The complete guide

What an AI agent really costs

The model bill is the part everyone sees. It is rarely the biggest part. A real AI agent carries about a dozen costs - safety, quality checks, human review, compliance, monitoring, integrations, keeping data and models current - and the ones people skip are exactly the ones that quietly sink the project. This guide walks through every cost the cost builder includes, what drives each one, and how they add up into a clear go / no-go answer.

Open the Builder →

Why it costs more than the model bill

A demo runs on the model alone. A real production agent runs on everything around it. The research is blunt: a large share of AI agent projects get cancelled before they pay off, and when they stall the cause is almost never the model itself - it is the safety, quality-checking, and integration work that nobody budgeted for. Pricing only the model bill is how a project looks cheap on a slide and falls apart in production.

This tool takes the opposite approach. You enter simple numbers for each cost - how many requests, how often a human steps in, how big your data is - and each cost is worked out by its own dedicated calculator behind the scenes. Nothing is a rough guess. Sections you have not filled in stay on example numbers and are clearly marked, so the total is always honest about how much is really yours.

Getting it built

What it takes to stand the system up: integrations and connectors to your existing tools.

Integrations and connectors

live engine

Prices the most underestimated build-phase line: connecting the agent to real business systems such as CRM, ERP, ticketing, and identity. Each connector is authentication, schema mapping, rate-limit and error handling, plus ongoing upkeep, not just plugging in an API. Shows one-time build, annual maintenance, amortized annual cost for TCO, and cost per integration.

Why integration is the blind spot
Agents rarely operate in isolation; they must reach CRM, ERP, ticketing, identity, and analytics. A simple CRM hookup balloons into weeks once you add auth layers, schema mapping, rate-limit handling, and error recovery. Integration plus QA commonly run 40-60 percent of total build cost.
The math
Build days per integration = base days x complexity factor. One-time build = systems x build days x loaded day rate, plus discovery. Annual maintenance = build labor x maintenance percent. Amortized annual = one-time / amortization years + maintenance.
Drivers in this layer (7)
Integrations
Count distinct systems: each CRM, ERP, ticketing, identity, or data source.
Complexity
Legacy and undocumented systems can double timelines; pick complex for those.
Build-days each
Standard SaaS connectors are ~1-2 weeks; complex ones more.
Loaded $/eng-day advanced
Salary plus overhead divided by working days, or your contractor day rate.
Maintenance %/yr advanced
Industry norm is 15-25 percent; higher for brittle or fast-changing APIs.
Discovery $ (one-time) advanced
Typical discovery runs $1.5K-$15K; prevents far more expensive missteps.
Amortize years advanced
Common horizons are 2-3 years.

Running it, day to day

What you pay to keep it running: model usage, the search that grounds it, storage and network.

AI usage (the model bill)

live engine

See exactly what an LLM workload will cost across 70+ models. Pick a model, enter your tokens per request and daily volume, get per-request / daily / monthly / annual cost. Caching and Batch API savings calculated automatically.

Enter every field
Input tokens = everything you send (system prompt + user message + retrieved docs + examples). Output tokens = what the model generates. Rough rule: 1 token ≈ 0.75 English words. If you don't know your exact counts, use Token Estimator to measure real samples first.
Choose a model
Models listed with input / output prices per 1M tokens. Frontier work → Claude Opus 4.8 (newest, May 2026), Gemini 3.1 Pro, or GPT-5.5. Production traffic → Claude Sonnet 4.6 or Gemini 3 Flash. High-volume cheap → Claude Haiku 4.5, GPT-5-nano, or DeepSeek V3.2. Cheaper isn't always better: slower/weaker models may produce longer outputs, eroding the price advantage.
Drivers in this layer (5)
Requests / day
For production apps use real telemetry. For projections: peak users × requests per user per day, plus 30% buffer for retries.
Input tokens / req
Do not guess from word count. Run actual prompts through token-estimator. Typical shapes: single chat turn 300-1500, RAG with retrievals 2K-8K, long-doc analysis 8K-100K.
Output tokens / req
Constrain output in your prompt: "respond in ≤ 200 tokens" beats hoping the model is brief. Typical: classification 10-50, chat reply 150-600, code generation 500-2000, long-form 1000-4000.
Model tier advanced
Enter your own figure; the default is a sensible placeholder.
Working days / mo advanced
Enter your own figure; the default is a sensible placeholder.

