Optimise your tokens before the meter arrives

Do you remember when Uber was cheap?

That’s not me doing a hackneyed stand-up routine from the 1970s where we hark back to the distant past and say how things used to be better. I’m talking about 2012. I am old enough to remember how you could get across London for £5 in a cab you ordered ON YOUR PHONE. Good god, it felt like a brave new world. Would I have booked the cab if it had cost more than a standard taxi? Maybe as a novelty. Maybe if I was over a barrel, but it’s £5. Cheaper AND easier, you say? What a fantastic new world!

Uber was founded in 2009 and did not post a full-year operating profit until 2023. For most of those years, investors (mostly SoftBank’s Vision Fund, Alphabet and Saudi Arabia’s sovereign wealth fund) paid for your ride.

That was a strategy which has now become pretty commonplace in tech “disruptors”, where they can use investment capital to scale very fast (compared to more traditional businesses who must grow sustainably without huge cash injections).

Local “mom and pop” taxi firms had to cover their costs from fares and so they simply could not compete with rides that were sold below cost. Then, with competition thinned out and habits formed amongst taxi users, the prices went up. Service standards got worse too, and pretty rapidly. A Columbia Business School study of around 37,500 trips found Uber’s price per mile rose 53% between early 2023 and early 2026, while wait times got 19% longer.

The big companies in AI are largely the same set of guys running largely the same playbook.

Right now you are in the honeymoon, adoption period where you are being aggressively courted, largely subsidised by investor money. Since OpenAI launched their paid subscription for ChatGPT in February 2023, it has been sold in a way similar to gym memberships. You pay a flat price and you can use almost as much as you want (within usage limits – I’m gonna need that squat rack soon buddy). There’s also a generous free tier aimed at bringing new users on board in a low friction way. This is a fairly common customer acquisition strategy and you can find it in everything from free samples in the aisle at a supermarket to Canva. It’s a cost paid by investors who expect their money back and then some in new business.

If you are just someone trying to get some work done, the noise can feel utterly relentless. The tech sector is speaking with a near-singular voice when it says that you should “use more AI” and “use the newest model”. You will be told that “tokens are basically free now” and that you should build new empires on this recently-discovered ground. Much of that noise comes from companies who really need you to spend. They have spent eye-watering sums building this stuff and they’re desperate for it to pay.

So this piece is for you, the underdog. We’re going to attempt to cut through the hype and the marketing and help you get the most from whatever you spend, whether that’s $20 a month or a hefty API bill.

It’s good business to spend your resources wisely, whatever level you’re operating at, and there are real simple ways you can do it.

Here’s what’s worked for me, building experiment code, internal and external tools with AI.

How careful do I need to be?

Careful enough to get value for money. For everything you spend, you’re getting some value and some waste. Your job is to minimise the waste and to maximise the value.

Flat-rate plans. This is how most people are using LLMs. Either free or with a subscription and with those you are not billed per token. Instead, they attempt to ration usage around either time or volume or both. ChatGPT Plus gives a set number of messages on flagship models every few hours and then it drops you to a smaller model until that window resets. Claude’s “Max” plans have a usage allowance that resets every five hours and also a weekly cap. Both share a common business model, which is that your usage is something of a black box, and you’re out of control on how many tokens are used until after they’ve already been spent.

Plus, the rations can move over time in response to the model owner’s business needs. In September 2026, OpenAI re-opened its previously closed $200 Pro plan (to new subscribers with less usage than before), cut it for existing subscribers from the end of October and then launched a $500 tier. They are offering less service for higher prices. Just like Uber did.

Pay-per-use API. This is what the AI companies would like your business to use. This makes it possible for prices to change and costs to increase far more readily than they can when you pay a flat rate. With API pricing, you pay for each token and when you burn more tokens, you pay more.

At today’s prices, Claude Opus 5.5 charges $4 per million input tokens and $20 per million output tokens. Sonnet 5.5 is $2 and $10. Haiku 5.5 is $0.10 and $0.50.

Even if you don’t change providers (and you definitely should) and you ONLY use Anthropic models, that’s a 40x difference in cost between the cheapest and most expensive current models.

So some simple rules you can follow:

• Are you chatting and building on a subscription? Token optimisation is less important here unless you’re hitting your limits but don’t expect this pricing model to last forever. Maybe don’t expect it to last a year! When Anthropic/OpenAI/xAI/Gemini investors start asking for their money back, then expect this model to cut back pretty rapidly, like Uber did.
• Running anything automated, agentic or customer-facing using the API? Optimise this stuff properly. Any sloppy building or inefficient loop is literally costing you money.
• Running a small team and paying for the seats? Costs multiply pretty fast and we’re not the prompt police, so it’s important to agree some ground rules nice and early-on and make sure your team understands what is expensive, whilst also giving the team room to innovate.

