AI Token Counter & LLM Cost Calculator

The AI Token Counter & LLM Cost Calculator is a free, privacy-first tool for estimating token usage and API cost across the most-used large language models in 2026 — including GPT-5, GPT-5 Mini, Claude Opus 4.7, Claude Sonnet 4.6, Gemini 2.5 Pro, Gemini 2.5 Flash, Llama 4, DeepSeek V3.1 and more. It is role-aware (system / user / assistant / tool), understands prompt caching and reasoning "thinking" tokens, and computes per-request and monthly costs against published per-million-token prices. Everything runs in your browser — no server round-trips, no signup, no telemetry on your prompts — so sensitive data stays on your device.

• Role-aware prompt composer (system / user / assistant / tool) with per-message token counts — exact o200k per block for OpenAI when the tokenizer is loaded
• Compare 30+ frontier and open-source LLMs from OpenAI, Anthropic, Google, Meta, DeepSeek, Mistral, xAI, Cohere and Alibaba
• Prompt-caching-aware math for Anthropic, OpenAI and Gemini cached-input pricing
• Reasoning / "thinking" token toggle with a tunable ratio for GPT-5-thinking, Claude extended thinking, o-series and DeepSeek R1
• Per-request and per-month cost columns, sortable by total cost, input price, or context size
• Context-window fit meter with green / amber / red warnings at 70% and 90% utilisation
• Quick presets — Flagship, Cheapest, Best for coding, Best for reasoning, Long context (≥ 500k)
• Drag-and-drop or file upload for .txt, .md, .json, .csv, .ts, .js, .py, .html, .xml, .yaml
• Shareable permalink that encodes the entire scenario (no account, no database)
• One-click CSV export for the per-model cost breakdown
• Zero network calls with your prompts — every token count and price calculation runs locally
• Open, versioned pricing dataset you can inspect, copy and send pull requests against

• Estimate monthly API cost before deploying an AI feature to production
• Pick the most cost-effective model for a given prompt across 30+ providers
• Size prompts against each model’s context window to avoid truncation
• Model the savings from turning on prompt caching for long system prompts
• Compare the true cost of reasoning models (GPT-5-thinking, o-series, Claude extended thinking, DeepSeek R1) when thinking tokens dominate output
• Build a per-customer pricing sheet for an AI SaaS feature before you launch
• Share a costed scenario with your team via a single URL — no login required
• Teach engineers how tokenization, pricing tiers and context limits actually work

• OpenAI GPT-family totals use the real o200k BPE (gpt-tokenizer) — same encoding as GPT-4o / GPT-5–class chat APIs for your pasted text. Claude, Gemini and Llama still use calibrated estimates (±~5%). Reconcile billing in each dashboard.
• Prompt caching can cut input cost by up to 90% for long, repeated system prompts. Set the "cached input fraction" slider to a realistic cache-hit ratio (0.5 – 0.9 for production chat agents).
• Reasoning models bill their hidden "thinking" stream as output. Enable the thinking toggle and tune the ratio (1.0 = thinking roughly equals visible output).
• Long contexts usually cost more, not just in tokens but in latency. Use the context-fit meter — aim to stay under 70% of the window for predictable performance.
• Copy the share link to lock the scenario into a URL — great for cost-review docs, Linear tickets and pull requests.
• Open with a preset via the query string, e.g. `/dev-tools/ai-token-counter/?preset=cheapest` or `?preset=flagship` (also `reasoning`, `long-context`, `coding`). A share URL that includes selected models in the hash overrides the preset.