Search and retrieval

live engine

Compare 9 embedding models on indexing + query-time cost for your RAG corpus.

Enter corpus stats
Docs = total documents in your RAG corpus. Tokens/doc = average document length in tokens (1500 typical). Queries/month = retrieval calls (user queries that hit the vector DB). Query tokens = avg user query length. These drive indexing (one-time) vs query-side (recurring) cost.
Pick embedding + re-indexing cadence
Embedding model choice matters for quality AND cost. Voyage and Cohere are retrieval-tuned (better recall). OpenAI 3-small is cheap and general. BGE/Nomic are open-source. Re-indexing cadence depends on how often your corpus drifts — stable knowledge bases: 1-2x/year. Fast-changing content: quarterly or more. Dimension truncation (OpenAI 3-large supports it) can cut storage ~50% with minor quality loss.
Drivers in this layer (6)
Docs indexed
Enter your own figure; the default is a sensible placeholder.
Queries / day
Enter your own figure; the default is a sensible placeholder.
Tokens / doc advanced
Enter your own figure; the default is a sensible placeholder.
Embedding tier advanced
Enter your own figure; the default is a sensible placeholder.
Full rebuilds / yr advanced
Enter your own figure; the default is a sensible placeholder.
Vector DB storage $/mo advanced
Enter your own figure; the default is a sensible placeholder.

Storage

manual entry

What it costs to keep your data: model files, training data, the vector database, logs and backups. A simple sum of how much you store (in TB) times the monthly price per TB.

Drivers in this layer (2)
Data stored (TB)
Enter your own figure; the default is a sensible placeholder.
Storage price ($ per TB per month)
Enter your own figure; the default is a sensible placeholder.

Network (data out)

manual entry

What it costs to move data out: sending results to users, copying between regions, and CDN traffic. Charged per TB of data sent out, which is easy to overlook until the bill arrives.

Drivers in this layer (2)
Data sent out (TB per month)
Enter your own figure; the default is a sensible placeholder.
Data-out price ($ per TB)
Enter your own figure; the default is a sensible placeholder.

Keeping it healthy

What it costs to keep it accurate over time: quality checks, fresh data, model upkeep and monitoring.

Quality checks

live engine

Prices continuous quality evaluation for an agent: a golden set of representative cases scored on every release by an LLM-as-judge. Shows the monthly eval cost, the cost per release run, and the agent-versus-judge split. This is the regression net that catches the silent works-Monday-fails-Wednesday drop when a provider updates the model.

Why eval is non-optional
Agent behavior is non-deterministic: the same prompt, tools, and model can drift across temperature sampling and silent provider weight updates. A customer-support router can go from 93 percent to 71 percent accuracy overnight with no test catching it. A golden set scored every release is the safety net.
The math
Per item cost = agent generation (input and output tokens at the agent rate) plus judge scoring (item, response, and rubric at the judge input rate, plus a small verdict). Per run = golden set size x per item. Monthly = per run x runs per month.
Drivers in this layer (6)
Golden-set items
Start around 100-300 covering your real failure modes; grow as you find gaps.
Eval runs / mo
Per PR plus per release plus nightly drift checks; 20-60 a month is common for active teams.
Input tokens / item advanced
Match your production prompt size.
Output tokens / item advanced
Use a typical completion length.
Judge rubric tokens advanced
A concise rubric is 200-600 tokens; longer rubrics raise judge cost.
Judge verdict tokens advanced
Keep it small; binary pass/fail with a one-line reason is often 30-80 tokens.

