Per-token prices have fallen roughly 80% over the past year, and yet most enterprise AI bills went up. Industry estimates put enterprise LLM API spend past $8 billion in 2025, on track to roughly double again this year. That is the paradox every finance manager is now living inside: the unit you buy keeps getting cheaper while the invoice keeps getting bigger. Cutting the bill is not about finding the provider with the lowest sticker price. It is about understanding how each provider charges, where the meter runs faster than you expect, and which three levers actually change the number at the bottom of the invoice.
This is a rate-card comparison as of July 2026, refreshed from the ground up because the models and prices in circulation a year ago are gone. GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 have all been retired or superseded. If a comparison you are reading still quotes those, it is costing you money, because it is anchoring your budget to a market that no longer exists.
What you actually pay for in 2026
All three providers still charge per token, and a token is still roughly four characters of English, about 0.75 words. Beyond that, the pricing model has picked up a few wrinkles that matter more than the headline rate.
Input, output, and the 5x-to-6x spread. Output tokens cost far more than input tokens across every provider. Claude Opus 4.8 charges $5 per million input tokens and $25 per million output, a clean 5x multiplier. GPT-5.6 Sol runs $5 input against $30 output, a 6x spread. This ratio is the single most important number for architecture decisions. An application that reads a lot and writes a little (classification, extraction, routing) has a very different cost profile than one that generates long documents, and the same monthly request volume can produce wildly different bills depending on which side of that ratio your workload sits on.
Cached input is now a first-class price tier. Every major provider charges a steeply discounted rate for input tokens it has already processed. Anthropic bills cache reads at $0.50 per million against a $5 base for Opus 4.8, roughly 90% off. OpenAI’s cached input on the GPT-5.6 family drops to $0.50 against a $5 base, again about 90% off. Google offers context caching with its own discount structure. If your prompts carry a large stable prefix (a system prompt, a policy document, a knowledge base snippet) and you are not caching it, you are paying full freight on tokens the provider is not actually reprocessing.
The tokenizer wrinkle almost nobody budgets for. Anthropic’s newer models, starting with Opus 4.7 and including Fable 5 and Sonnet 5, use a tokenizer that produces approximately 30% more tokens for the same text than the previous generation. The per-token price can look competitive while the effective cost per page of English runs higher, because there are more tokens in that page. When you benchmark providers, compare cost per completed task on your own prompts, not cost per token off the rate card. The rate card can lie by omission.
Context tiers. Large context windows are cheap to advertise and expensive to fill. Google’s Gemini 3.1 Pro charges $2 per million input tokens up to 200K tokens, then $4 above that threshold. Anthropic includes its full 1M-token window at standard pricing on Opus and Sonnet, with no surcharge above 200K. That difference is invisible until you build a document-analysis workload that routinely pushes past the threshold, at which point it becomes a line item.
The current rate card, July 2026
The table below covers the workhorse tiers, the models most enterprises actually route production traffic through. Flagship reasoning models and specialty tiers exist above these, but for cost planning this is the band that matters.
| Model | Input ($/1M) | Cached input ($/1M) | Output ($/1M) | Context | Where it fits |
|---|---|---|---|---|---|
| GPT-5.6 Sol (OpenAI) | $5.00 | $0.50 | $30.00 | Large | Flagship general workloads |
| GPT-5.6 Terra (OpenAI) | $2.50 | $0.25 | $15.00 | Large | Mid-tier production |
| GPT-5.6 Luna (OpenAI) | $1.00 | $0.10 | $6.00 | Large | Cost-sensitive volume |
| GPT-5.4 Mini (OpenAI) | $0.75 | $0.075 | $4.50 | Large | High-volume, light tasks |
| GPT-5.4 Nano (OpenAI) | $0.20 | $0.02 | $1.25 | Large | Classification, routing |
| Claude Opus 4.8 (Anthropic) | $5.00 | $0.50 | $25.00 | 1M | Complex reasoning, coding |
| Claude Sonnet 4.6 (Anthropic) | $3.00 | $0.30 | $15.00 | 1M | Most production workloads |
| Claude Haiku 4.5 (Anthropic) | $1.00 | $0.10 | $5.00 | 200K | Fast, lightweight tasks |
| Gemini 3.1 Pro (Google) | $2.00 / $4.00* | ~10% of base | $12.00 / $18.00* | 1M+ | Long-document processing |
| Gemini 3.5 Flash (Google) | $1.50 | discounted | $9.00 | Large | High-volume balanced |
| Gemini 3.1 Flash-Lite (Google) | $0.25 | discounted | $1.50 | Large | Cheapest per token |
*Gemini 3.1 Pro: lower rate applies to prompts at or under 200K tokens, higher rate above.
