Paste a prompt or attach PDF, DOCX, code, data files, and images. Provider results and costs update live without uploading your content.
Text, code, PDF, DOCX, PNG, JPEG, WebP, or GIF · 12 MB per file
Choose a provider card to inspect every available model.
OpenAI text uses exact local o200k_base BPE tokenization. Claude, Gemini, and DeepSeek use deterministic Provider-Calibrated UTF-8 Token Projections. Supported images follow each provider’s published visual-token formulas, including patch, tile, detail, and model-cap rules.
Every result identifies its method and accuracy level. Pricing is refreshed from Models.dev with the bundled registry available as an offline fallback.
Read the complete methodology →This token calculator measures text prompts, code, structured data, readable documents, and supported images against 35 current models from 9 providers: OpenAI, Anthropic, Google, DeepSeek, Qwen, Moonshot Kimi, xAI Grok, Z.ai GLM, Meta.
It reports input tokens, context-window fit, and estimated input cost. Text and files stay in the browser; only public model metadata may be refreshed from Models.dev.
Rates verified 2026-09-09
Cost per 1,000,000 tokens in USD, for input and output separately. Cached input is the discounted rate providers charge when a prompt prefix is reused. Paste a real prompt into the calculator above to turn these rates into a cost for your own workload.
| Model | Provider | Context window | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|---|---|
| GPT-6 Astra | OpenAI | 1,050,000 | $10 | $1 | $50 |
| GPT-5.6 Luna | OpenAI | 1,050,000 | $0.2 | $0.02 | $1.2 |
| GPT-5.6 Sol | OpenAI | 1,050,000 | $4 | $0.4 | $20 |
| GPT-5.6 Terra | OpenAI | 1,050,000 | $2 | $0.2 | $12 |
| Claude Fable 5.1 | Anthropic | 1,000,000 | $10 | $0.25 | $50 |
| Claude Opus 5 | Anthropic | 1,000,000 | $5 | $0.5 | $25 |
| Claude Sonnet 5 | Anthropic | 1,000,000 | $2 | $0.2 | $10 |
| Claude Opus 4.8 | Anthropic | 1,000,000 | $5 | $0.5 | $25 |
| Claude Haiku 4.5 | Anthropic | 200,000 | $1 | $0.1 | $5 |
| Gemini 3.8 Flash | 1,048,576 | $0.75 | $0.075 | $3.75 | |
| Gemini 3.7 Flash | 1,048,576 | $0.75 | $0.075 | $3.75 | |
| Gemini 3.5 Flash Lite | 1,048,576 | $0.3 | $0.03 | $2.5 | |
| Gemini 3.1 Flash Lite | 1,048,576 | $0.25 | $0.025 | $1.5 | |
| Gemini 2.5 Pro | 1,048,576 | $1.25 | $0.125 | $10 | |
| DeepSeek V4 Pro | DeepSeek | 1,000,000 | $0.435 | $0.003625 | $0.87 |
| DeepSeek V4 Flash | DeepSeek | 1,000,000 | $0.14 | $0.0028 | $0.28 |
| Qwen3.8 Flash | Qwen | 1,000,000 | $0.15 | $0.016 | $0.47 |
| Qwen3.8 Max | Qwen | 1,000,000 | $2 | $0.25 | $6 |
| Qwen3.7 Plus | Qwen | 1,000,000 | $0.5 | $0.05 | $3 |
| Qwen3.7 Max | Qwen | 1,000,000 | $2.5 | $0.5 | $7.5 |
| Qwen3.6 Flash | Qwen | 1,000,000 | $0.1875 | — | $1.125 |
| Kimi K3 | Moonshot Kimi | 1,048,576 | $3 | $0.3 | $15 |
| Kimi K2.7 Code | Moonshot Kimi | 262,144 | $0.95 | $0.19 | $4 |
| Kimi K2.6 | Moonshot Kimi | 262,144 | $0.95 | $0.16 | $4 |
| Grok 4.6 | xAI Grok | 500,000 | $2 | $0.5 | $6 |
| Grok 4.5 | xAI Grok | 500,000 | $2 | $0.3 | $6 |
| Grok 4.3 | xAI Grok | 1,000,000 | $1.25 | $0.2 | $2.5 |
| GLM-5.3-Flash | Z.ai GLM | 1,000,000 | $0.075 | $0.015 | $0.25 |
| GLM-5.3 | Z.ai GLM | 1,000,000 | $1.4 | $0.26 | $4.4 |
| GLM-5.2 | Z.ai GLM | 1,000,000 | $1.4 | $0.26 | $4.4 |
| GLM-5.1 | Z.ai GLM | 200,000 | $1.4 | $0.26 | $4.4 |
| GLM-5-Turbo | Z.ai GLM | 200,000 | $1.2 | $0.24 | $4 |
| Muse Spark 1.3 | Meta | 1,048,576 | $1.25 | $0.15 | $4.25 |
| Muse Spark 1.2 | Meta | 1,048,576 | $1.25 | $0.15 | $4.25 |
| Muse Spark 1.1 | Meta | 1,048,576 | $1.25 | $0.15 | $4.25 |
Rates come from the public model catalogue and refresh from the same live source the calculator uses. Long-context tiers, batch discounts, and provider-side overheads can change a real invoice. Provider documentation remains authoritative.