Monitoring

live engine

Prices the production tracing and monitoring layer for an agent (Langfuse, LangSmith, Datadog, Braintrust, Logfire, or self-host). The unit is the span: each LLM call, tool call, retrieval, and rerank. It computes per-span, per-trace, and self-host billing side by side, because the billing model, not the vendor, is what blows up the bill on agentic workloads.

Why the span is the unit
Observability for agents tracks a span per LLM call, tool invocation, retrieval, and rerank; together they form a trace. A simple call is 3-5 spans; a complex agent is 20-50. Fast HTTP 200s are not a proxy for correct output, so teams instrument to catch quality and cost regressions.
The math
Traced requests = requests x sampling. Spans = traced requests x spans per request. Per-span = spans x rate x retention. Per-trace = traces x rate x retention. Self-host = flat infra. The calculator shows all three.
Drivers in this layer (8)
Requests / mo
Use your real monthly request volume.
Spans / request
Simple calls are 3-5; complex agents are 20-50.
Billing model
Per-span for agents, per-trace for simple chains, self-host past the managed break-even.
% traced advanced
100 pct for full visibility; sample down to cut cost on high volume.
$/million spans advanced
Logfire-class is ~$2/M; Datadog and Langfuse-class run tens of dollars per million.
$/1k traces advanced
LangSmith-class is ~$2.50/1k at 14-day retention.
Retention multiplier advanced
1 for short (~14-day); ~2 for long (~400-day) retention.
Self-host $/mo advanced
Low at small scale; 500+ vCPUs at high ingestion push it to five figures.

Keeping data fresh

live engine

Prices the recurring data treadmill behind a RAG or agentic system: incremental re-embedding of changed documents each refresh, vector upsert and reindex, pipeline orchestration, the periodic full re-embed forced by an embedding-model or chunking change, and lightweight drift monitoring. This is the data-drift bill that pairs with eval's model-drift bill.

Why freshness is a recurring bill
Source content ages, so without scheduled rechunk, reembed, and reindex, retrieval relevance declines. Teams run incremental updates for the changed delta plus periodic full rebuilds. Embedding drift is the quiet killer: it breaks retrieval while everyone blames the model.
The math
Per run = changed tokens x embedding rate + changed tokens x vector-ops rate + pipeline per run. Monthly incremental = per run x runs. Full re-embed = whole corpus x (embedding + vector-ops), amortized over the re-embed interval. Plus monthly drift monitoring.
Drivers in this layer (7)
Corpus tokens
Estimate from document count x average tokens per document.
Refresh runs / mo
Fast-moving sources refresh daily (30); slower ones weekly (4).
% changed / run
Most KBs change 1-5 pct per cycle; bump it up for volatile sources.
Vector ops $/1M tok advanced
Derive from your vector-DB write and index pricing, or leave the default.
Pipeline $/run advanced
A re-index run commonly runs tens of dollars; scale with corpus and steps.
Full re-embed every (mo) advanced
Often 3-12 months; tie it to your embedding-model upgrade cadence.
Drift monitoring $/mo advanced
Light monitoring is a few hundred a month; raise for large or critical KBs.

Model upkeep

live engine

Prices the recurring cost of responding to model drift: forced migrations when a provider deprecates your model, prompt re-baselining and regression after each swap, periodic re-tuning to counter quality drift, and steady-state prompt-ops. This is the fix bill that pairs with eval-hallucination-cost, which is the detect bill, on the model side.