People tend to burn more where they’re not responsible personally for the cost, or usage isn’t metered or visible. One advantage of using coding tools rather than chat is you can see context usage and token burn.

A lot of this problem is overkill when selecting models – using less advanced models when breaking down a chain of steps is just one way of reducing costs. I use multiple models (opus, sonnet, haiku) for a daily ingest of synthetic white papers and it works – because the model is picked for the task.

Being aware of when your queries are burning piles of tokens is usually a good indicator of needing to find a better way to do that thing. Claude will often go down rabbit holes, burning loads of tokens for nothing.

– Craig Sullivan, Experimentation Consultant at Optimise or Die

Tokens vs credits

A token is a unit that the model actually reads and writes. It varies, but a nice shorthand is “roughly three-quarters of a word”.

A credit is not analogous to a token, and is a unit that has been invented by a vendor, which can obfuscate what you’re paying for. I am being cynical here, but with good reason. Credits abstract away what you’re paying. AI development environment Cursor is a good example of this in practice. In June 2025, Cursor swapped a fixed number of “fast requests” for a monthly “credit pool” that was roughly equal to your subscription price. Third-party models like Claude and GPT were then charged at API prices, so a big context window or a premium model was ripping through credits much faster than users expected.

Plus, the incentives make it so you can be nudged into decisions you otherwise wouldn’t have wanted to take. SpaceX purchased Cursor on 14th August 2026 and at the time that happened said the team would “make Grok the world’s most useful AI”. Cursor’s larger credit pool now covers Grok and Cursor’s own Composer model with Claude, GPT and Gemini sitting in a separate, smaller pool billed at API prices. Cursor no longer publishes how big that pool is.

Users on Cursor’s forum have also reported Grok switching itself back on after they’d disabled it, and new chats opening on Grok. They also said Grok became the fallback when they hit their limits. Cursor staff originally called it a bug, then later admitted: “we do run some model nudges to encourage you to try out Grok 4.6.”

How to handle credit-based platforms:

1. Try to find the exchange rate. Work out what a credit buys on the models you want to use. If the vendor will not tell you, well, that’s a red flag. You can try and measure this from the outside by running a known task and watching your dashboard but I’d actually advise that you should consider changing vendors if they won’t tell you what your money buys you.
2. Check what counts as context. Attached files, chat history and “helpful” auto-included code all count. Some harnesses do awful context-eating things too, like dumping images into the conversation as raw base64 text which can turn a single screenshot into hundreds of thousands of tokens. Huge bills usually come from out-of-control context windows rather than individual prompts.
3. Use a cheap default. Most credit platforms have an in-house or “auto” model that costs less. Use it for normal stuff and save premium models for challenging or high-value work.
4. Set a spend limit. Do this before you’re burned.

Token optimisation across different disciplines

The generalist advice is the same everywhere. Spend tokens where you need judgement and not where it is not. That might look different by role.

Marketing. My personal opinion here is that LLMs are not particularly strong for marketing and that most human-facing copy should be written and conceived by marketers. If you are going to use it to help you market, I would consider using it to go and source information that you can consume to help you find the hook, rather than specifically generating your marketing copy.

Development. Context is what costs so much money here. Agents reading an entire repo to change something are remarkably expensive and also the context bleeding into the outcome leads to rubbish websites and bloat. Point agents at files that matter, keep a short and well-written project instructions file so you don’t re-explain your stack and decisions every single session and learn to start fresh conversations for features and changes regularly instead of having a beefy mega thread in which you discuss everything.

CRO. We have an advantage here, or we should, in that we already think in hypotheses and evidence. Use a cheap model for sorting, tagging, summarising research (session recording notes, survey verbatims, support tickets). Build systems to be deterministic where possible. Design your systems to give some low-brain dogwork to AI and then use the time that creates for humans to do things they’re uniquely good at. Thinking through complex problems. Recognising underlying themes that are important. Understanding how psychology intersects with different messages and systems. Don’t try to replace those parts with a machine. The outcomes are faster and worse.

Tokenmaxxing is a perverse incentive

In spring 2026, some of the big tech companies decided that a good way to measure AI adoption would be to see who could set fire to the most money.

That practice was dubbed “tokenmaxxing”. It was rewarding people for consuming as many tokens as possible and it was remarkably stupid, even for Silicon Valley.

Jensen Huang (NVIDIA CEO who makes money from people using as much AI as possible) said he would be “deeply alarmed” if a $500,000 engineer did not spend at least $250,000 a year on tokens. A Sequoia (Venture Capital firm heavily invested in tech) partner told the Wall Street Journal, “we all should be tokenmaxxing.”

Meta (a company that sees its future very much as an AI product after the commercial failure of the Metaverse and Meta Glasses) ran an internal leaderboard with ranks like “Token Legend”.

