1M-context flagship prior to GPT-5.6; still the widely deployed standard.
Claude Opus 5
Anthropic
$5.00
$25.00
1,000k
Frontier
Agentic tool use
Medium
Top production Opus tier; best for agentic reasoning and long-horizon tool use.
Claude Opus 4.8
Anthropic
$5.00
$25.00
1,000k
Frontier
Stable Opus 4.x
Medium
Last stable Opus 4.x before Opus 5; 1M context at the same $5/M input rate.
Claude Opus 4.7
Anthropic
$5.00
$25.00
1,000k
Frontier
Pinned production
Medium
Pinned production Opus for teams that value output stability over the latest weights.
Claude Fable 5
Anthropic
$10.00
$50.00
1,000k
Frontier
Long-context frontier
Medium
Anthropic’s premium tier above Opus 5; 1M context at Opus-5-fast pricing.
Claude Opus 4.1
Anthropic
$15.00
$75.00
200k
Frontier
Enterprise legacy
Medium
Long-serving enterprise Opus; 200k context at premium $15/M pricing.
GPT-5.5 Pro
OpenAI
$30.00
$180.00
1,050k
Reasoning
Deep analysis
Slow
Highest-effort GPT-5.5 tier; the pre-5.6 flagship for reasoning workloads.
GPT-5.4 Pro
OpenAI
$30.00
$180.00
1,050k
Reasoning
Hard math / code
Slow
High-effort variant of GPT-5.4; long-context reasoning without 5.5 pricing.
Prices reflect each provider's public list price, auto-refreshed daily via the OpenRouter API. Open-weights rows are marked as estimates (varies by host: Together, Groq, Fireworks, etc.). Last updated — methodology & sources for per-provider details.
How to reduce your LLM API costs
Six levers that consistently bring monthly LLM bills down by 30–70% in production.
Pick the right tier for the job
Use a cheap fast model (GPT-4o mini, Gemini Flash, Claude Haiku) for routing, classification, and most production tasks. Reserve frontier models for the hard 5–10% of calls.
Cache prompt prefixes
Anthropic, OpenAI, and Google all expose prompt caching. A long system prompt or RAG context replayed across requests can be served at ~10% of normal input price.
Cap the output, not the input
Output tokens are 3–5× more expensive than input. Set max_tokens, use stop sequences, and ask the model for structured JSON instead of prose whenever possible.
Route by complexity
Send simple queries to a fast model and only escalate to a frontier model when a confidence check or eval fails. A two-tier router cuts cost 40–70% in production.
Compress context before sending
Summarise long histories, dedupe retrieved chunks, and strip boilerplate from system prompts. Most teams over-fill the context window by 2–3×.
Batch & stream where you can
OpenAI and Anthropic Batch APIs run within 24h at ~50% off. Streaming doesn’t change the bill but lets you cancel mid-flight when the user navigates away.
Understanding LLM API pricing
LLM API pricing is almost always quoted per million tokens, separately for the prompt you send in (“input”) and the text the model writes back (“output”). A token is roughly 3–4 characters of English, so 1,000 tokens ≈ 750 words. The total bill for any given call is simply: input_tokens × input_rate + output_tokens × output_rate.
Three modifiers can change that base number meaningfully. Prompt caching lets you replay long system prompts or retrieved context at ≈10% of the normal input rate, which is enormous for RAG. Batch APIs (OpenAI, Anthropic) trade synchronous latency for a 50% discount on jobs that can wait up to 24 hours. And fine-tuning generally costs 1.5–3× the base inference rate, so the math only works if you’re saving on prompt length or quality at scale.
Most teams underestimate output cost. Output tokens are typically 3–5× more expensive than input tokens, and chatty models pad responses unless you cap them. The single highest-leverage change is usually setting max_tokens aggressively and asking for structured JSON instead of prose.
Methodology & sources
How this calculator stays fresh
Prices auto-refresh daily via the OpenRouter API, with each provider’s official pricing page as the underlying source of truth. Pricing last updated: .
Live pricing · 57/57 rows from OpenRouter
Calculation methodology
The estimated monthly bill you see on the right of the calculator is a deterministic function of four inputs — the selected model, input tokens per request, output tokens per request, and requests per day. The exact formula is:
A 30-day month, matching how every major provider reports usage on invoices.
List prices published by each provider — no volume discounts, committed-use credits, or negotiated enterprise rates.
No prompt caching, batch discounts or fine-tuning surcharges (those levers are covered separately in the “How to reduce cost” section).
Token counts are treated as-billed: providers charge for reasoning traces and tool-call outputs where applicable, so numbers can undershoot in agentic workloads.
Open-weights rows (Llama, Mixtral, Qwen, self-hosted) are marked as estimates because effective cost depends on the host you pick and your GPU utilisation.
Caveats
Prices change frequently. Always confirm live rates on the official provider page linked below before contracting.
Region-specific rates (Vertex AI EU, Bedrock cross-region) can differ by 10–30% from the base US price shown here.
Fine-tuned model inference is billed at a premium (typically 1.5–3× base). This calculator prices the base model only.
Update history
A running log of every refresh — the topmost entry is regenerated automatically on each ISR revalidation; older entries capture the broader release cadence.
Automated daily refresh
Live prices pulled from the OpenRouter API and merged onto the canonical model list.
Applied to the calculator, comparison table and JSON-LD dateModified in the same build.
Q3 2026 refresh
Re-verified every row against each provider’s public pricing page.
Added Mistral Small 3 (Feb 2026 release) and refreshed DeepSeek V3 rates after the March discount.
Widened Google Gemini 2.0 Flash context note to reflect the 1M-token GA.
Q2 2026 refresh
Re-priced OpenAI o1 and o1-mini after the June rate cut.
Marked Claude 3 Opus as legacy pending Anthropic’s deprecation notice.
DeepSeek V3 launch update
Added DeepSeek V3 and DeepSeek Coder V2 at launch prices.
Refreshed Llama 3.1 family rates from Together AI / Groq.
January 2026 refresh
Backfilled release years and category tags on every row.
Introduced the `isEstimate` flag for self-hosted and open-weights rows.
Initial publish
First public release of the calculator with 22 models across 8 providers.
Established the "prices per 1M tokens" convention shown throughout the page.
Data sources & last verified dates
Live prices come from the OpenRouter API on a daily ISR schedule. The provider rows below are the underlying source of truth we (and OpenRouter) mirror — where a provider does not publish first-party API rates (Meta’s open-weights family, self-hosted rows) we cite the third-party host we sourced the number from and flag those rows as estimates in the comparison table.
Provider
Official pricing page
Last verified
OpenRouter API (automated feed)Live
Primary source. All prices in the table are refreshed daily from this endpoint via ISR; providers below are the underlying source of truth OpenRouter mirrors.
Pricing pages are published and controlled by each provider — we don’t mirror them. If you spot a stale rate, open the relevant source above, then let us know so we can push a mid-quarter update.
Frequently asked questions
Ready to scale with confident pricing?
Plug your real traffic into the calculator above, then dig into our directory of 1,000+ AI tools and frameworks to ship your next product.