Why model maintenance is recurring
Providers retire models on a roughly 12-18 month cadence, often with weeks of warning. Even a same-family upgrade shifts output: studies show large accuracy swings from formatting alone, so prompts must be re-baselined and regression-tested. One documented forced migration needed 400+ hours of prompt re-engineering.
The math
Migration labor = migrations per year x eng-days per migration x loaded day rate. Re-tuning = re-tunes per year x cost per re-tune. Prompt-ops = days per month x 12 x day rate. Annual = the sum.
Drivers in this layer (6)
Forced migrations / yr
With 12-18 month lifespans across several models, 1-3 a year is common.
Eng-days / migration
A documented case ran two engineers for two weeks (~20 days); simpler swaps less.
Prompt-ops days / mo
Active production agents often need 1-4 days a month even without a migration.
Loaded $/eng-day advanced
Salary plus overhead divided by working days, or your contractor day rate.
Re-tunes / yr advanced
Many teams re-tune 0-2 times a year; 0 if you rely on prompting plus RAG.
Cost / re-tune $ advanced
Retraining commonly runs $2K-$20K; use fine-tuning-cost for token-level detail.

Handling the risks around it

What it costs to keep it safe and dependable: guardrails, human review, compliance, vendor risk and downtime.

Safety guardrails

live engine

Prices the runtime safety layer in front of an agent: safety, prompt-injection, and PII screening passes run on a cheap classifier model for every checked call, plus an optional guardrail-platform fee. Shows the monthly guardrail bill, cost per 1,000 screened calls, and the percent it adds on top of your agent token spend.

Why guardrails cost money
A production agent screens inputs and outputs for safety, prompt injection, and PII before acting. Each screen is an inference call on a classifier model, so the safety layer has its own per-call token cost on top of the agent itself.
The math
Screened calls = calls x screened pct. Cost per pass = (tokens checked x input rate + verdict tokens x output rate) / 1M. Monthly = screened x passes per call x cost per pass, plus any platform fee.
Drivers in this layer (6)
Agent calls / mo
Use your real monthly call volume.
% screened
High-risk surfaces screen 100 pct; low-risk internal flows may screen less.
Screens / call
Match your policy; more passes mean more cost and latency.
Screen input tokens advanced
Roughly the size of the input or output being screened.
Screen output tokens advanced
Keep it small; a label plus reason is often 10-30 tokens.
Guardrail platform $/mo advanced
Zero for self-hosted; use the vendor quote for a managed platform.

Human review

live engine

Prices the human-in-the-loop review layer of an agentic workload: the monthly review labor, the cost per reviewed action, the reviewer headcount it implies, and whether the incidents that review catches pay for the labor. It is the single most-omitted line item in agentic TCO.

Why review is the hidden tax
Agentic governance means a human checks high-stakes actions before they ship. That labor scales with escalation volume and loaded reviewer cost, and it never appears on a token bill, so most TCO estimates miss it entirely.
The math
Reviews/mo = actions x escalation pct. Labor = reviews x (minutes/60) x loaded hourly rate. Reviewer FTEs = review-hours / ~160 productive hours per month.
Drivers in this layer (7)
Agent actions / mo
Use your real monthly action or decision volume.
% escalated
High-stakes flows (financial, PII, policy) escalate more; low-risk flows escalate little.
Minutes / review
Time a few real reviews; simple approvals are 1-2 min, complex ones 5-10.
Reviewer $/hr advanced
Use your loaded rate, not base pay, typically 1.3-1.5x salary per hour.
Cost / missed incident advanced
Blend refunds, remediation, and reputational or compliance exposure for your domain.
Incidents / 1k unreviewed advanced
Estimate from past error rates on high-risk actions.
% incidents caught advanced
Mature review with clear rubrics catches 80-95 pct; rushed review less.

Compliance and audit

live engine

Prices the annual compliance and governance overhead of an agentic deployment across the five real cost centers: compliance staff, audit-log storage, conformity assessment and certification, external audit, and a governance platform. EU AI Act risk tier is the master switch that scales the rigor and triggers certification.