A dear friend of mine working at a tech company even confided in me that, in a mid-year performance review, they had it intimated to them that, were they not to consume more tokens, their job might be at risk.

I would suggest that if you are in charge of setting targets at your business around adoption, then you shouldn’t follow this tokenmaxxing advice, as the call is coming from inside the house.

Even companies on the selling side couldn’t stomach the bill when it was time to pay.

Amazon, which sells AI compute through AWS, scrapped its internal leaderboard in May after staff were reported to be padding their token counts with pointless tasks. Microsoft, which sells Copilot, started cancelling internal Claude Code licences, with The Verge reporting cost as part of the reason. Uber reportedly blew its entire 2026 AI budget by the end of April.

Anyone who has run a CRO programme has seen this movie before. It is Goodhart’s law writ large. When a measure becomes a target, it stops being a good measure. Celebrating token spend is like celebrating “experiments launched” when all your experiments were minor copy changes entirely inconsequential to your journeys or your users. It is a vanity metric that only really serves the vendor.

The advice we would have given before 2023 hasn’t changed really. Measure outcomes. Shipped features, validated hypotheses, hours saved, revenue. If your token count is decreasing while those go up, you have won, and if the biggest companies in the world cannot justify burning tokens just for the heck of it, neither should you.

The real cost of AI, and who is really paying

Follow the money and token pricing makes a lot more sense.

Combined 2026 AI spending by the big hyperscalers has topped $700 billion, according to Reuters.

Microsoft, one of the few actually selling AI at scale, reports an AI revenue run rate of $37 billion. That is a very big hole and your usage is one of the ways they plan to fill it.

The model labs are in the same boat. Anthropic told investors it could be cash flow positive by 2027. Its leaked IPO filing shows a $42 billion net loss for 2025 on $4.6 billion of revenue, and $518 billion committed to compute. Recent “profitable” quarters rely on accounting measures critics have called a swindle. Making that money back needs a lot of customers paying a lot more.

Flat-rate plans are a loss leader designed to get as many users onboard before prices go up. In May 2026, Anthropic announced it would move programmatic usage off of subscription limits and onto credits and then full API billing. After developer backlash, it paused the change on the day it was due. In October 2026, it added “monthly API credits” to Max and Team plans. At the moment, it’s an “extra”, seemingly to encourage more API usage, but time will tell if this is just a softer slide towards API pricing or not.

Be wary of any “tokens keep getting cheaper” line coming from the industry too. Older capability does get cheap relatively fast but the best models keep launching at same or higher prices and they burn far more tokens per task while they “think”.

Price per token is probably the wrong number to watch. Watch “cost per finished job” instead. 

Gartner also reported that costs for agentic workflows are set to rise more than fivefold in two years (because agents burn 5 to 30 times more tokens than a chatbot for the same task).

Gartner’s advice is to move from flat fees to usage-based pricing. The meter is coming, folks, whether you like it or not, so don’t assume the subscription will work forever.

So, as a user, plan for three things:

1. Subscriptions will tighten. Expect lower limits, higher tiers or both.
2. Automation will be metered. Anything that runs without a human in the loop is the obvious place to start charging more.
3. The real cost is not just money. Data centres need energy and water, and the models were trained on a lot of people’s work without permission. Ethics and restitution for that is ongoing, but waste here has a human and environmental cost as well.

Use as much as you need, but as little as you can. The ‘cost’ associated is more than monetary and that will also be an inevitable consequence once we’re addicted.

– Paul Randall, Associate Director of Strategy at Speero

Concrete tips from my workflows

Practicing what I preach. Not everything leaves the building for me. I run Ollama on a home server next to n8n and a Qdrant vector database. That’s dork speak for saying that boring, high-volume jobs, like tagging, summarising and sorting, go to a local model and cost me electricity. In the near future, I plan to make that electricity entirely solar. I also maintain independence in my tooling as much as possible which lets me move to cheaper/better/more sustainable models as allows.

