Three AI Models Got Cheaper. Nobody’s Bill Went Down.

The week everything got cheaper

For a few days at the end of September, AI suddenly looked a lot cheaper.

Anthropic shipped Claude Sonnet 5.5 at the same price as Sonnet 5, making it half the per-token cost of Opus 5.5 while beating Opus on a major coding benchmark.  OpenAI followed with GPT-6.1 Sol, promising near-Astra intelligence at a fraction of the price. Google then put Gemini 4 Argon on the board at the same headline rate as both of them.

The coverage understandably reads like a clearance sale.

For context, AI models are generally charged by the amount of text they process and generate: the first figure below is the cost per million input tokens, and the second is the cost per million output tokens.

Here's the week, in the only form that matters when you're the one paying.

Three of those four lines made frontier AI look cheaper. One went sharply in the other direction: a new $500-a-month tier, 150% above the previous top consumer plan. Announced in the same week, that's the first clue that something more interesting is going on than a price war.

I build and run software on these models. Fourteen live products, a one-man shop in Watford, every invoice landing on my own desk. So I did what anyone in that position does when four major pricing announcements arrive in a few days - I went and looked at what they'd actually done to my bills.

Nothing.

They had done nothing to my bills.

The number that spoils the story

Buried in the Sonnet 5.5 coverage is a figure that undoes the headline. Running at maximum effort, Sonnet 5.5 generated roughly 193,000 tokens on a single task - the most Artificial Analysis has ever measured, and about 60% more than Opus 5.5 uses for the same work.

Do the arithmetic. Sonnet 5.5 costs half as much per token as Opus 5.5, but in that test it consumed 60% more tokens. Half the price × 1.6 times the tokens = 80% of the cost. So in that configuration, the model advertised at half the token price was only about 20% cheaper for the task - and that's before you count retries, failed tool calls and the context you may have to re-send on every turn.

It gets better, or worse, depending on your mood. Anthropic's own framing of the same release is that Sonnet 5.5 uses fewer tokens per task than Sonnet 5 and therefore cuts cost per task without an API price change. Help Net Security reported exactly that, alongside a FrontierCode result at roughly one-fifteenth of Sonnet 5's cost per task.

Both claims can be true. They're measured at different effort settings, on different benchmarks and against different comparators. One is "cheaper than the model above it", the other is "cheaper than the model it replaces", and neither tells you "cheaper than what you are running today".

That's the whole problem in one release.

The sticker price is a fact. The cost of getting your actual job done is a measurement, and nobody can take it for you.

I'm not accusing anyone of fiddling the numbers. Anthropic published the token counts. Artificial Analysis published theirs. The information is all there, in public, on the same day. It just doesn't survive compression into a headline - and the headline is what most people will act on.

Why the rate card stopped predicting the bill

There was a period, roughly 2023 to 2024, when per-token pricing was a decent proxy for cost. You sent a prompt, got a completion, and the relationship between the two was fairly predictable, so you could estimate a month's spend on the back of an envelope.

That period is over, for four compounding reasons:

1.      Reasoning effort is now a dial.  The same model doing the same task can consume dramatically different numbers of tokens depending on how hard you let it think. Sonnet 5.5's 193,000-token task was running at maximum effort. The rate card tells you what each token costs; it does not tell you how many the model will decide it needs.

2.      Agents keep carrying context forward. A simple model call might send a prompt and receive an answer. An agent can make repeated model and tool calls while carrying an accumulating history. So a twelve-step task doesn't necessarily cost twelve times the first step - later steps may be processing much of what happened in the earlier ones as well. Caching helps, and Sol's cached input at $0.10 against $2 standard is a 95% discount that genuinely matters, but only on the portion that actually hits the cache.

3.      Failures still cost money. A tool call that returns malformed JSON, a timeout, a refusal, a loop that needs killing - you've still paid for the AI work without getting the result you wanted. On a mature product, this should be a relatively small percentage. During development, or on something you shipped last week, it can become a surprisingly large part of the bill.

4.      The cheapest model is not always the cheapest route. A weaker model that needs three attempts is more expensive than a stronger one that works first time.  It also costs you latency, which, on something like a voice product, is often what the customer actually notices.