A note on introductory pricing: Anthropic’s Claude Sonnet 5 carries introductory rates of $2 input and $10 output per million tokens through August 31, 2026, after which it moves to $3 and $15, matching Sonnet 4.6. If you are budgeting a Sonnet 5 workload into the fourth quarter, plan for the higher number, not the introductory one.
The cheapest sticker price in the room belongs to Gemini 3.1 Flash-Lite at $0.25 input and $1.50 output. That does not make it the cheapest answer. Cheapest per token and cheapest per completed job are different questions, and the gap between them is where most budget surprises live. A model that is half the price but needs a second pass, a human review, or a retry on a third of requests is not half the price.
Why cheaper tokens have not shrunk your bill
The deflation is real. So is the bill going up. Both are true at once, and the reason is consumption, not price.
Agentic workloads are the driver. A single chatbot answer might consume a few thousand tokens. An agent that plans a task, calls tools, reads results, and iterates can burn 5 to 30 times that for the same business outcome, according to Gartner’s read on the shift. When a task that used to be one API call becomes a loop of fifteen, an 80% price cut does not save you money. It gets swallowed whole and then some. Inference is now estimated to consume around 85% of enterprise AI budgets, and a majority of AI projects still run 30% to 50% over their original cost estimates.
This is the part a static rate-card comparison misses entirely, and it is the part that matters most to a CFO. The question is not “which provider has the lowest per-token price this quarter.” Prices will keep falling. The question is “which architecture keeps my token consumption from outrunning the price cuts.” I spent two decades in IT operations, including years running telecom infrastructure in Saskatchewan where every circuit had a metered cost, and the lesson transfers cleanly: the cheapest per-unit rate is worthless if nobody is watching the volume. AI inference is a metered utility now. Treat it like one.
The three levers that actually move the number
Once you accept that consumption is the problem, the optimization stops being about vendor selection and starts being about how you run whichever vendor you pick. Three levers do most of the work.
Batch processing cuts non-urgent work in half. OpenAI, Anthropic, and Google all offer a batch or asynchronous tier at roughly 50% off standard rates, with turnaround measured in hours rather than milliseconds. Anything that does not need a real-time answer (overnight document processing, bulk classification, report generation, backfills) belongs here. Most enterprises run 100% of traffic through the real-time endpoint out of habit and pay double for the portion that could have waited.
Prompt caching pays for itself after one reuse. When a large chunk of your prompt is stable across requests, caching it drops the read cost to roughly 10% of the base input rate on both OpenAI and Anthropic. The math is decisive. Anthropic charges a 1.25x premium to write a 5-minute cache and 0.1x to read it, so caching turns profitable after a single reuse. For a customer-support application replaying the same 4,000-token policy prompt across a million monthly requests, this is the difference between a rounding error and a real line item.
Model tiering routes each task to the cheapest model that can do it. Not every request needs the flagship. Classification and routing run fine on Gemini 3.1 Flash-Lite or GPT-5.4 Nano at a fraction of flagship cost. Reserve Opus 4.8 or GPT-5.6 Sol for the reasoning-heavy escalations where quality directly moves a business outcome. An AI gateway that automates this routing can cut inference costs meaningfully without touching quality on the requests that matter. Combined, these three levers can compress total inference cost by roughly 5x to 10x versus running every request through a flagship model in real time.
A worked example makes the scale concrete. Take 100,000 monthly support conversations averaging 3,700 tokens each. Run entirely through a flagship model in real time, that is a five-figure monthly bill. Route the 70% of routine interactions to a Haiku- or Flash-Lite-class model, cache the shared policy prompt, and batch the overnight analytics, and the same volume lands at a fraction of the cost while the flagship still handles the escalations. The savings came from operations, not from switching logos.
