.9 | 77.3 |
| MathematicsCan it solve difficult quantitative problems with checkable answers? Strong math is useful evidence of reasoning, but does not guarantee better writing or tool use. | 97.1 | 96.4 | 93.3 |
| Data analysisCan it join and reformat tables and reason about event sequences? Useful for agents that process structured business information. | 80.3 | 81.2 | 79.3 |
| LanguageCan it interpret word relationships, reconstruct plots and correct typos? This is a narrow language test, not a full measure of writing quality or multilingual coverage. | 86.3 | 85.3 | 81.2 |
| Instruction followingCan it follow requested constraints while rewriting, simplifying, summarizing and composing text? Relevant to agents that must return a specific format. | 65.7 | 68.6 | 70.0 |
| Input / output per 1M | $4 / $20 | $2 / $10 | $0.094 / $0.6 |