AI Model Pricing in 2026: A Complete Guide

AI API bills are usually more predictable than they first appear — once you understand how providers actually charge for tokens. This guide walks through the mechanics, the current rates across OpenAI, Anthropic, and Google, and the handful of levers that make the biggest difference to your monthly bill.

Quick answer

All three major providers bill per token, with output tokens costing 4-8x more than input. As of August 2026, flagship models (Claude Opus 5, GPT-5.5, GPT-5.6 Sol) run $5.00 input / $25.00-$30.00 output per million tokens; mid-tier models (Claude Sonnet 5, Gemini 3.1 Pro) run $2.00-$3.00 input / $10.00-$15.00 output; budget models (Claude Haiku 4.5, GPT-5.6 Luna, Gemini 3.1 Flash-Lite) cost under $1.50 input. Prompt caching, batch processing, and routing simple tasks to cheaper models are the three biggest levers for cutting your bill, often by 50% or more combined.

Key takeaways

• Output tokens cost 4-8x more than input tokens on every major provider — a generation-heavy workload costs far more than a read-heavy one at the same total token count.
• Claude Sonnet 5’s introductory $2.00 / $10.00 pricing ends August 31, 2026, reverting to $3.00 / $15.00 on September 1.
• OpenAI’s GPT-5.6 family (Sol, Terra, Luna) launched July 9, 2026, then got a mid-cycle price cut on July 30 — Luna alone dropped 80%.
• Prompt caching cuts repeated-context input costs by roughly 90% across Anthropic, OpenAI, and Google.
• Batch processing is a flat 50% discount on every major provider in exchange for asynchronous, non-real-time delivery.
• Routing easy tasks to a budget model and reserving the flagship for hard reasoning typically cuts blended costs 30-40% without a noticeable quality drop.

What determines AI API pricing

Every major provider bills per token, not per request. A token is roughly four characters of English text, so 1,000 tokens works out to about 750 words (see our tokens guide for the full mechanics). Two things set the price for a given model: which tier it sits in (budget, mid, or flagship), and whether the tokens are input or output.

Input vs. output tokens: why the split matters

Input tokens are everything you send the model: your prompt, system instructions, conversation history, retrieved context, and tool definitions. Output tokens are what the model generates back.

Output tokens cost significantly more than input tokens on every provider — typically four to six times as much. Anthropic keeps a consistent 5x ratio across its whole lineup, which makes back-of-envelope budgeting easy. OpenAI’s ratio varies more by tier, from about 4x on nano models up to 6x on the flagship. Google’s Gemini models generally sit in the 6-8x range.

The practical takeaway: a read-heavy workload (summarizing documents, answering questions about existing text) is far cheaper than a generation-heavy workload (writing long-form content, producing detailed reports) at the same total token count. If you’re building something that generates long output, that’s where most of your bill will come from.

Current pricing across the major providers

Rates below are standard, non-cached, non-batch pricing as of August 2026. Every provider changes prices periodically, so treat this as a snapshot rather than a permanent reference — our AI Model Cost Calculator pulls from the same figures and is easier to keep bookmarked than a screenshot of this table.

OpenAI

• GPT-5.6 Sol (current flagship, launched July 9, 2026) — $5.00 input / $30.00 output per million tokens
• GPT-5.5 (previous flagship, still available at the same rate) — $5.00 / $30.00
• GPT-5.6 Terra (mid-tier, cut 20% on July 30, 2026) — $2.00 / $12.00
• GPT-5.4 (mid-tier, half the price of 5.5 on both sides) — $2.50 / $15.00
• GPT-5.6 Luna (budget tier, cut 80% on July 30, 2026) — $0.20 / $1.20
• GPT-5.4 Nano (older budget tier) — $0.20 / $1.25

GPT-5.6 is gradually superseding GPT-5.4 as OpenAI’s default mid- and budget-tier lineup; the 5.4 models remain callable but aren’t seeing further price cuts. OpenAI also offers Batch and Flex processing at roughly 50% off standard rates for workloads that can tolerate delay, and a Priority tier at a premium for latency-sensitive production traffic.

Anthropic

• Claude Opus 5 (flagship, complex reasoning and agentic coding — replaced Opus 4.8 at the same rate) — $5.00 / $25.00
• Claude Sonnet 5 (best price-to-performance in the lineup) — $2.00 / $10.00 introductory through August 31, 2026, then $3.00 / $15.00
• Claude Haiku 4.5 (fastest and cheapest current-generation model) — $1.00 / $5.00

Anthropic’s pricing is unusually predictable: every current model keeps the same 5x output-to-input ratio, and prompt caching cuts cached input to roughly 10% of the standard rate.

