Compare measured token efficiency, a fixed API cost scenario, and the evidence behind model decisions. Every board shows its inputs, sources, and limits.
7 tokenizer families across 3 reproducible samples.
View token board Fixed workload39 source-tracked models for 1M input + 300K output tokens per month.
View cost board Evidence ledgerSource coverage is shown without turning uneven reports into a fake quality score.
Inspect the evidenceLower measured token count for the same text means lower token volume in this sample set. It does not predict model quality.
| Rank | Tokenizer | Samples | Total tokens | Tokens / 100 chars | Basis |
|---|---|---|---|---|---|
| 1 | OpenAI Current | 3 | 114 | 27.2 | o200k_base |
| 2 | DeepSeek V4 Pro | 3 | 117 | 27.9 | Pinned tokenizer JSON |
| 3 | Llama 3 | 3 | 118 | 28.2 | Pinned tokenizer JSON |
| 4 | OpenAI Legacy | 3 | 121 | 28.9 | cl100k_base |
| 5 | Qwen 3.5 | 3 | 129 | 30.8 | Pinned tokenizer JSON |
| 6 | Mistral Medium 3.5 | 3 | 132 | 31.5 | Pinned tokenizer JSON |
| 7 | Gemma 3 | 3 | 135 | 32.2 | Pinned SentencePiece model |
Checked 2026-07-19. Plain text only; no chat templates, roles, tools, images, or API wrappers. The exact source samples and per-sample results are available in the tokenizer comparison.
1M input + 300K output tokens per month. Cost only; no equal-quality assumption.
| Rank | Model | Input / 1M | Output / 1M | Workload cost | Checked |
|---|---|---|---|---|---|
| 1 | GPT-OSS 20B on GroqGroq | $0.07 | $0.30 | $0.16 | 2026-07-15 |
| 2 | Llama 4 Scout on GroqGroq | $0.11 | $0.34 | $0.21 | 2026-07-15 |
| 3 | Gemini 2.5 Flash-LiteGoogle | $0.10 | $0.40 | $0.22 | 2026-08-05 |
| 4 | DeepSeek V4 Flash non-thinkingDeepSeek | $0.14 | $0.28 | $0.22 | 2026-08-05 |
| 5 | Command R 08-2024Cohere | $0.15 | $0.60 | $0.33 | 2026-07-17 |
| 6 | GPT-4o miniOpenAI | $0.15 | $0.60 | $0.33 | 2026-08-05 |
| 7 | GPT-OSS 120B on GroqGroq | $0.15 | $0.60 | $0.33 | 2026-07-15 |
| 8 | Qwen3-32B on GroqGroq | $0.29 | $0.59 | $0.47 | 2026-07-15 |
| 9 | GPT-5.6 LunaOpenAI | $0.20 | $1.20 | $0.56 | 2026-08-05 |
| 10 | gpt-5.4-nanoOpenAI | $0.20 | $1.25 | $0.58 | 2026-08-05 |
| 11 | DeepSeek V4 ProDeepSeek | $0.43 | $0.87 | $0.70 | 2026-08-05 |
| 12 | Gemini 3.1 Flash-LiteGoogle | $0.25 | $1.50 | $0.70 | 2026-08-05 |
| 13 | Gemini 3.5 Live Translate PreviewGoogle | $0.25 | $1.50 | $0.70 | 2026-08-05 |
| 14 | Gemini 2.5 FlashGoogle | $0.30 | $2.50 | $1.05 | 2026-08-05 |
| 15 | Gemini 3.5 Flash-LiteGoogle | $0.30 | $2.50 | $1.05 | 2026-08-05 |
| 16 | Gemini 3 Flash PreviewGoogle | $0.50 | $3.00 | $1.40 | 2026-08-05 |
| 17 | gpt-5.4-miniOpenAI | $0.75 | $4.50 | $2.10 | 2026-08-05 |
| 18 | Kimi K2.6Moonshot AI | $0.95 | $4.00 | $2.15 | 2026-07-17 |
| 19 | Kimi K2.7 CodeMoonshot AI | $0.95 | $4.00 | $2.15 | 2026-07-17 |
| 20 | Claude Haiku 4.5Anthropic | $1.00 | $5.00 | $2.50 | 2026-08-05 |
| 21 | Gemini 3.6 FlashGoogle | $1.50 | $7.50 | $3.75 | 2026-08-05 |
| 22 | Mistral Medium 3.5Mistral | $1.50 | $7.50 | $3.75 | 2026-07-17 |
| 23 | Grok 4.5xAI | $2.00 | $6.00 | $3.80 | 2026-07-17 |
| 24 | Gemini 3.5 FlashGoogle | $1.50 | $9.00 | $4.20 | 2026-08-05 |
| 25 | Gemini 2.5 ProGoogle | $1.25 | $10.00 | $4.25 | 2026-08-05 |
| 26 | Kimi K2.7 Code High-SpeedMoonshot AI | $1.90 | $8.00 | $4.30 | 2026-07-17 |
| 27 | Claude Sonnet 5Anthropic | $2.00 | $10.00 | $5.00 | 2026-08-05 |
| 28 | Command ACohere | $2.50 | $10.00 | $5.50 | 2026-07-17 |
| 29 | Command R+ 08-2024Cohere | $2.50 | $10.00 | $5.50 | 2026-07-17 |
| 30 | GPT-4oOpenAI | $2.50 | $10.00 | $5.50 | 2026-08-05 |
This is a billing comparison for one reference workload, not a quality or “best model” ranking. Change the assumptions in the API cost calculator before making a routing decision.
These rows show how much evidence is documented for each model. More sources do not prove better quality.
| Model | Official facts | Community reports | Controlled StackLens test | Last reviewed |
|---|---|---|---|---|
| Claude Fable 5 | 2 sources | 5 sources | Not run | 2026-07-13 |
| Claude Sonnet 5 | 2 sources | 5 sources | Not run | 2026-07-13 |
| Gemini 3.5 Flash | 2 sources | 5 sources | Not run | 2026-07-13 |
| GPT-5.6 Sol | 1 source | 5 sources | Not run | 2026-07-13 |
| GPT-5.6 Terra | 1 source | 3 sources | Not run | 2026-07-13 |
| GPT-5.6 Luna | 1 source | 3 sources | Not run | 2026-07-13 |
Official documentation, user reports, and controlled tests are not interchangeable. Read the methodology before using any signal in a production decision.
StackLens does not currently publish a weighted quality score. Public leaderboards use different prompts, dates, models, and scoring rules, so combining them into one number would create false precision.
Find a model for a taskOpen a focused board when you need one dimension, then validate it against your own workload and quality bar.
No. The cost board compares token charges for one fixed workload and does not measure quality, latency, retries, or completed-task success.
The tokenizer board sums measured token counts across the published English, Chinese, JSON, and code samples. The exact text, tokenizer basis, source, and check date remain visible.
The available public sources do not use one shared controlled task set. StackLens keeps official facts, community reports, and independent tests separate rather than inventing a combined quality score.