[tokcalc demo]
Switch model → multi-GPU → long-context capacity planner → Build vs Buy → Reference catalog → Live GPU pricing
tokcalc turns your LLM traffic, context length, latency SLOs, cache behavior,
and model choice into a defensible serving topology and cost plan —
with transparent formulas and cited benchmarks.
It's the engineering-grade pre-deployment decision layer for LLM inference.
Not another static "tokens per second" calculator.
Try it now
tokcalc.vercel.app — no signup, no tracking, no paywall.
tokcalc is an open-source LLM serving capacity planner with 44 models,
31 accelerators, and 16 quantization formats. It estimates model fit, KV/state
memory, throughput, latency, topology, and cost through a shared catalog and
calculation engine used by the website, the MCP server, and the CLI.
Pick a model, a GPU, and a workload. Get an instant capacity plan:
- Generation speed (tok/s)
- Estimated first-token latency + inter-token latency (ITL)
- VRAM budget with KV-cache sizing
- Multi-GPU topology recommendation (Single GPU → TP×2/4/8 → Context Parallel)
- Monthly cost + break-even vs API pricing
- Live GPU pricing across Azure + AWS + GCP + Vast.ai marketplace (fetched in parallel)
- Shareable URL — your config encoded in the URL hash, send to colleagues
- Copy as Markdown — paste the full result into GitHub issues / Slack / docs
GitHub Action — plan on every PR
If your repo declares a serving plan, tokcalc can check it. Drop a
.tokcalc.json at the repo root:
Then add this to .github/workflows/plan.yml:
Every PR that touches the plan gets a comment with the VRAM budget, KV-cache
sizing, throughput, TTFT/ITL, and cost — so a reviewer can check the model/GPU
choice without leaving the diff.
Pin to a release tag (as above) rather than @main, so the action can't
change under you between runs.
The action runs tokcalc's own src/lib/token-calc.ts from this repository, so
the numbers it reports are the same ones the site and the MCP tools produce.
There is no second implementation to drift.
Any field can be overridden per-run, or you can skip the config file entirely:
Run the same thing locally without the action:
Exit code is 0 when the plan fits, 2 when it does not, 1 on a usage error
— so it works as a CI gate.
Why tokcalc?
The market has dozens of "tokens per second" calculators and self-host-vs-API
break-even tools (induwara.lk, gigagpu, kickllm, cloudparity, curlscape,
profitable.ai). None of them are unified capacity planners — and none show
their math.
What tokcalc answers that competitors can't
"Can I serve Qwen 2.5 72B at 128K context on 2× H100 with 20 concurrent users?"
"How many H200s do I need for 1,000 req/min with P95 TTFT < 2s?"
"Does FP8 or AWQ save more money once quality, KV cache, and engine support are included?"
"At what daily volume does an H100 beat GPT-4o pricing?"
"What's the cheapest H100 right now across Azure / AWS / GCP / Vast.ai?"
"What happens to cost and latency if an agent makes 8 model calls, has 3 tool calls, and its context grows by 5K tokens each turn?"
"Would prefix caching, continuous batching, or PD disaggregation save more for this workload?"
Features
Calculator tab
Build vs Buy tab
Independent calculator (separate state) that compares:
- Self-host: model + GPU + quant + utilization + batch → $/M tokens + monthly cost
- API: 13 providers (OpenAI / Anthropic / Gemini / Groq / DeepSeek / Mistral / Together)
- Verdict: Self-host cheaper / API cheaper / Not enough volume — with break-even reqs/day
Reference tab
A tab on the homepage (tokcalc.vercel.app → Reference), not a separate route.
6 sub-tables — fully transparent, every record source-linked where available:
- Models (44 entries: Llama 4 Scout/Maverick, Qwen 3 family, DeepSeek V3/R1, Pixtral, BGE-M3, GPT-OSS 20B/120B, Kimi K3, GLM-5.3, ...)
- Accelerators (31 entries: H100/H200/B200/B300, AMD MI300X/MI325X, Intel Gaudi 3, TPU v5p/Trillium, Groq LPU, Cerebras WSE-3, Apple M2/M3/M4 Ultra, ...)