Google Gemini

• Gemini 3.1 Pro (flagship reasoning, 1M context) — $2.00 / $12.00 up to 200K tokens of input, doubling to $4.00 / $18.00 beyond that
• Gemini 3.6 Flash (frontier-adjacent speed, native grounding — replaced 3.5 Flash on July 21, 2026) — $1.50 / $7.50
• Gemini 3.1 Flash-Lite (budget tier) — $0.25 / $1.50

Google isn’t the only provider with a context-length cliff, though the exact terms differ. Gemini 3.1 Pro doubles its rate — $2.00 / $12.00 to $4.00 / $18.00 — once a request crosses 200K input tokens, and that higher rate applies to the entire request, not just the tokens past the threshold. OpenAI’s GPT-5.5 and GPT-5.6 models apply a similar penalty at a higher threshold: cross 272K input tokens and the whole request bills at 2x input / 1.5x output. Anthropic’s current lineup (Opus 5, Sonnet 5, Haiku 4.5) doesn’t use stepped context pricing — the quoted rate applies regardless of how much context you send. If your workload regularly sits near one of these thresholds, trimming duplicate retrieval chunks or stale conversation history before the request is often cheaper than accepting the higher tier.

Worth designing around if you’re working with long documents.

How to estimate your monthly AI costs

The formula is the same regardless of provider:

(monthly input tokens ÷ 1,000,000 × input price) + (monthly output tokens ÷ 1,000,000 × output price)

The hard part isn’t the math — it’s estimating your token volume before you have real usage data. A reasonable starting point: multiply your expected daily requests by your average input and output tokens per request, then multiply by 30. Once you have even a few days of real traffic, replace the estimate with actual numbers from your provider’s usage dashboard, since actual usage is often higher than planning estimates once conversation history, retries, and tool definitions are factored in.

A worked example

Say a customer-support tool sends Claude Sonnet 5 an average of 2,000 input tokens (conversation history plus a system prompt) and receives 400 output tokens back, across 15,000 requests a month. That’s 30 million input tokens and 6 million output tokens monthly.

At Sonnet 5’s introductory rate: (30 × $2.00) + (6 × $10.00) = $60.00 + $60.00 = $120.00 per month. After the rate reverts to $3.00 / $15.00 on September 1: (30 × $3.00) + (6 × $15.00) = $90.00 + $90.00 = $180.00 — a 50% jump worth planning for in advance rather than discovering on the next invoice. If half the input tokens are a repeated system prompt eligible for caching, the cached portion drops to roughly 10% of the standard input rate, pulling the post-September bill down to $139.50 — most of the way back toward the pre-reversion cost. For comparison, that same 30 million input / 6 million output token workload on Claude Haiku 4.5 would cost $60.00 total — a third of Sonnet 5’s standard rate. Whether that trade is worth it depends entirely on whether Haiku’s quality holds up on this specific support workload, which is exactly the kind of validation the earlier section on choosing a model recommends before committing production traffic.

Five ways to cut your AI API bill

Prompt caching

If your application repeatedly sends the same system prompt, few-shot examples, or document prefix, caching lets subsequent requests read that content at a fraction of standard input price — roughly 90% off on Anthropic and comparable savings on OpenAI and Google. This is usually the single biggest lever for chatbots and agents with stable instructions.

Batch processing

Every major provider offers an asynchronous batch mode at roughly 50% off standard pricing, with results returned within 24 hours. Good for anything that doesn’t need a real-time response — nightly data processing, bulk content generation, evaluation runs.

Model routing

Not every request needs your most expensive model. Routing simple classification, extraction, and short-response tasks to a budget model while reserving the flagship for genuinely hard reasoning can cut blended costs 30-40% without a noticeable quality drop for most users.

Trim what you send, not just what you generate

Conversation history, retrieved document chunks, and tool definitions all count as input tokens, and they accumulate fast in multi-turn applications. Dropping stale conversation turns, sending only the relevant excerpt of a retrieved document instead of the whole file, and pruning unused tool definitions from the request all reduce input cost directly — and, on providers with a long-context cliff, can keep a request under the threshold that triggers a full-request rate increase.

Cap output length where verbose responses aren’t needed

Since output tokens cost 4-8x more than input, an uncapped or loosely-prompted model that writes longer than necessary is one of the more expensive habits to leave unaddressed. Setting a max-output-tokens limit and asking explicitly for concise responses in the system prompt both help, particularly on models with visible reasoning tokens, where the reasoning itself bills as output even though the user never sees it.

Which model should you choose?

Cheapest-per-token is rarely the right question on its own. A model that requires more retries, longer prompts to get the same result, or produces output you have to edit by hand can end up costing more in total than a pricier model that gets it right the first time. Treat the numbers in our AI Model Comparison Tool as your starting filter, then validate quality on your actual use case before committing production traffic to any single model.