Why compliance is a cost center, not a checkbox
Regulated and high-risk AI requires bias audits, human-oversight workflows, fundamental-rights impact assessments, and auditor-grade record-keeping. That work is staff time, log storage, and external assessment, not just a software license.
The math
Annual = compliance staff (FTEs x loaded cost x tier rigor) + audit-log storage (GB per month x retention x rate x 12) + conformity assessment (high-risk only, per system, amortized) + external audit + governance platform.
Drivers in this layer (9)
EU AI Act risk tier
Most chatbots are minimal/limited; credit, hiring, clinical, and consequential agentic actions are high.
AI systems in scope
Count distinct production systems; certification cost scales with this for high-risk.
Compliance FTEs
Startups often run 0.25-1; larger programs run several.
FTE loaded $/yr advanced
Salary plus benefits and overhead, typically 1.3-1.5x base.
External audit $/yr advanced
External audits often run ~25 percent of recurring compliance cost.
Compliance platform $/yr advanced
SMB tools EUR 5K-50K; enterprise platforms can exceed EUR 100K.
Audit log GB / mo advanced
Estimate from current log volume; high-traffic agents generate more.
Log retention (mo) advanced
High-risk record-keeping commonly spans 12-60 months; check your obligations.
Log storage $/GB-mo advanced
Use your cloud storage rate; cold/archive tiers are cheaper.

Vendor lock-in cushion

manual entry

A cushion for relying on a single model vendor: the cost of an abstraction layer, a second supplier, or a fallback provider so one price change or shutdown cannot strand the system. Enter your best estimate of that monthly insurance.

Drivers in this layer (1)
Premium $/mo
Enter your own figure; the default is a sensible placeholder.

Downtime (lost revenue)

manual entry

The revenue you lose when the system is down. Hours unavailable in a month times the revenue you would normally earn each hour. It is not a bill you pay, but it is a real cost of running the service.

Drivers in this layer (2)
Hours down per month
Enter your own figure; the default is a sensible placeholder.
Revenue lost per hour down ($)
Enter your own figure; the default is a sensible placeholder.

Other costs

The smaller things that are easy to forget but add up.

Data prep and other extras

manual entry

The smaller things that never make the headline model bill: preparing and labeling data, the occasional rollback, and miscellaneous one-off costs. Each is small; together they add up.

Drivers in this layer (1)
Data prep + misc ($ per month)
Enter your own figure; the default is a sensible placeholder.

How it decides: the readiness checklist

Adding up the costs gives you a monthly total. That alone does not tell you whether to go ahead. So the tool reduces the expected payoff by the chance the project gets cancelled before it pays off - and roughly two in five AI agent projects do not make it. Every item you can honestly tick on the readiness checklist lowers that chance, because the things that keep a project alive are within your control.

The readiness checklist
  • Focused on a few high-value tasks (not everything at once)
  • We measured the 'before', so we can prove the improvement
  • We check the AI's quality regularly
  • Safety and compliance work is budgeted
  • One person clearly owns this project
  • Our data is clean and kept up to date

You get one of three answers. GO: the payoff clearly beats the cost, even after allowing for risk - build it. RISKY: the payoff works and it pays back, but the margin is thin or the project is not ready - de-risk and finish the checklist before you commit budget to scale. NO-GO: even before risk, the payoff does not cover the full cost - the case is not there yet, usually a sign the task is too low-value or the hidden costs were never budgeted.

Run your numbers →

Methodology & sources

Each cost is computed by its own validated calculator, so the ones tied to live pricing (model usage, embeddings, guardrails, quality checks, data-freshness) track the pricing catalog automatically, and yearly costs (compliance, integrations, model upkeep) are shown per month. The tool also divides the total by your volume to show cost per request, per 1,000 tokens, per user, and - the most useful number for most teams - cost per business result, such as per ticket resolved or per document processed. The risk-weighting and gate logic are shared with the express control tower, so the deep and express paths always agree on a verdict.

The framing draws on published industry research: Gartner's projection that a large share of agentic projects are abandoned by 2027; Bain's payback benchmarks showing measurable workflows recover their investment in months while diffuse ones take far longer; MIT Sloan's finding that stalled programs frequently had no evaluation in place and lost meaningful accuracy over time; and McKinsey and BCG/Forrester estimates of multi-month ramp periods and multi-year returns. The quality safety-net - quality checks catch problems while data-refresh and model-update budgets fix them - reflects a recurring lesson: catching problems without budgeting the fix is how accuracy quietly decays.