Put those together, and the most predictive number is cost per completed job, measured on your own traffic.

Everything upstream of that is a component price, and component prices don't give you a useful bill of materials when you don't yet know how many of those components each job will consume.

Launch pricing is marketing

Gemini 4 Argon arrived at $2 input and $10 output per million tokens, rising to $4 and $20 at general availability. Google is telling you, in advance and in writing, that today's price is not the price you should build your long-term economics around.

I've a lot of respect for that, actually. It's more honest than the alternative: a quiet repricing eighteen months in, once everybody's product depends on you. But it should settle an argument that comes up on almost every build: no, you cannot price your SaaS off this month's rate card.

If your margin only works at preview pricing, you don't have a margin. You have a promotion, and someone else controls when it ends.

This isn't a hypothetical risk for a small software company. It's an easy way for an AI-native product to get itself into trouble. You model your unit economics on a launch price, sign annual contracts at a number that works against that price, then the provider normalises its rates. Suddenly every customer you signed is underwater, and you cannot raise prices mid-term.

The provider is fine. You are not.

What I do instead is dull, but it has saved me twice. Every product gets costed at the worst plausible token price, not the current one. In practice, I assume double. Argon has just given us a useful example of what that kind of normalisation can look like.

• If the business case only closes at today's price, the product does not get built. If it still closes at double, anything better than that is margin I didn't promise anybody.

It's the same discipline as pricing a fixed-price consultancy job on your realistic worst-case estimate rather than your best. Nobody has ever gone broke from a job that came in under.

Tokens got cheaper. Seats got dearer.

The part of DevDay nobody built a headline around:  OpenAI launched a $500-per-month Pro tier and, in the same breath, halved the usage allowance on the existing $200 plan. Existing subscribers paid the same and got less. If you wanted the higher level of access you had on Monday, you were suddenly looking at $500 rather than $200.

So, in the same keynote:

• Astra-level capability was being pitched through the API at a fifth of the token price, while
• the top human seat became two and a half times more expensive.

Those are not contradictory moves. They are two sides of the same market.

The underlying AI is starting to look more like a commodity. Three vendors landed around the same $2/$10 headline rate within 48 hours. That's what commodity markets tend towards: convergence rather than differentiation on the underlying unit price.

Seats are different. A seat is sold to a person with a workflow, a history, a habit and, quite often, a company card. Once AI becomes embedded in the way somebody works, they're buying considerably more than tokens.

The cheapest way to bring advanced AI into a business may increasingly be to build it into the product that actually does the job, rather than buying a premium AI licence for every person who might need access to it.

A clinic with six staff doesn't need six $500 seats so somebody can answer the phone properly. It needs one voice agent doing that particular job, using the AI only when it's needed and potentially costing just a few pounds a day.

That gap - between paying for general-purpose AI access per person and using AI to perform a specific job - is a substantial commercial opportunity for small software companies.

And this week, that gap got wider.

Most commentary read the same announcements as evidence that the big labs are squeezing everyone else out. I think they may be creating more room for us.

What everybody else makes of it

I'm not the only person who thinks the rate card has stopped telling us very much on its own. The analyst community has actually been clearer about this than much of the tech press.

In CIO's coverage of the price-performance race, three analysts approach the same change from slightly different directions:

Charlie Dai of Forrester sees it as structural rather than promotional: "Frontier AI is entering a prolonged price-performance race driven primarily by inference efficiency gains."

• In other words, this is bigger than a temporary sale. The underlying technology is becoming more efficient, and the major providers are competing aggressively on how much capability they can deliver for the money.

Sanchit Vir Gogia of Greyhound Research gets even closer to my argument: "The only defensible measure is total cost per accepted, policy-compliant outcome."

• That sentence is doing a lot of work.
• Accepted excludes the outputs you threw away. Policy-compliant excludes the ones that created another problem. What remains is the work that was actually completed successfully - which is a much more useful thing to measure than how cheaply you bought a million tokens.

Anushree Verma of Gartner takes the argument further, warning of a "race to the bottom" in AI pricing as providers fight for market share.

There is a bear case too, and it's not daft. Writing in July, VKTR argued that price compression is arriving at exactly the wrong moment for AI labs heading towards public markets, quoting investor Gavin Baker: "A world where there are only 2-3 dominant frontier labs with 90% inference margins is net negative for every other layer."