1. Right model, right job. The easy way to do this without intimately knowing every model is to just choose the midweight one, drop to a small one for sorting, tagging, extraction and formatting and step up to the big one for really challenging work. The hardcore way to do this is to build workflows specifically with the smallest possible model in mind.
2. Write it down once. Good practice for everyone this. Write a short instructions file (stack, conventions, tone, what not to do) that is loaded at the start of every chat. Keep it tight, though, as it’s sent along with every prompt.
3. Start new conversations. Long threads inject their whole history with every reply, so context-bloat grows with every prompt. When the task changes, start fresh. If you need to, have the previous chat write you a short handover that you then paste into the next chat. Edit it to remove the junk before you paste it into a new chat.
4. Give it less to read. Point at the files that matter or paste the relevant rows. It may not need the whole spreadsheet in the context window.
5. You can still edit by hand. If you need to change the padding of an element in a webpage, or edit a block of copy, learn to find that copy and rewrite it by hand. It’s usually faster for minor edits, plus it’s utterly wasteful to prompt towards it. I guarantee you, you’ll write better copy than Claude this way.
6. Turn repeat work into skills or templates. If you have asked for the same thing, in the same way, three times, then you have something worth thinking about as a process. You can save this as a skill or a template so it feels less random each time.
7. Use caching and batching on the API. One for the API heads. These are settings you should think about when writing your work. Think about this a bit like how you can get cheaper electricity overnight when things are in less demand.
8. Ask for less output. Output tokens are priced higher. Tell Claude to “say less, bro”.
9. Let deterministic code do deterministic work. Good advice for living as well as for coding. Deterministic code for repeatable jobs. If regex, SQL queries or a script can do it the same way every time, then you should not be paying a language model to guess.
10. Watch the dashboard weekly. Know what you’re spending and keep an eye on it.

Go small, go open

The cheapest token will almost always be the one you don’t buy from a hyperscaler.

Open-weight models are models whose weights are published for anyone to download and run. On a laptop, on a server or with a cheap hosting provider. It sounds complex, but it’s remarkably simple once set up. There is no meter, no five-hour window, no surprise price rise, and your data never leaves the building.

If Uber taught us anything, it is not to build your business on someone else’s subsidy.

These models are not toys anymore. Alibaba’s Qwen3.6 and Qwen3.8 (27B), Google’s Gemma 4 and OpenAI’s gpt-oss 20B are all free to download, and gpt-oss runs in about 16GB of memory. For lighter machines, Gemma 4 E2B and Microsoft’s Phi-4-mini run happily on modest hardware. This area is moving fast, so if you’re reading this a few months after publication, it may have moved on again. Keep your ear to the ground.

Tools like Ollama and LM Studio make running them about as hard as installing an app.

Better still, most jobs do not need a big model at all. A March 2026 study of 16 models across five tasks found small models of 0.5 to 3 billion parameters gave the best balance of accuracy, speed, memory and latency on every task tested. Classifying feedback, tagging survey responses, extracting fields, routing support tickets: a small, targeted model will do that all day for the cost of electricity.

A properly cost-effective system usually looks like this:

1. Plain code for anything with a right answer.
2. A small local or open-weight model for high-volume, repetitive work.
3. A frontier model, called sparingly, for the hard reasoning that actually needs it.

Note: Read the licence, as these models have terms you need to read and understand before you use them commercially. Also, open weight =/= open source. The biggest open models will still need some hefty hardware.

But for the little guy, owning your tools will always be better than renting them.

Housing small models on your machine for things like embeddings, straightforward inference, and the like is great. If you’re easily spooked, a simple early fix is to make sure you’re using skills/connectors/apps the model can work with.

So many Claude tokens spent converting Excel docs to Libre Office, watching it create a script to add a Substitute formula to a column, then converting it back again. Avoid all that interim spend and just add the connector.

– Michael Bates, Head of Organic Media @ Herd

You do not have to marry one vendor either. OpenRouter gives you one API key and one bill for models from most of the big labs and a long tail of alternatives, and it passes through the providers’ prices with no markup on usage (it takes a small fee when you buy credits). Switching models becomes a one-line change, and if one provider falls over, it falls back to another. When I built some tools for Convert, we built them on this technology for this reason.

OpenRouter also has an Auto Router which uses a lightweight classifier to work out what kind of task you are asking for (debugging, maths, writing and so on), then it picks a model within the cost band you have set. You can restrict it to only models you trust, you can cap price per request and there is no extra fee on top of the model you end up using.

None of that works if your tools won’t let you choose and your vendor locks you in. The harness, the app or agent you actually work in is what ultimately decides which models you can reach. Use Claude Code and you are choosing only from Anthropic’s menu. Codex keeps you locked in with OpenAI. Cursor now nudges you towards Grok and will lose OpenAI’s models in November. When SpaceX bought Cursor, I consciously stepped away from that tool because I don’t want to give Elon Musk money.

Open-source harnesses let you break out of that vendor lock-in. OpenCode supports more than 75 providers, plus local models through Ollama and LM Studio. Cline, Aider, Kilo Code and Goose all let you bring your own key and plug in OpenRouter or run models on your own machine.

My current harness of choice is Zed. I think it’s pretty cool.

Spend like it matters

This whole article is about why you should be sensible with what you have, the same way you are with money, time and energy. The vendors really want you to be locked in and to maximise usage, whereas you really want outcomes.

Spend your time, energy and money investing in outcomes and you will be fine, whatever happens to the pricing.

Or, to put it in strictly CRO terms, you can stop counting the tests you launched and start reflecting on what you have learned.