This guide and calculator are a budgeting and decision aid, not investment advice. Figures left on defaults are placeholders, not your data; refine the layers that matter before taking a number to a board.

📖 Data sources & methodology 158 text models · 9 embeddings · 37 vision · 55 audio · 8 vector DBs across 10 vendor pages · last verified 2026-07-21

Methodology

  • All prices are USD per 1 million tokens, current as of 2026-07-21.
  • Vendor-published values have no mark. Inferred/extrapolated values are marked with * and listed below.
  • Batch API discounts are 50% off standard rates across providers that offer Batch mode.
  • Prompt caching discounts vary by provider (typically 80-90% off cached input tokens).
  • Regional data-residency surcharges (Anthropic 1.1x, OpenAI 1.1x, Google regional tiers) are NOT included in base rates.
  • Long-context pricing tiers apply when input exceeds model threshold.
  • Embedding prices are input-only (no output tokens generated).

Primary sources

Last-verified date is the most recent successful daily snapshot (aicost_pricing_snapshots) or, when no snapshot exists yet, the latest successful crawler run (aicost_crawler_runs). 10 of 10 vendors are currently verified. Aggregator services (TokenCost, AI Pricing Guru, etc.) are not listed.

Anthropic
2026-07-21
https://www.anthropic.com/pricing
Daily snapshot since Sep 2023 · 624 days captured
Anthropic Docs
2026-07-21
https://platform.claude.com/docs/en/about-claude/pricing
Daily snapshot since Sep 2023 · 624 days captured
OpenAI
2026-07-21
https://openai.com/api/pricing/
Daily snapshot since Sep 2023 · 625 days captured
Google AI
2026-07-21
https://ai.google.dev/gemini-api/docs/pricing
Daily snapshot since Dec 2023 · 600 days captured
Google Vertex
2026-07-21
https://cloud.google.com/vertex-ai/generative-ai/pricing
Daily snapshot since Dec 2023 · 600 days captured
DeepSeek
2026-07-21
https://api-docs.deepseek.com/quick_start/pricing
Daily snapshot since May 2024 · 539 days captured
xAI
2026-07-21
https://x.ai/api
Daily snapshot since Nov 2024 · 457 days captured
Mistral
2026-07-21
https://mistral.ai/pricing
Daily snapshot since Dec 2023 · 598 days captured
Cohere
2026-07-21
https://cohere.com/pricing
Daily snapshot since Sep 2023 · 624 days captured

Inferred values (marked with * in calculator tables)

Derived from industry conventions, not directly published by the vendor. Typical conventions: cached input = 10% of base (90% off), Batch API = 50% of base (50% off).

