Cafiyn Pulse
22 providers, verified monthly

Free AI cost calculator across Claude, GPT-5, Gemini, DeepSeek, and 36 more.

Every AI provider has a pricing page. None of them tell you your bill. Enter your users, calls, tokens, audio minutes, and image count. Get a live projection across 79 models, a recommendation for your scale, and a shareable card.

The problem

Provider pricing pages lie by omission.

OpenAI shows you $5 per million tokens. Anthropic shows you $3 per million. Google shows you $1.25. Every page implies a small number, and none of them tell you what your actual monthly bill will be, because none of them know how many users you will have, how many turns per session, or how big your context windows are.

The gap between “$3 per million tokens” and “$4,200 in November” is where most AI startups get surprised. It shows up as a Slack ping from the finance person, a panic switch to a cheaper model, prompt caching added in a hurry, and sometimes a hard price hike passed to end users. It is preventable with two minutes of math up front.

The Cost Comparator turns provider pricing pages into a live model of your bill. Change the slider, watch the bill change. Compare 8 providers at once. See which model wins at 1,000 users and which one wins at 100,000.

What it covers

Every provider that matters in 2026.

79 models across five categories, updated weekly by an automated verification job. If a provider prices you use is not here, add a custom model in one click.

Text (LLMs)
Claude Haiku 4.5, Sonnet 5, Opus 4.8. GPT-5, GPT-5 mini, GPT-4.1, GPT-4.1 mini, o3-mini. Gemini 2.5 Pro, Flash, Flash-Lite. DeepSeek V3 and Reasoner. Grok 4 and mini. Mistral Large 2 and Small 3. Llama 3.3 70B and 3.1 405B. Cohere Command R+. Perplexity Sonar Pro.
Text-to-speech
ElevenLabs Multilingual v2 and Turbo v2, OpenAI TTS and TTS HD, Google Cloud TTS Neural2 and Studio, Deepgram Aura, Azure Neural TTS. Priced per minute.
Speech-to-text
Whisper, Deepgram Nova 2 and Nova 3, AssemblyAI Universal-2, Google Cloud STT, Azure STT. Priced per minute, both real-time and batch.
Image generation
DALL-E 3 and DALL-E 3 HD, Stable Diffusion 3.5 and SDXL, Flux Pro and Flux Schnell. Priced per image, with SDXL and Flux Schnell as ultra-cheap draft options.
Video generation
Runway Gen-3, Pika, Luma, Kling, and Sora tiers. Priced per second of output. Great for storyboards and marketing shorts.
Custom models
Bring your own pricing for private LLMs, fine-tunes, or providers not yet in the roster. Same live math applied.
What you will find out

Answers you will get in two minutes.

These are the specific questions the calculator resolves. Every one of them has cost a founder time or money to answer manually.

  • What is the cheapest capable model for my task in 2026 (and how much cheaper is it than my current choice)?
  • If I switch from GPT-5 to Claude Sonnet 5, how much does my monthly bill change?
  • What is the true cost of adding a “deep research” premium mode that uses Opus 4.8 or DeepSeek Reasoner?
  • If I add speech input (Whisper vs Deepgram Nova 3), what does that add per user?
  • At what point does the DeepSeek V3 or Gemini Flash-Lite savings justify a fallback route instead of a single-model pipeline?
  • What is my worst-case bill if a viral moment 10x my usage overnight?
  • If I set a $500 monthly ceiling, which providers will I need to cap first?
Head-to-head

Sample matchups the calculator handles.

ModelInput per 1MOutput per 1MBest for
Claude Fable 5.1$10.00$50.00Deepest reasoning, research and analysis
GPT-6 Astra$10.00$50.00OpenAI flagship, escalation tier only
Claude Opus 5$5.00$25.00Hard reasoning and agentic loops, a third of Opus 4.8
GPT-5.6 Sol$4.00$20.00Premium tier below the flagship
Claude Sonnet 5$2.00$10.00Strongest mid-tier for chat and tool use
GPT-5.6 Terra$2.00$12.00Closest replacement for the retiring o3-mini
Gemini 3.1 Pro$2.00$12.00Long context, multimodal, Google ecosystem
Grok 4.6$2.00$6.00Low output price, doubles above 200k tokens
GPT-5$1.25$10.00Reliable workhorse, now priced like a mid-tier
Claude Haiku 4.5$1.00$5.00Cheapest capable Anthropic option
Gemini 3.8 Flash$0.75$3.75Fast and cheap, introductory rate ends 31 Dec 2026
DeepSeek V4 Pro$0.66$1.98Very cheap, halves again off-peak
Mistral Large$0.50$1.50European hosting, cut four-fold from Large 2
GPT-5 mini$0.25$2.00Cheap OpenAI classifier and router
GPT-5.6 Luna$0.20$1.20Cheapest tier in the GPT-5.6 family
Prices verified 2026-09-05, read from the same roster the calculator uses. Every change is dated and sourced in the price change log.
How it works

How to model your bill in five steps.

Step 1
Sign up free (10 sec)
One email. One 6-digit code. No card, no password. Your work is saved to your account.
Step 2
Enter your usage
Sliders for users per day, calls per user, tokens per call. Optional: audio minutes, images per month, video seconds.
Step 3
Read the recommendation
The engine picks a primary + premium fallback + cheap floor based on your scale. Explanation attached.
Step 4
Compare live
Check any providers to compare; the projection bars re-rank in real time.
Step 5
Turn on alerts (optional)
Email alert if projected monthly cost crosses your threshold. Default $500.
Save more

Five levers that cut most AI bills 30 to 60 percent.