- Quantization (16 formats: FP16/BF16, GGUF Q2_K→Q8_0, GPTQ, AWQ, EXL2, FP8, NVFP4)
- API pricing (13 models with input/cached/output + retired/current status)
- Cloud GPU pricing (all accelerators with $/hr > 0 + typical providers — static curated)
- 🔴 Live pricing (real-time fetch from Azure + AWS + GCP + Vast.ai in parallel — LIVE badges + timestamps)
The long-context capacity planner (the differentiator)
This is the formula the Perplexity research brief called "the most important
tokcalc should visibly expose":
$$
\text{KV bytes/request} = 2 \cdot L \cdot T \cdot H_{\text{kv}} \cdot D_h \cdot B
$$
For dense attention, prefill cost grows superlinearly with context length:
$$
\text{prefill FLOPs} = \underbrace{2 \cdot N \cdot T}{\text{linear}} + \underbrace{T^2 \cdot H{\text{kv}} \cdot D_h \cdot L}_{\text{attention}}
$$
tokcalc shows you:
- Max concurrent users at 4K / 8K / 16K / 32K / 64K / 128K / 256K / 512K / 1M context
- KV memory per request at each context length
- Prefill time (with superlinear attention correction beyond 32K)
- Required topology (Single GPU → TP×2/4/8 → TP×8 + Context Parallel)
- RingAttention citation when CP is needed
Live GPU pricing (NEW)
The 🔴 Live pricing sub-tab in Reference fetches real-time GPU prices from 4 providers in parallel via Promise.allSettled:
Each provider shows a LIVE badge with timestamp + cache state. The comparison table shows the cheapest price per GPU across all 4 providers (highlighted in emerald) + per-provider breakdown.
If any provider fails (e.g., GCP_API_KEY not set), the others still work — graceful degradation per provider.
The math, transparently
Every number above comes from a formula you can inspect. No black boxes.
Decode tokens/sec (memory-bandwidth bound)
$$
\text{decode tok/sec} \approx \frac{\text{HBM BW} \cdot \eta_{\text{mem}} \cdot \text{quant_eff}}{\text{model size}}
$$
Where:
HBM BW = GPU memory bandwidth (e.g., 3350 GB/s for H100 SXM)η_mem = 0.65 = typical real-world memory utilization (35% overhead)quant_eff = dequantization efficiency multiplier (1.0 for FP16, 1.5 for FP8 on H100, 0.85 for INT4)model size = active_params × bytes_per_param (uses ACTIVE params for MoE, not total)
Refs: PagedAttention paper (arxiv.org/abs/2309.06180)
Prefill tokens/sec (compute bound)
$$
\text{prefill tok/sec} \approx \frac{\text{GPU FLOPS} \cdot \eta_{\text{compute}}}{2 \cdot \text{active params}}
$$
Where:
GPU FLOPS = dense FP16/BF16 TFLOPS (sparse values not used)η_compute = 0.50 = typical compute utilization- Factor of 2 = one multiply + one add per parameter per token
For long context (>32K), the superlinear attention correction above applies.
Continuous batching multiplier (workload-specific)
$$
\text{aggregate tok/sec} = \text{decode tok/sec} \cdot \text{batch size} \cdot \text{continuous batching multiplier}
$$
Critical caveat: There is no universal continuous batching multiplier. vLLM
reported 14–24× vs HF Transformers (extreme), 2.2–2.5× vs TGI. SOSP paper
finds 2–4× typical vs FasterTransformer/Orca. tokcalc defaults to a
conservative 1.5× and lets you tune.
Refs:
Prompt caching economics (Anthropic / OpenAI)
For shared prefix of length $T_p$, suffix of length $T_u$, output $O$,
cache hit rate $h$:
$$
\text{API input cost} = N \cdot \left[ (1-h) \cdot T_p \cdot P_{\text{write}} + h \cdot T_p \cdot P_{\text{read}} + T_u \cdot P_{\text{input}} \right]
$$
Anthropic multipliers (verified 2025-2026):
- 5-minute cache write: 1.25× base input
- 1-hour cache write: 2.0× base input
- Cache read: 0.1× base input (90% savings)
OpenAI: cached input discounted 50%, no separate write fee.
Refs:
Multi-GPU topology recommendation
$$
\text{total needed} = \text{model weights} + (\text{KV per request} \cdot \text{batch size})
$$
Walk the smallest topology that fits:
- Single GPU: total ≤ VRAM × 1
- Tensor Parallel ×2/4/8: total ≤ VRAM × N (weights + KV split evenly)
- TP×8 + Context Parallel: total > VRAM × 8 — use RingAttention to shard KV across nodes
Refs: RingAttention paper
Self-host vs API break-even
$$
\text{self-host $/M tokens} = \frac{\text{GPU $/hr}}{3600 \cdot \text{effective tok/s} \cdot \text{utilization}} \cdot 10^6
$$
$$
\text{break-even req/day} = \frac{\text{monthly self-host cost}}{30 \cdot \text{API cost per request}}
$$
The decisive term is effective utilization — not peak throughput. A GPU
running at 10% utilization pays 10× more per token than the theoretical minimum.
