Best overall
Closed model
Anthropic
- Intelligence
- 63.1
- Speed
- 58.8tok/s
- Input / output
- $5 / $25
- Context
- 1M
Best fit
State-of-the-art reasoning and complex agentic workflows
AI · LEADERBOARD
Escolha o modelo certo para o trabalho.
Uma comparação prática dos modelos de IA que equipes podem implantar agora.
Snapshot: 16 models compared
Independent editorial snapshot · not live platform telemetry
Dated editorial snapshot
Review target: weekly and after major releases. Rankings reflect this dated snapshot. OpenRouter usage below has its own source timestamp.
The short answer
Three leaders, three different decisions.
Best open model
Open weights
Kimi
- Intelligence
- 59.7
- Speed
- 35tok/s
- Input / output
- $3 / $15
- Context
- 1.05M
Best fit
Sovereign, open-weight deployments needing maximum intelligence
Best value
Open weights
DeepSeek
- Intelligence
- 51.8
- Speed
- 115tok/s
- Input / output
- $0.14 / $0.28
- Context
- 1.31M
Best fit
Cost-sensitive open inference and coding workloads
OpenRouter usage
OpenRouter LLM usage ranking
Top models by prompt, completion, and reasoning tokens processed on OpenRouter. Showing the first 20 only.
Top 20OpenRouter data · cached hourlyUpstream · Oct 20, 2026
- 1.DeepSeek V4.1 Flash (batch)by DeepSeek27.3T tokens32%
- 2.GLM 5.3 Flash (batch)by Z.ai11.6T tokens10%
- 3.DeepSeek V4.1 Flash 1017by DeepSeek11.5T tokensnew
- 4.MiMo-V2.6-Flashby Xiaomi9.05T tokens22%
- 5.Hy4 previewby Tencent8.16T tokens0%
- 6.Claude Opus 5.5 (batch)by Anthropic6.8T tokens21%
- 7.GPT-6 Luna (batch)by OpenAI6.21T tokens13%
- 8.Step 5 Previewby StepFun5.82T tokens61%
- 9.GLM 5.3 (batch)by Z.ai5.32T tokens27%
- 10.MiniMax M3.1 Flash (free)by MiniMax4.97T tokensnew
- 11.DeepSeek V4 Flash 0731by DeepSeek4.71T tokens7%
- 12.Nemotron 3 Ultra (free)by Nvidia4.5T tokens39%
- 13.Muse Spark 1.3 Contributorby Meta3.26T tokens17%
- 14.Hy3by Tencent3.15T tokens4%
- 15.DeepSeek V4 Flash 0423by DeepSeek2.77T tokens13%
- 16.Claude Sonnet 5.5 (batch)by Anthropic2.47T tokens7%
- 17.Claude Haiku 5.5 (batch)by Anthropic1.96T tokens14%
- 18.GPT-6.1 Sol (batch)by OpenAI1.89T tokens18%
- 19.Kimi K3 (batch)by Moonshot AI1.83T tokens28%
- 20.GPT-5.6 Luna (batch)by OpenAI1.69T tokens6%
From benchmark to decision
Which model for which job?
Nine recurring production scenarios, with the trade-offs that matter before a team commits to a model.
Selected use case
01 / 09
Decision view
Customer support
Why it fits
- 01High measured throughput
- 02Low per-token cost
- 03Strong intelligence for support
How the ranking works
Transparent enough to disagree with.
The ranking combines public benchmark signals with price, measured throughput, context and deployment constraints. It is a decision aid, not a universal definition of model quality.
- Dataset snapshot
- 25 Aug 2026 · 10:41 TST
- Review cadence
- Revalidated weekly and after major releases
Value score
(intelligence − 25)² × min(1, speed ÷ 30) ÷ (input cost + output cost × 3)
Output is weighted ×3 to reflect output-heavy production use; intelligence is squared so cheap but weak models do not dominate.
- Intelligence
- Public benchmark and reasoning signals form the default order.
- Speed
- Output tokens per second; provider and region still affect real latency.
- Cost
- Public input and output price per million tokens, before enterprise discounts.
- Deployment
- Context, weight access and hosting constraints determine production fit.
Public references
Prices and performance can change without notice. Validate the finalists against your own prompts, infrastructure and risk requirements before procurement.
FAQ
Questions worth asking before the benchmark wins the argument.
The first row and Best overall card show the current intelligence-first leader. The best choice can change once price, latency, hosting and task-specific evaluation enter the decision.
The Best open model card is recalculated from models with verified downloadable weights, using the same intelligence-first ordering as the main table.
The Best value card is recalculated from the published value formula whenever verified intelligence, speed or base token pricing changes.
No. Context capacity only describes how much can be submitted. Retrieval design, instruction quality, attention behaviour and evaluation still determine whether that context produces a useful answer.
A scheduled pipeline revalidates the dataset weekly and after major releases. Failed or conflicting source checks keep the last verified snapshot online instead of publishing partial data.
The default order uses the public intelligence signal. Speed, price, context, openness and a separate value formula remain visible so teams can challenge the default order.
Tenten does not sell model tokens through this page. Public sources are linked, assumptions are shown, and we recommend a task-specific evaluation before any production decision.
TENTEN AI · MODEL EVALUATION
The leaderboard is the start
Bring the workflow. We’ll test the shortlist.
A model only becomes a good decision after it survives your documents, permissions, latency budget and failure cases.