Where the list price bends
Published rates are retail. Enterprise agreements move them, but the leverage points have shifted in 2026.
Marketplace billing changes the procurement path. Claude models now bill through AWS Marketplace and Microsoft Foundry using consumption units that convert token usage into a single line item on your existing cloud invoice, with negotiated discounts applied at conversion. If you already have committed spend with AWS or Azure, routing AI consumption through the marketplace can fold it into agreements you have already negotiated rather than opening a separate vendor relationship. Google’s Vertex AI does the equivalent for Gemini and, notably, also serves Claude models, so a single Vertex contract can cover multiple providers.
The negotiation levers that still work. Competitive bids remain the strongest tool, and they are more credible now that three providers ship genuinely substitutable workhorse tiers. Multi-year committed-use discounts trigger the best pricing but should carry exit clauses, because the model you commit to today will be two generations old within a year at the current release cadence. The same discipline that governs any cloud contract negotiation applies here: never sign to a single provider’s roadmap when the whole market reprices every quarter.
One caution from the fractional-COO work I do through Ops Harmony: enterprises consistently over-commit to annual minimums based on pilot-phase consumption, then discover production usage moved in a different direction. Commit to the floor you are certain of, buy the rest on demand, and revisit at renewal. An annual commitment sized to an optimistic forecast is just prepaid waste.
A decision framework for 2026
The single-provider strategy is dead for cost-conscious teams. The three workhorse tiers are close enough in capability and price that lock-in buys you nothing and costs you leverage. A practical posture:
- Default to a dual-provider setup. One flagship for reasoning and generation where quality is the product, one low-cost tier for the high-volume routine traffic that makes up most of your request count. The two do not need to come from the same vendor.
- Instrument before you optimize. You cannot route, cache, or batch what you cannot see. Per-request logging with token counts and cost attribution by team and use case is the prerequisite for every other move. This is the “inform” phase of the FinOps Foundation framework applied to AI cost management, and most organizations skip it and wonder why their bill is opaque.
- Re-benchmark quarterly. Prices fall and models turn over fast enough that last quarter’s optimal routing is this quarter’s overspend. Build the review into a standing cadence, not a fire drill triggered by a bad invoice.
- Codify the rules. A written AI spending policy with budget alerts, per-team rate limits, and default-cheap model routing keeps the optimization from decaying the moment the person who set it up moves on.
The providers will keep leapfrogging each other on price and capability, and that is fine. Your cost outcome does not depend on picking this quarter’s winner. It depends on running whichever provider you pick like the metered utility it has become.
Common questions
Which AI API is cheapest for enterprise use in 2026?
On sticker price, Google’s Gemini 3.1 Flash-Lite is lowest at $0.25 input and $1.50 output per million tokens, with OpenAI’s GPT-5.4 Nano close behind. But cheapest per token rarely means cheapest per completed task. For workloads where a budget model needs review or retries, a mid-tier model like Gemini 3.5 Flash, GPT-5.6 Luna, or Claude Haiku 4.5 often costs less all-in.
How much does a flagship model cost per month for a real application?
Consumption dominates, so ranges are wide. A customer-facing chatbot handling 100,000 monthly conversations at roughly 3,700 tokens each lands in the low five figures per month on a flagship tier at 2026 rates. Document-heavy or agentic workloads that loop through multiple model calls per task can run several times higher. Pilot first to establish your own token-per-task numbers before you commit an annual budget.
Do batch and caching discounts stack?
Yes. All three providers offer roughly 50% off for batch processing and steep discounts for cached input, and on Anthropic and OpenAI the two combine. A non-urgent workload with a large stable prompt prefix can see both discounts at once, which is how teams reach the 5x-to-10x total reduction that operations, not vendor switching, delivers.
What is the difference between Claude, GPT, and Gemini pricing at the flagship tier?
As of July 2026, Claude Opus 4.8 runs $5 input and $25 output per million tokens, GPT-5.6 Sol runs $5 and $30, and Gemini 3.1 Pro runs $2 to $4 input and $12 to $18 output depending on context length. Watch the Anthropic tokenizer, which produces about 30% more tokens per page than the previous generation, so compare on cost per task using your own prompts. For a fuller breakdown across every tier, see the guide to LLM API costs.