Comparison with adjacent tools
Roadmap
Shipped
- ✅ 44 models, 31 accelerators, 16 quantization formats
- ✅ Continuous batching, reasoning tokens, prompt caching
- ✅ Estimated first-token latency / ITL split, long-context superlinear attention
- ✅ Long-context capacity planner + topology recommendation
- ✅ Build-vs-Buy calculator (13 API providers with retired/current status)
- ✅ Reference catalog (6 sub-tables including live pricing)
- ✅ Live GPU pricing (Azure + AWS + GCP + Vast.ai parallel fetch with LIVE badges)
- ✅ Observed benchmark calibration (paste vLLM/SGLang/TRT-LLM JSON → verdict)
- ✅ MCP server v0.3.4 — 10 read-only tools + 1 calibration writer, stdio + Streamable HTTP transports
- ✅ Single-core MCP architecture — one
src/lib/mcp/core.ts, shimmed into both the
hosted endpoint and the npm package, so the two surfaces cannot drift
- ✅ Published to the MCP Registry and kept
current by an OIDC workflow on every
v* tag
- ✅ Correct MoE accounting — resident (full) weights and active (per-token)
parameters reported separately, instead of conflating them
- ✅ Strict tool schemas — Zod 4 native JSON Schema; unknown parameters are rejected
with the valid names suggested back
- ✅ Public hosted MCP endpoint at
tokcalc.vercel.app/api/mcp with bearer auth + Upstash Redis rate limiting
- ✅ Self-serve API key generation at
tokcalc.vercel.app/mcp (email → instant key)
- ✅ 4 SEO landing pages:
/compare/h100-vs-h200, /compare/gguf-q4-k-m-vs-q5-k-m, /self-host-vs-openai-api, /mcp (install docs)
- ✅ Share URL + localStorage persistence
- ✅ GitHub Action — drops a capacity + cost plan straight onto your PR
- ✅ Copy result as Markdown (for GitHub issues / Slack / docs)
- ✅ Dark mode toggle
- ✅ Plain-English glossary (28 terms with hover tooltips)
- ✅ OG image + Twitter card + social metadata
- ✅ Plausible Analytics (privacy-friendly) + Sentry error monitoring
Next 30 days
- ⏳ Versioned price + benchmark provenance — source URL + retrieval date on every rate and benchmark
- ⏳ P/D disaggregation planner (separate prefill + decode pools)
- ⏳ Engine-aware presets (vLLM / SGLang / TensorRT-LLM / llama.cpp) — UI integration
- ⏳ Multi-turn / agentic economics — growing context and tool-call turns in the cost model
Next 90 days
- ⏳ Multi-LoRA capacity planner (Punica / S-LoRA economics)
- ⏳ VLM image-token accounting (per-model patch/tile tokenization)
Deliberately not doing
i18n (Chinese / Japanese / Korean) — permanent translation maintenance across
44 models × 31 accelerators × 11 tools, and traffic is currently overwhelmingly CI and agent
installs rather than human readers. Revisit only if CJK traffic shows up in analytics.OAuth 2.1 / PKCE — auth infrastructure for a product with one user. Self-serve key
issuance plus Upstash rate limiting already covers the real need. Revisit on demand.Training/fine-tuning FLOPs, energy/carbon, embedding mode — each of these is a
separate product in a separate domain. "Capacity planner" is a sharp position; widening
it into a general AI-infra calculator suite costs more than it earns.
Open core model
tokcalc is open core — the calculator and catalog are open source; the
cloud/data/team features are paid.