As a general starting point: route high-volume, low-complexity tasks (classification, tagging, short summaries) to a budget-tier model; use a mid-tier model for most production workloads; and reserve flagship models for the requests where getting it right the first time is worth the premium. If you’re deciding between specific providers rather than tiers, our Claude vs. ChatGPT vs. Gemini comparison breaks down the differences beyond price.

Common mistakes when budgeting for AI costs

Estimating from input tokens alone

Because output costs 4-8x more than input, a workload that looks cheap based on prompt size alone can blow past budget once you factor in what the model actually generates. Long-form generation, detailed reports, and reasoning-heavy tasks with long chain-of-thought output are the usual culprits.

Defaulting every request to the flagship model

Classification, short extraction, and simple formatting tasks rarely need a $5.00 / $25.00 model. Routing that traffic to a budget tier is usually the single biggest lever available — see our multi-model routing workflow for how to set it up.

Ignoring the introductory-pricing expiration date

Providers regularly launch new models at a temporary introductory rate, then revert to standard pricing on a fixed date — Claude Sonnet 5’s $2.00 / $10.00 rate through August 31, 2026 is a current example. Budgets built on an introductory rate without a reminder for the reversion date routinely get blindsided by a 50% cost jump.

Skipping prompt caching on repeated context

If your application resends the same system prompt, few-shot examples, or document prefix on every call, that’s a roughly 90% cost reduction sitting unused. It requires minimal code changes on every major provider and is one of the few optimizations with essentially no tradeoff.

Frequently asked questions

Why did Claude Sonnet 5 launch below Anthropic’s usual $3.00 / $15.00 rate?

It’s a temporary introductory rate through August 31, 2026 — the first time Anthropic has priced a production Sonnet tier below its long-standing $3.00 / $15.00 anchor. Budget for the reversion to $3.00 / $15.00 on September 1 if you’re building sustained usage on Sonnet 5.

Is a newer, more expensive model always worth switching to?

Not automatically. Test the new model against your actual workload before migrating — a price-matched replacement like GPT-5.6 Sol for GPT-5.5 is a safe swap, but a genuinely pricier upgrade only pays for itself if it measurably reduces retries, shortens prompts, or removes manual editing work downstream.

How much does prompt caching actually save in practice?

Roughly 90% off the input rate for cached tokens across Anthropic, OpenAI, and Google. For an application that resends a large system prompt or document on every call, that can cut total input cost by more than half once averaged across a real request pattern.

Do all providers charge the same for context above 200K tokens?

No — this varies by provider and is worth checking before committing to a model for a long-document use case. Google’s Gemini 3.1 Pro doubles its rate to $4.00 / $18.00 once a request crosses 200K input tokens; the same behavior doesn’t apply uniformly across every OpenAI and Anthropic model, so check the specific model’s pricing page rather than assuming.

Should I budget in flat monthly terms or per-token?

Per-token, even if you report the total as a monthly number internally. Token-based billing means your actual cost scales directly with usage, so a monthly estimate is only as good as your token volume estimate — see our guide to setting an AI budget for your team for how to build that estimate and revisit it against real usage.

What happened to older models like GPT-5.4 and Claude Opus 4.8?

They’re both still usable, but newer models have effectively superseded them at the same or better price. GPT-5.6 Sol matches GPT-5.5’s rate, and Claude Opus 5 replaced Opus 4.8 as Anthropic’s recommended default at the identical $5.00 / $25.00 rate — there’s little reason to specifically choose the older version for new projects.

Does switching providers mean rewriting my whole integration?

Usually not entirely. Several providers and third-party routers offer OpenAI-compatible endpoints, so swapping the base URL and model name handles basic text completion in many cases. Provider-specific features — extended thinking, native tool schemas, prompt caching syntax, multimodal handling — don’t map over automatically, so budget time to test those separately rather than assuming a drop-in replacement.

Do enterprise or volume agreements change these prices?

The rates in this guide are standard self-serve API pricing, publicly listed by each provider. Enterprise agreements with committed volume can include different terms, but those aren’t publicly published and vary by negotiation — treat the numbers here as your self-serve baseline, and talk to the provider’s sales team directly if you’re evaluating a large, sustained commitment.

Next steps

AI pricing moves fast, but the structure behind it doesn’t: input and output tokens, tiered models, and a handful of discount levers you either use or leave on the table. Bookmark our AI Model Cost Calculator to run your own numbers whenever you’re evaluating a new model or workload — it pulls from the same rates as this guide, and we keep both updated as providers change pricing. For a printable side-by-side reference, see our AI Model Pricing Comparison Chart.