The calculator shows you the price. These levers change the shape of the bill.

  1. Cache prompts and responses. Anthropic prompt caching, OpenAI cached input, and simple response memoization typically shave 30 to 50 percent off a chatty workload.
  2. Route by difficulty. A cheap classifier (GPT-4.1 mini, Haiku 4.5, DeepSeek V3) triages, then only hard cases go to Opus 4.8 / Sonnet 5 / GPT-5.
  3. Set per-user daily caps. Prevents a small number of heavy users from consuming your entire budget.
  4. Cut token counts with prompt engineering. Removing rehashed system prompts, compressing tool schemas, streaming instead of buffering, and tightening JSON output all reduce spend.
  5. Batch long-running work. Overnight image generation on Flux Schnell or SDXL is roughly 20x cheaper than daytime interactive generation on Flux Pro.
Related
By workload

Building something specific? Start there instead.

FAQ

Common questions.

Is this AI cost calculator really free?

Yes. Free forever. No credit card at signup, no trial, no premium tier. You can compare all 79 models, use the recommendation engine, and turn on budget alerts without paying anything.

Which AI providers and models are included?

Text: Anthropic (Claude Haiku 4.5, Sonnet 5, Opus 5, Fable 5.1), OpenAI (GPT-6 Astra, GPT-5.6 Sol / Terra / Luna, GPT-5 and mini and nano, o3, o3-mini, o4-mini, GPT-4o), Google (Gemini 3.8 / 3.7 / 3.5 Flash, 3.5 Flash-Lite, 3.1 Pro), DeepSeek (V4 Pro, V4 Flash), xAI (Grok 4.6, Grok 4.3), Mistral (Large), Meta (Llama 3.3 70B via Together), Groq (GPT OSS 120B and 20B), Cohere Command R+, Perplexity (Sonar, Sonar Pro, Sonar Reasoning Pro). Text-to-speech: ElevenLabs Multilingual v2 and Turbo v2, OpenAI TTS and TTS HD, Google Cloud TTS Neural2 and Studio, Deepgram Aura, Azure Neural TTS. Speech-to-text: Whisper, Deepgram Nova 2 and Nova 3, AssemblyAI Universal-2, Google Cloud STT, Azure STT. Image: DALL-E 3 and 3 HD, Stable Diffusion 3.5 and SDXL, Flux Pro and Schnell. Video: Runway, Pika, Luma, Kling, and Sora tiers.

How are prices kept up to date?

Two layers. A GitHub Action runs every Monday and checks every price we publish against the provider's own pricing page, opening an issue when one no longer matches; it reports rather than edits, because a scraper that guesses wrong would publish a wrong price with the authority of an automated check. On top of that, a person re-verifies the roster against every provider page monthly. Every change either pass finds is published with its date and source at /cost-comparator/pricing/changes, including our own corrections.

What is the difference between Claude Sonnet 5 and GPT-5 for cost?

Claude Sonnet 5 is $2 per million input tokens and $10 per million output. GPT-5 is $1.25 and $10. For a 1,000-token in, 500-token out prompt, Sonnet 5 costs about $0.007 and GPT-5 about $0.00625, so GPT-5 is roughly 11 percent cheaper, not more expensive. An earlier version of this page had both figures wrong and said the opposite; the correction is recorded in the price change log.

What is the cheapest capable LLM in 2026?

GPT-5 nano is $0.05 / $0.40 per million tokens, and Gemini 2.5 Flash Lite is $0.10 / $0.40, though Google now lists the 2.5 line as legacy. Among current models, GPT-5.6 Luna is $0.20 / $1.20 and Mistral Large is $0.50 / $1.50, which is unusually cheap for a model of that tier. DeepSeek V4 Flash is $0.22 / $0.66 off-peak but doubles during peak hours. The right answer depends on your quality bar: the calculator ranks by projected monthly cost given your own usage.

What is the recommendation engine?

The Cost Comparator reads your usage sliders (users per day, calls per user, tokens per call, audio minutes, images per month, seconds of video) and routes based on volume: under 5k calls / month use premium models because cost is a rounding error; 5k to 100k use a mid-tier workhorse plus premium for hard cases plus a cheap fallback; over 100k cost-optimize hard, cache aggressively, use routing.

Can I add a custom model that is not in the roster?

Yes. Click "Add custom model" and enter your own pricing. Useful for your own fine-tunes, private LLMs, or providers not yet supported.

Can I export or share my projection?

Yes. Every projection generates a shareable card (1200x630 PNG) with your min, max, and cheapest option, and a link back to a live view. Great for sharing with a co-founder or dropping into a board update.

How is this different from a spreadsheet?

A spreadsheet is a snapshot. This is live. Prices are drift-checked weekly and re-verified monthly. You add a new provider and every projection updates. You want to compare 8 models at once, you check 8 boxes. You want to see what a 3x traffic spike does to your bill, you drag a slider.

Open the tool.

You can compare all 79 models, use the recommendation engine, and turn on budget alerts without paying anything.

Open the free AI cost calculator