Why this structure
- Trust: Open-source formulas build credibility vs opaque competitors
- Community: Contributors can submit models, accelerators, quants, benchmarks
- Defensibility: Trademark + cloud features + URL-share viral loop protect against forks
- Revenue: Cloud tier funds ongoing development + pricing/benchmark data maintenance
Contributing
We welcome contributions! See CONTRIBUTING.md for:
- How to add a model (with HuggingFace config.json as source)
- How to add an accelerator (with critical guardrails for B200/B300 null FP16 fields)
- How to add a quantization format (with measured file sizes as source)
- How to submit a benchmark (3-tier confidence model)
- How to improve a formula (cite the source, no magic numbers)
Most-needed contributions
- 🟢 New models (Qwen 3, Mistral Large 4, Llama 4 variants as they release)
- 🟢 New accelerators (B300, AMD MI400, Apple M5 Ultra when shipping)
- 🟢 New quantization formats (mxFP8, MXFP4, BitNet 2)
- 🟢 Real benchmark data (run vLLM benchmarks and submit with provenance)
- 🟢 Translations (especially Chinese, Japanese, Korean)
MCP server — use tokcalc from AI agents
tokcalc ships an MCP (Model Context Protocol) server that lets AI agents (Cursor, Claude Desktop, Cline) call tokcalc during design reviews.
npm package: @tokcalc/mcp-server (latest: v0.3.4)
Public endpoint: https://tokcalc.vercel.app/api/mcp (Streamable HTTP, bearer auth, rate-limited)
Install docs: https://tokcalc.vercel.app/mcp
11 tools — 10 read-only + 1 calibration writer
The ten planning tools have no side effects, no cloud credentials, no
deployments. record_measured is the one writer: it stores a measurement so
later estimates for that model+GPU+quantization can self-calibrate. Every tool
declares this in the MCP wire annotations (readOnlyHint / destructiveHint /
idempotentHint / openWorldHint), so a client can tell a read from a write
before calling it.
Tool parameters are strict — unknown keys are rejected rather than ignored, and
the error names the valid ones. The parameters are model and gpu (not
modelId / gpuType).
Option A: Public hosted endpoint (recommended — no install)
Point Cursor or Claude Desktop at the public endpoint. Get an instant API key at /mcp:
Features:
- ✅ Bearer API key auth (constant-time comparison via
crypto.timingSafeEqual)
- ✅ KV-backed rate limiting via Upstash Redis (30 req/min per IP, 120 req/min per key)
- ✅ Self-serve key generation — enter email at
/mcp, get instant key (5 per IP per day limit)
- ✅ Keys expire after 90 days of inactivity
- ✅ SHA-256 key hashing (never stored raw)
Option B: Self-hosted stdio (for local Cursor / Claude Desktop)
No API key needed for stdio (local install). Cursor spawns the process via npx.
Option C: Self-hosted HTTP (for air-gapped / custom infra)
Set MCP_API_KEY, UPSTASH_REDIS_REST_URL, UPSTASH_REDIS_REST_TOKEN env vars for auth + rate limiting.
Example agent prompt
"I need to serve Llama 3.3 70B at 32K context for 50 concurrent users. What GPU topology do you recommend, and how much will it cost per month?"
The agent calls list_models → list_gpus → recommend_topology → estimate_capacity → returns a structured plan with VRAM, throughput, latency, cost, and provenance.
Development & testing
GitHub Actions runs all of the above (except the published-tarball mode) on every push — see .github/workflows/ci.yml.
Tech stack
- Framework: Next.js 16 with App Router (Turbopack)
- Language: TypeScript 5
- Styling: Tailwind CSS 4 + shadcn/ui (New York)
- Charts: Recharts
- State: React hooks (useState + useEffect + useMemo)
- Theme: next-themes (dark mode default)
- MCP SDK:
@modelcontextprotocol/sdk v1.30.1 (Streamable HTTP transport, stateless mode)
- Auth:
crypto.timingSafeEqual (constant-time bearer key validation)
- Rate limiting:
@upstash/ratelimit + @upstash/redis (sliding window, 30/min per IP, 120/min per key)
- Analytics: Plausible (privacy-friendly, no cookies)
- Error monitoring: Sentry
- Database: None for core; Upstash Redis for API key storage + rate limiting
Project structure
Acknowledgments
tokcalc builds on the work of:
- vLLM team — PagedAttention, continuous batching (arxiv.org/abs/2309.06180)
- llama.cpp / ggml-org — GGUF format and quantization variants
- MLPerf / MLCommons — standardized inference benchmark methodology
- Stanford HELM — efficiency-aware model evaluation framework
- Anthropic / OpenAI / Google — published prompt-caching pricing rules
- Model Context Protocol team — MCP Streamable HTTP transport spec (2025-03-26)
- NVIDIA / AMD / Intel / Google / Groq / Cerebras — published hardware specs
- Upstash — Redis + rate limiting infrastructure
Every formula has a citation. Every model/GPU/quant entry has a source URL
where available. If you spot an unsourced claim, please open an issue.
Star history
[Star History Chart]
