AI Token Prices in 2026: What Claude, Gemini, ChatGPT, DeepSeek, Qwen and Kimi Actually Cost

The headline event this month is Kimi K3, which lands at roughly Opus-class performance for Sonnet-class money. It is the most interesting price-to-performance story of 2026 so far, and it has a section of its own below. The bill only becomes real at scale, or when you start running agents.

Why I'm Writing This

I run a digital marketing and AI technology business, and the question I get more than any other right now is some version of "which AI is cheapest?" Usually from someone who has just seen an API invoice, or who is trying to work out whether building an AI content workflow will cost them a fortune.

The honest answer is that most marketers are confused about AI pricing for one simple reason: nobody sells it in units that make sense. You do not buy "an article" or "a campaign". You buy tokens. And unless you understand what a token is, the numbers on a pricing page are meaningless.

So this is the piece I wish existed. No jargon, no benchmark marketing. Just what a token is, how the pricing works, what the five labs everyone is comparing actually charge in July 2026, and what that means for the kind of work marketers actually do.

What Is a Token, Actually?

A token is a chunk of text that the AI reads and writes in. It is not quite a word and not quite a letter. It sits somewhere in between. The rough rule that works well enough for planning: one token is about four characters of English, and 1,000 tokens is about 750 words.

Common short words ("the", "and", "is") are usually one token each. Longer or unusual words get split into pieces. "Marketing" might be one token; "Trendfingers" might be three.

Every time you use an AI model through its API, two things get counted:

• Input tokens - everything you send in. Your prompt, your instructions, any background documents, the conversation so far.
• Output tokens - everything the model writes back to you.

You pay for both, at different rates. That is the entire billing model. There is no monthly subscription hiding underneath the API; you pay per token, metered per request, and if you send nothing you owe nothing.

How Token Pricing Actually Works

Three quirks explain almost every surprise on an AI bill.

Output costs more than input

Across nearly every model, generating text is four to six times more expensive than reading it. GPT-5.6 Sol charges $5 to read a million tokens and $30 to write a million. Claude Opus 4.8 is $5 in and $25 out. This matters more than it sounds. A job that reads a long document and writes a short summary is cheap. A job that takes a short brief and writes 2,000 words of copy is, relatively, expensive. Same token count, very different bill.

Everything is priced per million tokens

A million tokens sounds enormous, and for a single task it is. It is roughly 750,000 words, or about ten full-length novels, or somewhere around 750 blog posts. That scale is exactly why individual AI tasks feel almost free, and why costs still creep up on people who run thousands of them.

Long prompts can cost extra

Several models now charge a higher rate once a single request goes past a threshold, usually around 200,000 tokens. Gemini 3.1 Pro, for example, doubles its input rate above 200K tokens. Most marketing tasks never get anywhere near that, so you can usually ignore it, but it is worth knowing before you feed an entire website into a single prompt.

The July 2026 Price Table: Flagship Models

These are the top-tier models from each of the five labs, priced per million tokens. USD is the native billing currency; GBP equivalents are at roughly $1 = £0.75.

The headline story is the spread. On output tokens, GPT-5.6 Sol is roughly 34 times more expensive than DeepSeek V4 Pro. That is not a rounding difference. It is a fundamentally different cost structure, and it is the single biggest thing to understand about the 2026 market.

The two American frontier models (GPT-5.6 Sol and Claude Opus 4.8) have converged on almost identical pricing, with Claude slightly cheaper on output. Google sits noticeably below both. The Chinese labs are playing a different game entirely, and as of this week that game changed again.

The Kimi K3 Story: Opus Performance at Sonnet Prices

If you only read one section of this article, make it this one. Moonshot AI released Kimi K3 on 16 July 2026, and it is the most significant thing to happen to AI pricing this year.

The specification is genuinely striking: 2.8 trillion parameters in a Mixture of Experts design, making it the largest open-weight model ever released. A 1-million-token context window with no long-context surcharge. Native text, image and video input. Always-on reasoning. And Moonshot has committed to publishing the full weights by 27 July, meaning you will be able to self-host it.

The price is $3.00 input / $15.00 output per million tokens (£2.25 / £11.25), with cached input dropping to $0.30 (£0.23). That is the exact same sticker price as Claude Sonnet 5's standard rate, from a model competing a tier above.

Does It Really "Beat Everything"?

This is where I want to be careful, because the launch-day hype and the evidence do not quite line up. The independent evaluation from Artificial Analysis, which ranks models across agentic work, coding, scientific reasoning and long-context tasks, scores K3 at 57 on its Intelligence Index — fourth place overall. Here is the actual leaderboard:

So K3 does not beat everything. It sits behind Claude Fable 5 and GPT-5.6 Sol, and lands roughly level with Opus 4.8. Moonshot's own benchmarks show K3 beating Opus 4.8 and GPT-5.5 across most tests, which is true, but those runs used different agent harnesses for different models, so they are not a clean like-for-like comparison. It also topped the Frontend Code Arena, a human-preference ranking, with a 76% pairwise win rate, which is a real and impressive result.

