LLM API Pricing Comparison 2026: OpenAI vs Claude vs Gemini vs Grok vs Mistral

Picture a developer who launches an AI product in October 2026 on Google's Gemini 4 Argon. At $2 per million input tokens and $10 per million output, it's one of the best-value flagship models available.

On 1 December, the price doubles to $4 and $20.

Nothing about the product changes. The customers don't change. The bill does. Google calls it introductory pricing, and it's printed right there on the pricing page.

In the book, I described Stripe's classic offer to growing merchants: commit to us for twelve months and we'll take less of your money. Model vendors play a version of the same game. Build on our low price today, and your switching costs will do the rest.

That's why you need a current, side-by-side view before you pick a model, and why you should check it again before every price review. Here it is.

LLM API prices at a glance (October 2026)

Prices checked 27 October 2026, per million tokens, standard (non-batch) rates. They change often; see the AI Pricing Index for the live tracker.

Each vendor has a full teardown, covering consumer plans too: OpenAI, Anthropic, Google, SpaceXAI and Mistral.

What does a real task cost on each model?

List prices are hard to compare because input and output are priced differently. So here's one standard request, priced on every model: 2,000 input tokens and 500 output tokens, roughly a short customer email plus some instructions and a reply.

Read down the "per 100,000 requests" column. The same workload costs anywhere from $45 to $4,500 a month. That's the most important number on this page.

Now, a quality caveat. These are price comparisons, not quality comparisons. A cheaper model that gets the answer wrong, or needs three retries, isn't cheaper. Test on your real tasks.

Claude vs ChatGPT: how does Anthropic's token pricing compare with OpenAI's?

This is one of the most searched questions about AI pricing, and in October 2026 the answer is surprisingly simple. At matching tiers, the list prices are the same.

Both take 50% off for batch jobs. Both run their consumer chat plans at $20 a month.

So where does the real difference come from?

• Caching mechanics. OpenAI discounts cached input by 90–95%. Anthropic charges about 5% of the input price for cache hits, but charges 1.25x input to write a five-minute cache and 2x for one hour. If your cache hit rate is high, Anthropic's frontier caching is cheaper; if you rarely reuse a prefix, the write fees add up.
• Long context. Anthropic offers 1M-token context at standard prices on Claude 4.6 and later (except Haiku 5.5). Check the long-prompt rules on any model you plan to feed big documents.
• Token counts. Each vendor uses its own tokenizer, so the same text can count as a different number of tokens. Run your own prompts through both and compare the bills.
• Tools. Anthropic charges $10 per 1,000 web searches plus tokens. Code execution is free alongside web search or fetch, otherwise 1,550 free hours a month then $0.05 per container-hour.

Gemini, Grok and Mistral: where's the value?

Google Gemini has the widest range and the most moving parts. The Flash-Lite models are among the cheapest capable models anywhere, and many Flash models have a free tier. But watch three things: the scheduled doublings on Gemini 4 Argon and 3.8 Flash, long-context surcharges on the Pro models above 200,000 tokens, and grounding with Google Search at $14 per 1,000 requests after 5,000 free a month. Priority processing costs about 1.8x standard.

SpaceXAI's grok-4.7 lists at $2 in and $6 out, which makes its output notably cheaper than the other workhorse models. Tool calls are reported at $5 per 1,000. Note the vendor's volatility: SpaceXAI retired older API models in May 2026 and routed them to newer ones, and it raised text-to-speech from $4.20 to $15 per million characters in Q2 2026, according to AlloyPress.

Mistral Large costs $0.50 in and $1.50 out for an open-weight flagship, a fraction of the US labs' workhorse prices. Mistral also sells private deployments and EU hosting, which is the real reason many buyers choose it.

Open models on inference platforms. If you run open-weight models through a host such as DeepInfra, you'll usually pay per million tokens with options to trade speed for price: DeepInfra charges 1.5x for "priority" and 0.8x for "flex." See the DeepInfra teardown.

The hidden line items that change the ranking

The list price is only the start. These are the multipliers that decide your real bill.

For a customer-support bot that resends the same 5,000-token instructions on every reply, caching alone cut the bill in my token pricing worked example from $166 to $71 a month.

Which LLM API is cheapest for your use case?

Here's how I'd shortlist.

The best answer is usually more than one model. Route each task to the cheapest one that passes your quality test.

How to compare LLM costs properly

1. Use your own prompts. Take 50 real tasks and run them through each shortlisted model.
2. Measure actual tokens, including cached input and reasoning tokens, from the vendor's usage report.
3. Include retries and failures. A model that needs a second attempt 20% of the time costs 20% more.
4. Price at announced future rates, not launch rates.
5. Add the extras: search, tools, residency and storage.

What does this mean for your own pricing?

If you're building on these APIs, don't pass the list price straight through to your customers. That makes you a reseller with no margin, and the AI wrapper guide explains why that's a fragile place to be.

Use these prices as your cost floor. Then price on the value of the job, with an allowance and an overage, the way I lay out in the AI pricing guide. And when vendors cut their prices, as Anthropic has done repeatedly with Opus, keep some of the saving as margin rather than cutting your own price on day one.

One more thing from the book: minimum commitments work in both directions. Model and infrastructure vendors give better rates for committed spend (Deepgram's Growth plan, for example, cuts unit rates about 10–15% for a prepaid commitment). Once your volumes are steady, ask.

What to do next

1. Shortlist two or three models using the use-case table above.
2. Run your real tasks through each and record tokens, quality and retries.
3. Cost them at today's and announced future prices with the AI Token Cost Calculator.
4. Design for caching and batch from the start.
5. Check your margin with the AI Gross Margin Calculator before you publish a price.

And put a reminder in your calendar for the first of every month. In this market, a price you checked in October can be wrong by December.

Frequently asked questions

Get both books free

The Subscription Playbook (ebook + audiobook) and Feed a Starving Crowd, instantly. In Australia? Claim the free hardcover.

Leave this empty

First name

Email

CountryAustraliaNew ZealandUnited StatesUnited KingdomCanadaOther

What best describes you?Business owner (services)Business owner (retail / e-commerce)SaaS or AI founderHealth or allied health practiceTrades or home servicesAgency or consultantAccountant or adviserEmployee / marketerStudent / other

Yes, I'd like a free 30-minute Subscription Profit Review with a BPAG Accountants adviser (Australian businesses).

Your details go to Subscription Playbook and our sponsor BPAG Accountants Pty Ltd, who may contact you about the book and, if you ticked the box, your review. Unsubscribe any time. Privacy policy.

Send me the free books

Robert Coorey, MBA

Robert is the author of The Subscription Playbook and Feed a Starving Crowd and a Director at BPAG Accountants in Sydney. More about Robert.