Computation provenance remains visible, distinguishing exact BPE tokenization from deterministic provider-calibrated projection.
Upload a text-based PDF or DOCX document instead of copying its contents page by page. TokenCalculator.dev extracts readable text inside the browser, combines it with anything in the prompt editor, and immediately recalculates every provider result.
Plain text, Markdown, CSV, JSON, JavaScript, TypeScript, Python, Go, Rust, Java, C/C++, HTML, CSS, XML, YAML, TOML, SQL, and shell files are also supported up to 12 MB per file. Scanned PDFs need OCR and are intentionally rejected rather than returning a misleading zero-token result.
Privacy: document contents never leave the device. No prompt or uploaded file is sent to a token-counting server.
Add a PNG, JPEG, WebP, or GIF to compare its visual-token footprint across supported OpenAI, Claude, and Gemini models. The calculator reads width and height locally and applies the selected provider’s documented tile, patch, or media-token rule.
Image tokens are added to extracted document text and pasted prompt text, so a multimodal request can be planned as one workload. If a selected model has no documented vision rule, the result says image unavailable instead of presenting a fabricated estimate.
This is an image-token and cost estimate, not OCR. Pixel data is not uploaded, retained, or used to extract words from an image.
Representative planning workflows now; attributable developer reviews are added only after their wording and permission are verified.
Extract a text-based PDF locally, measure its complete payload, and compare context fit and input cost across OpenAI, Claude, Gemini, and DeepSeek.
Combine image dimensions with prompt text, then apply documented tile, patch, or media-token rules for each supported vision model.
Attach code, configuration, and documentation files as one workload before selecting a model with enough context capacity.
Measure structured data and instructions together to compare provider input rates before committing a recurring workload.
Tell us what you measured, which providers you compared, and what decision the result helped you make. Published reviews include attribution and require explicit permission.
Registry reviewed 2026-09-09
| Provider | Text method | Image method | Result label |
|---|---|---|---|
| OpenAI | Local model BPE | Published tile rules | Exact BPE / provider formula |
| Claude | Provider-Calibrated UTF-8 Projection | Published patch rules | Deterministic projection |
| Gemini | Provider-Calibrated UTF-8 Projection | Published media rules | Deterministic projection |
| DeepSeek | Provider-Calibrated UTF-8 Projection | Unavailable when undocumented | Deterministic projection |
Yes. Attach PDF, DOCX, code, data, PNG, JPEG, WebP, or GIF files. Text extraction, image-dimension reading, and token measurement run locally in your browser.
No. Your text, documents, code, and images stay in your browser and are not sent to TokenCalculator.dev or another calculation server.
A token is a small unit of content processed by a language model. It may be a whole word, part of a word, punctuation, whitespace, or—in multimodal models—a unit assigned to an image. Different providers can split the same input differently.
Input cost is calculated as input tokens divided by one million, multiplied by the selected model’s input price per million tokens. Provider tiers, cached input, tools, and message wrappers can affect final billing.
No. Text, PDF and DOCX extraction, image-dimension reading, and token calculations run in your browser. Zero prompts, document contents, or uploaded images are sent to TokenCalculator.dev or another calculation server.
OpenAI text counts use the matching local BPE encoding. Other providers are clearly labeled estimates unless they publish a deterministic local formula. Image counts use provider-published rules.
Attach a text-based PDF or DOCX file and the browser extracts its readable text locally, then counts it with every selected provider model. Scanned or encrypted PDFs require OCR and are not supported.
Yes. Add a PNG, JPEG, WebP, or GIF. For supported vision models, the calculator reads its pixel dimensions locally and applies the provider-published image token formula. It reports unavailable instead of inventing a value when a provider has no documented rule.
Providers can add message wrappers, system instructions, tools, or model-specific processing. The accuracy badge and methodology explain what this calculator includes.