If access to powerful AI keeps getting cheaper, that's excellent news for people building products on top of it. It's potentially less comfortable for the companies spending billions building the underlying models, particularly if their margins are being squeezed at the same time.

Which brings me back to that $500 seat.

Where I part company with some of the commentary is the conclusion. The usual fear is that cheaper AI means the big labs will eventually eat the application layer and leave little room for smaller software companies.

I think it could work the other way around.

If the underlying AI keeps getting cheaper while general-purpose access to it becomes more expensive per person, there's increasing value in building that intelligence directly into a product that does a specific job.

The big AI companies don't necessarily have to be our competitors.

Increasingly, they may be our suppliers.

How I actually cost a product

None of this is theoretical for me. GeekyBee runs fourteen-odd live products using metered AI, and what keeps them viable is not picking the cheapest model.

It's routing.

Take our comment moderation platform. Most of what arrives on a public social account is trivially fine or trivially not - a heart emoji, a tour date question, obvious spam. Sending every one of those to one of the most powerful AI models available would be like hiring a barrister to read your post.

• So the first pass goes to a small, fast model that can handle the bulk of traffic for almost nothing. What it's genuinely unsure about escalates to a more capable hosted model. The small minority that is actually ambiguous - sarcasm, coded abuse, a complaint that might become a legal problem - escalates again. And anything the system considers genuinely consequential can ultimately be put in front of a human.
• The effect on the bill isn't marginal; it's the difference between a viable product and a hobby.
• A flat rate of £39.99 a month for a private account only works because the average comment costs a fraction of a penny to classify and only the difficult ones consume meaningful amounts of AI.

That architecture is also why this week's price announcements did almost nothing for me.

If the expensive model is only handling 3% of your volume, even cutting its price in half only moves your overall AI cost by around 1.5%. The headline says "half price". Your invoice says "barely noticed".

The second habit is measuring the right unit.

• Not tokens per month.
• Cost per completed job.

And that means something different for every product:  cost per moderated comment, per answered call, per site crawled, per property scored.

It doesn't take long to build that measurement into a product, and once you do, it starts telling you things a rate card never will - which product is quietly subsidising another, whether your routing is actually saving money, and whether a particular customer on a flat-rate plan is costing more to serve than they are paying you.

One of mine was.

I would never have found it by reading pricing announcements.

What to do about it this week

If you run software on top of these models, or you're paying someone who does, five things are worth an afternoon.

1.      Work out your cost per completed job. One number per product, based on your own traffic and your own logs. Until you have it, every price announcement is a guess, not information.

2.      Check your reasoning settings. If you set reasoning effort to maximum during development and never revisited it, you may be paying for more reasoning than the task actually needs. Test lower settings against real work and find the lowest setting that still delivers reliably.

3.      Check whether you're making use of caching.  If the AI repeatedly needs the same instructions, background information, or conversation history, you may not need to pay full price to process it every time. Cached input at $0.10 compared with $2 standard is a 95% discount, so how you structure repeated information can make a significant difference to the bill.

4.      Cost everything at double today's price. Argon has just given us a useful example of what a doubling between introductory and full pricing can look like. If the business case needs preview pricing to work, it doesn't work.

5.      Route, do not automatically upgrade. The instinct when something underperforms is to reach for a more powerful model. Often the cheaper answer is to classify the work first and send only the difficult cases to the expensive model.

The broader point is simpler than any of that.

This week, three of the biggest AI companies made powerful AI look dramatically cheaper, while access at the human-seat level moved sharply in the other direction. The headlines treated them as separate announcements.

I think they're part of the same change.

• The underlying AI is becoming cheaper and increasingly interchangeable, while convenient, general-purpose access to that intelligence is becoming something companies can charge a premium for.
• If you're buying seats for everyone, that could become increasingly expensive. If you're building products, however, cheaper underlying AI gives you an opportunity to put increasingly capable intelligence directly into the job that needs doing.

That gap is where small software companies can compete. Either way, the number that matters isn't on anybody's pricing page.

It's in your own logs.

And almost nobody has gone to look.

Research checked and links verified on 5 October 2026.