The honest framing is this: K3 is not the smartest model available, but it may be the best value at the top of the market. Artificial Analysis put its cost per task at $0.94, against $1.04 for GPT-5.6 Sol and $1.80 for Opus 4.8. Near-frontier capability at roughly half the cost of Opus is the story, not "it beats everything".

The Catches

Three things temper the enthusiasm:

• Hallucination got worse. K3's accuracy improved substantially over its predecessor, but its hallucination rate rose to around 51% from 39% on the AA-Omniscience test. For marketers producing factual content, that matters more than a benchmark score. Fact-check its output.
• Reasoning is always on, at maximum. At launch, K3 only accepts the maximum reasoning effort setting. Since reasoning tokens bill as output at $15 per million, a task that thinks hard and retries a few times can cost far more than the sticker rate suggests. This is the agent problem again, sharpened.
• It is much pricier than its own predecessor. Kimi K2.6 was $0.95 / $4.00. K3 is roughly three to four times that. Within the Chinese cohort, K3 is now the expensive option: DeepSeek V4 Pro is about a sixth of its input cost.

The era of reflexively cheap Chinese models is ending, and K3 is the clearest signal of that shift. For UK marketers, the same data-residency questions apply as with DeepSeek and Qwen. But the open-weight release on 27 July changes the calculation: if you can self-host it, the data never leaves your infrastructure and the per-token cost drops to whatever your hardware costs to run.

The Value Tier: What You Actually Run

Here is the part most "which AI is cheapest" articles skip. You rarely need a flagship model for marketing work. Drafting copy, rewriting, summarising, classifying, tagging, answering support questions - these are handled comfortably by the mid-tier models at a fraction of flagship cost.

A note on Claude Sonnet 5, because the timing matters if you are reading this in July: that $2 / $10 is introductory pricing that runs through 31 August 2026, after which it rises to the standard $3 / $15. So if you are budgeting a Claude-based workflow for the autumn, use the higher number.

For most content and automation jobs, this table is the one to plan around. Gemini 3.5 Flash and Claude Sonnet 5 are the sensible Western defaults. If pure cost is the priority and you are comfortable with a Chinese provider, DeepSeek V4 Flash is close to an order of magnitude cheaper than anything from the US labs.

What Does 1,000,000 Tokens Actually Buy You?

Abstract numbers are hard to feel, so let us make it concrete. One million output tokens is around 750,000 words. In marketing terms, that is roughly:

• 750 blog posts of 1,000 words each, or
• About 5,000 product descriptions, or
• Tens of thousands of ad variations, meta descriptions or social captions.

So when Gemini 3.5 Flash charges $9 to produce a million output tokens, it is charging around $9 to write the equivalent of 750 blog posts. That is about $0.012 (a bit over a penny) per post in raw model cost. Your flat white this morning cost more than a week of AI drafts.

This is the thing marketers consistently get wrong in both directions. They either assume AI content is expensive (it is not, per unit) or they assume it is free (it is not, at volume, and definitely not once agents are involved).

What a Real Marketing Task Costs

Let us price a realistic job: generating one 1,000-word blog draft from a brief. Assume about 2,000 input tokens (your brief, tone notes and a couple of examples) and about 1,500 output tokens (the draft itself).

Even the most expensive model on the list produces that draft for under $0.06 (under six pence). The cheapest does it for about $0.0007 (two-thousandths of a penny). So where does the money actually go? Three places.

• Volume: $0.05 (about four pence) a draft is nothing until you are generating fifty thousand a month.
• Long context: feeding whole codebases, transcripts or document libraries into a prompt multiplies input tokens fast.
• Agents: this is the big one. An AI agent that loops, calls tools, re-reads its own context and retries can quietly push 400,000 to 2,000,000 tokens through the API to complete a single task.

That is where a "cheap" model becomes a real invoice, and why the per-token rate matters far more for automation than for the odd blog draft.

The China Question: Why Are DeepSeek and Qwen So Cheap?

It is a fair thing to be suspicious about. When something is 30 to 100 times cheaper than the competition, marketers reasonably ask what the catch is. The short answer is architecture and strategy, not magic.

DeepSeek, Qwen and Kimi all lean heavily on a design called Mixture of Experts, where only a small fraction of the model's parameters actually fire for any given token. Qwen's newest generation activates around 4% of its parameters per token. That keeps the cost of running the model genuinely low, and those savings get passed through. DeepSeek cut its V4 Pro pricing by 75% in mid-2026 and now sits as the cheapest serious API on the market.

Worth noting, though, that "Chinese means cheap" is no longer a reliable rule. Kimi K3 arrived this month at $3 / $15, three to four times its own predecessor and firmly in Western mid-tier territory. As the Chinese labs close the capability gap, they are starting to price like it. Treat each model on its own numbers rather than by nationality.

