Cut your AI token bill by up to 95% - for free

Friends,

your weekly AI briefing is here - designed to help you respond to AI, not react to the noise. No curveballs. No chaos. Just clarity.

4700 business leaders read this newsletter every week because someone they trust shared it with them. Together we can create the tomorrow we all want to live in, so please do share this with someone you trust your future with.

Time sensitivr: Sherpas AI just launched Future Founder - teenagers use real AI skills to explore career paths while building their own startup, not a simulation. It's what last week's Knowledge Was Power report highlighted gives teenagers a strong start to their career. Applications are open now at sherpas-ai.com. (Disclosure: I'm a co-founder of Sherpas AI. It is a social enterprise with a mission to create a generation of Ai fluent changemakers who can solve the worlds hardest problems.)

You can watch my podcast for AI for Equity with Jenny Garrett OBE and Lead-Sunshine Garrett to hear more about why I do this work. It was a wonderful conversation and great to sit down with a very cold drink in a cool place!

📰 This was the week that was...

This was the week the token bill came due - and the industry finally admitted the maths was never about price.

Uber burned through its entire 2026 AI budget by April, Microsoft revoked developer licences it had granted only months earlier, and one company reportedly ran up a $500 million bill after forgetting to set usage limits. Yet per-token prices keep falling - a study of 2.4 billion enterprise API calls found blended cost per million tokens down 67% year on year, from $18.40 to $6.07. Volume outgrew every budget model.

GitHub Copilot's move to token-based billing closed its first full metered month this week, and developers running autonomous coding sessions reported bills 10 to 50 times their old flat-rate plans - not from misuse, but because each agentic session makes dozens of model calls, resending the accumulated conversation every time.

Culture is catching up faster than tooling: Every reports that bragging rights have flipped from maxing out consumption to efficiency, with "revenue per million tokens" now circulating as a successor to revenue-per-employee. The FinOps Foundation finds 98% of organisations now actively manage AI spend, up from 31% two years ago.

None of that is a reason to slow down - it's a sign the discipline is arriving on schedule. More on the fix below.

Let's get into it.

🔥 Urgent Priorities

✅ No fires to fight this week

✅ If you run agentic tools (Copilot, Claude Code, Cursor or similar), check whether a hard spending cap is switched on - several vendors now bill by default with no ceiling unless you set one

✅ Time to ask a procurement question you've probably never asked your AI vendor - this week's Strategic Insight, below

This isn't a week for panic. It's a week for a five-minute check of your billing settings.

🎯 Strategic Insight

Who in your AI arrangement is actually incentivised to bring your bill down?

Tension: Most leaders read this month's shock bills as "AI got too expensive" - and quietly assume their frontier lab will help them use less of what it bills them for.

Optimistic insight: That assumption was never true - and it isn't a crisis, it's a design problem. A frontier lab makes money on tokens, so it has no structural reason to want you using fewer of them; that is simply their incentive. The fix isn't a discipline memo asking your team to be careful - it's contractually and operationally separating who owns capability (the lab) from who owns cost (you). None of the three lanes below is universally right - picking one, on purpose, is the whole game.

• Frontier Partner - go direct to a lab's frontier model where raw capability moves the P&L, and accept being at the cutting edge requires an acceptance of spiraling costs.

• Orchestrated - put a routing layer between your business and the models, so cheap models handle routine work and expensive ones are saved for what actually needs them.

• Sovereign - run models on infrastructure you control, trading capex spend for total visibility (the sovereign end of last week's local-AI spectrum).

What's shifting: The Orchestrated lane has gone from theoretical to routine in about a year - tools like LiteLLM and OpenRouter now sit between businesses and models as standard, and a free tool that strips 60-95% of wasted spend before it reaches the model (see Geek Out) means the case for doing this builds itself. Strategist Stuart Winter-Tear has separately argued that the industry spent two years blaming the model for disappointing results, and has now converged on the same conclusion: the answer was in the organisation, not the model. Choosing a lane is that kind of decision, not a technology purchase.

Why this matters now: Successful AI adoption increases usage - the bill grows fastest exactly when your rollout is working. Have this sorted before you discover you need it.