Vendor / Model Field Why it’s inferred
Anthropic — Claude Sonnet 4.6 cachedInput Derived at 10% of input rate — Anthropic publishes 90% cache-hit discount on this tier.
Anthropic — Claude Sonnet 4.5 cachedInput Derived at 10% of input rate; same 90% cache-hit convention as Sonnet 4.6.
Anthropic — Claude Sonnet 4.5 batchInput Derived at 50% of standard input — Anthropic documents uniform 50% Batch discount.
Anthropic — Claude Sonnet 4.5 batchOutput Derived at 50% of standard output — Anthropic documents uniform 50% Batch discount.
Anthropic — Claude Haiku 4.5 cachedInput Derived at 10% of input rate — Anthropic 90% cache-hit discount convention.
OpenAI — GPT-5.4 Mini cachedInput Derived at 10% of input — OpenAI documents automatic 90% discount on cache hits across GPT-5.x tier.
OpenAI — GPT-5.4 Nano cachedInput Derived at 10% of input — OpenAI 90% cache-hit convention.
OpenAI — GPT-5.4 Nano batchInput Derived at 50% of input — OpenAI Batch API uniform 50% discount.
OpenAI — GPT-5.4 Nano batchOutput Derived at 50% of output — OpenAI Batch API uniform 50% discount.
OpenAI — GPT-5.4 Pro cachedInput Derived at 10% of input — OpenAI 90% cache-hit convention.
OpenAI — GPT-5.4 Pro batchInput Derived at 50% of input — OpenAI Batch API uniform 50% discount.
OpenAI — GPT-5.4 Pro batchOutput Derived at 50% of output — OpenAI Batch API uniform 50% discount.
OpenAI — GPT-5.2 cachedInput Derived at 10% of input; no residency uplift.
OpenAI — GPT-5.2 batchInput Derived at 50% of input.
OpenAI — GPT-5.2 batchOutput Derived at 50% of output.
OpenAI — GPT-5 cachedInput Derived at 10% of input.
OpenAI — GPT-5 batchInput Derived at 50% of input.
OpenAI — GPT-5 batchOutput Derived at 50% of output.
OpenAI — GPT-5.5 Pro cachedInput Derived at 10% of input — OpenAI does not publish a cached rate for *-pro models; using the family convention.
OpenAI — GPT-5.5 Pro batchInput Derived at 50% of input.
OpenAI — GPT-5.5 Pro batchOutput Derived at 50% of output.
OpenAI — GPT-5.2 Pro cachedInput Derived at 10% of input — pro-tier convention.
OpenAI — GPT-5.2 Pro batchInput Derived at 50% of input.
OpenAI — GPT-5.2 Pro batchOutput Derived at 50% of output.
OpenAI — GPT-5.1 batchInput Derived at 50% of input.
OpenAI — GPT-5.1 batchOutput Derived at 50% of output.
OpenAI — GPT-5 Pro batchInput Derived at 50% of input.
OpenAI — GPT-5 Pro batchOutput Derived at 50% of output.
OpenAI — GPT-5 Nano cachedInput Derived at 10% of input.
OpenAI — GPT-5 Nano batchInput Derived at 50% of input.
OpenAI — GPT-5 Nano batchOutput Derived at 50% of output.
Google — Gemini 3 Flash cachedInput Derived at 10% of input — Google caching discount convention ~90%.
Google — Gemini 3.1 Flash-Lite cachedInput Derived at 10% of input — Google caching convention.
Google — Gemini 3.1 Flash-Lite batchInput Derived at 50% of input — Google Batch API uniform 50% discount.
Google — Gemini 3.1 Flash-Lite batchOutput Derived at 50% of output — Google Batch API uniform 50% discount.
Google — Gemini 2.5 Pro cachedInput Derived at 10% of input.
Google — Gemini 2.5 Flash cachedInput Derived at 10% of input.
Google — Gemini 2.5 Flash-Lite cachedInput Derived at 10% of input — Google caching convention.
Google — Gemini 2.5 Flash-Lite batchInput Derived at 50% of input — Google Batch API uniform 50% discount.
Google — Gemini 2.5 Flash-Lite batchOutput Derived at 50% of output — Google Batch API uniform 50% discount.
Google — Gemini 2.0 Flash cachedInput Derived at 25% of input per Google 2.0 family caching rates.
Google — Gemini 2.0 Flash batchInput Derived at 50% of input — Google Batch API uniform 50% discount.
Google — Gemini 2.0 Flash batchOutput Derived at 50% of output — Google Batch API uniform 50% discount.
Google — Gemini 2.0 Flash-Lite cachedInput Derived at 10% of input — Google caching convention.
Google — Gemini 2.0 Flash-Lite batchInput Derived at 50% of input — Google Batch API uniform 50% discount.
Google — Gemini 2.0 Flash-Lite batchOutput Derived at 50% of output — Google Batch API uniform 50% discount.
xAI — Grok 4 (legacy) cachedInput Extrapolated at 25% of base.

Pricing is cross-verified against the LiteLLM community registry when available. Daily snapshots are kept in aicost_pricing_snapshots; every change is logged to aicost_price_changelog with old & new values for full audit trail. Read the full methodology →