The real considerations are not about quality, which is now competitive on most everyday tasks. They are practical:

• Data residency. Qwen's cheapest rates are on its Singapore endpoint; the mainland China endpoint is cheaper still but stores data in China. For UK businesses handling client or customer data, that is a GDPR conversation worth having before you route anything sensitive through it.
• Ecosystem and tooling. The Western labs have more mature libraries, integrations and support. The Chinese models are catching up fast but are less battle-tested in production stacks.
• They are open-weight. DeepSeek and Qwen publish their model weights, and Moonshot has committed to releasing K3's by 27 July, so you can self-host and pay nothing per token beyond your own hardware. Note that K3 at 2.8 trillion parameters needs at least 64 accelerators to run, so this is a data-centre proposition, not a Mac Mini one. The smaller open models are a different matter, and I covered that route in detail in the best offline LLM guide.

For non-sensitive, high-volume work (bulk product copy, classification, internal drafting), the Chinese models are hard to argue with on price. For anything touching regulated client data, the decision is about more than the sticker rate.

The Catches Nobody Puts on the Pricing Page

A few things that quietly change your real cost:

Caching is the biggest lever you are probably ignoring. Every major provider now discounts repeated input. If your prompts share a stable prefix (a fixed system prompt, a standard brief template, a reference document), the repeated part bills at roughly 10% of the normal input rate, sometimes less. DeepSeek's cache-hit rate drops input by around 98%. For any workflow that reuses the same instructions across thousands of calls, caching is the difference between a sensible bill and a silly one.

Batch processing halves everything. If you do not need answers in real time (overnight report generation, bulk content jobs, data enrichment), the batch API on every major provider takes 50% off both input and output for jobs that complete within 24 hours.

Tokenizers differ, so the same text is not the same token count. Newer models from Anthropic and others use updated tokenizers that can turn the same English text into meaningfully more tokens. Two models at an identical sticker rate can produce different real costs on the same job. Sticker price is a starting point, not the final word.

What This Means for Marketers

If you take three things from this:

First, stop shopping for the flagship. The frontier models are for the hardest reasoning and agent work. For content, rewriting, summarising and classification, the value tier (Sonnet 5, Gemini 3.5 Flash, GPT-5.6 Luna, or the Chinese models) does the job at a small fraction of the cost.

Second, the per-token rate barely matters for occasional use and matters enormously for automation. If you are drafting the odd blog post, pick the model you like and forget the price. If you are building an agent or a high-volume pipeline, the difference between $0.28 and $30 per million output tokens is the difference between a viable product and a runaway invoice.

Third, turn on caching and batch before you switch models to save money. Most teams chase a cheaper model when they could halve their bill on the model they already trust, just by structuring prompts for caching and moving non-urgent jobs to batch.

Frequently Asked Questions

What is the cheapest AI model in July 2026?

Among serious, current models, DeepSeek V4 Flash at $0.14 input / $0.28 output per million tokens is the cheapest. Qwen has even cheaper lightweight tiers. Among Western labs, GPT-5.6 Luna ($1 / $6), Claude Haiku 4.5 ($1 / $5) and Gemini's Flash-Lite range are the budget options.

Which is cheaper, Claude or ChatGPT?

At the flagship tier they are close. Claude Opus 4.8 is $5 / $25 and GPT-5.6 Sol is $5 / $30, so Claude is slightly cheaper on output. In the value tier, Claude Sonnet 5 at an introductory $2 / $10 currently undercuts GPT-5.6 Terra ($2.50 / $15), though Sonnet 5 rises to $3 / $15 on 1 September 2026.

Is Kimi K3 better than Claude and ChatGPT?

Not quite, though it is close. Independent testing from Artificial Analysis scores K3 at 57 on its Intelligence Index, fourth overall, behind Claude Fable 5 (60) and GPT-5.6 Sol (59), and roughly level with Claude Opus 4.8 (56). Where it wins is value: about $0.94 per task against $1.80 for Opus 4.8. Moonshot's own benchmarks show it beating Opus 4.8 and GPT-5.5, but those runs used different agent harnesses, so treat them cautiously.

How much does Kimi K3 cost?

$3.00 per million input tokens and $15.00 per million output (£2.25 / £11.25), with cached input at $0.30 (£0.23). That is flat across its full 1-million-token context with no long-prompt surcharge. Built-in web search bills separately per call.

Why does output cost more than input?

Generating text is more computationally demanding than reading it. Across nearly every model, output runs four to six times the input rate, which is why generation-heavy tasks (writing long copy) cost more than reading-heavy ones (summarising a document) at the same token count.

Is it safe to use DeepSeek or Qwen for business?

On price and quality, they are strong. The main considerations for a UK business are data residency and GDPR, since the cheapest Chinese endpoints may store data in China. For non-sensitive, high-volume work they are excellent value; for regulated client data, review where the data goes first, or self-host the open-weight versions.

How much does it cost to write one blog post with AI?

In raw model cost, between roughly $0.0007 (two-thousandths of a penny) on DeepSeek V4 Flash and about $0.055 (six pence) on GPT-5.6 Sol for a 1,000-word draft. The model cost is almost never the real expense; your time editing it is.