Takeaway / Action: Before you sign or renew any frontier AI deal, ask your vendor and your own team one question: "who in this arrangement is incentivised to reduce my token consumption?" If the answer is nobody, that's your case for the Orchestrated lane - not a bigger budget line.

🤓 Geek Out

The Free Tool That Shrinks Your AI Bill Without You Lifting a Finger

Most of what you pay for isn't the AI's answer - it's everything it has to read first: old messages, files, logs, pasted-in reference material. Headroom is a free tool that quietly shrinks all of that before it reaches the model, cutting 60-95% off your token bill for the same answers. Nothing is lost - the full originals stay saved, and the model can pull them back if it genuinely needs to. Independent tests show accuracy holds up, and on some measures it actually improves.

Why it matters: This is the Orchestrated lane made real - a way to cut your bill without touching accuracy. If you're paying per token for repetitive tool calls or long logs, this is what catches it before the invoice does.

👉 Action: Run this on one repeated AI workflow: ask it to break down what it's spending at each stage - which model, roughly how many tokens in and out, what each step is for, and whether that spend was actually necessary (flag anything it can't be sure of as an estimate). Find the step where the spend doesn't match the difficulty of the work, then change one thing - trim what you're feeding it, tighten your instructions, or switch that step to a cheaper model - and try again. It's Every's method for auditing agent spend; you still need a human to review it, because AI isn't good at spotting its own waste.

👉 Read about Headroom

A Shorter Way to Feed Data to AI - Same Information, 40% Fewer Tokens

If your business feeds AI things like reports, spreadsheets or system data, it's probably being sent in a much wordier format than it needs to be. TOON is a free way to write the same information in far fewer words - it strips out the repeated formatting that computer systems normally pad every entry with. Independent tests show it uses around 40% fewer tokens than the standard format most systems already use, with no drop in accuracy - if anything, slightly better. I've used it myself on data going into prompts, and the saving is real: a rare case where the cheaper option is also the better one.

Why it matters: If your business feeds structured data - reports, tables, system exports - into AI, the format you use is a lever on your bill you're probably not pulling yet.

👉 Action: Next time you're building a prompt around structured data, try TOON instead of the usual format and compare the token count - it's a straight swap, not a rebuild.

👉 Read about TOON

Feeding Text to a Model as a Picture, Not a Sentence

DeepSeek's OCR research asks an odd question: what if a model saw a page as an image instead of reading typed words? DeepSeek-OCR compresses roughly ten text tokens into a single "vision" token while holding 97% precision, and DeepSeek-OCR-2 - the next version - was open-sourced in January 2026. It's the furthest-out signal here: if reading is one of the biggest silent costs in a long AI conversation, re-engineering how the model reads may matter more than any prompt trick.

Why it matters: This is a preview of where the token meter itself may be rebuilt from the ground up - worth knowing about now, well before it reaches your everyday tools.

👉 Action: Nothing to implement yet - file this one under "watch," and expect optical compression to start showing up inside mainstream tools within the year.

👉 Read about DeepSeek-OCR

🎨 Weekend Playground

This weekend, open PlotLines - ten out-of-copyright classic novels, from Frankenstein to Mrs Dalloway, played out journey by journey on a genuine period map. Chris Moran (LinkedIn), the Guardian's editorial innovation lead, vibecoded it over a weekend while his daughter studied David Copperfield.

Why this matters: Moran's own reflection is the more interesting bit - the map was nice, but the real learning was the back-and-forth with the machine and the rabbit holes it opened (he lost an afternoon to Victorian coaching routes). That's a better template for a weekend with a young person than the tool itself: build something together, then chase one tangent as far as it goes.

👉 Mission:

• Pick a novel on PlotLines with a young person in your life and follow its journey across the map together

• Ask what surprises them about the route - distances, place names, how people actually travelled then

• Chase one rabbit hole as far as it goes - a place, a mode of transport, a historical detail neither of you knew - and see where an hour of curiosity takes you

📢 Share the Optimism

If The AI Optimist helps you think more clearly, forward it to someone else navigating the shift.

And here's the question I'm genuinely curious about this week: if you had to name the one AI cost in your business that nobody is currently accountable for, what would it be? Reply and tell me - I read every message and I'll come back to you personally.

Stay strategic, stay generous.

Hugo & Ben