Cheap Tokens, Expensive Mistakes

If you signed off an AI budget in the spring, the numbers underneath it have already moved. Not because the models got worse, but because the price of a good answer has fallen through the floor in the space of a few weeks. I read release notes the way other people read the football results, and the last month has been the busiest I can remember. Two things are happening at once, and only one of them is good news for your budget.

The price of a good answer is falling

Start with the repricing, because it is the part everyone is talking about.

- Anthropic shipped Haiku 5.5 on 7 October, quoting input at a tenth of what Haiku 4.5 charged for prompts up to 100k tokens and output at a tenth too, with the average run working out around 75% cheaper. Sonnet 5.5 arrived a week earlier, 30% faster and up to 30% cheaper.

- Google put Gemini 4 Argon out on 13 October at an introductory rate, with cached input at 95% off.

- OpenAI shipped GPT-6 Astra, then a further Sol update, then GPT-6 for everyone on 7 October, putting the model in front of the 1.2 billion people who use ChatGPT each week.

That is three vendors, one month, and a step change in what a token costs. A year ago you could plan around annual price movements. Now the ground shifts every few weeks, and it shifts in your favour.

What that means for a business is more interesting than the numbers themselves. Any business case you built in 2025 on 2025 token prices is now wrong, and it is wrong on the side that makes you look overly cautious. The opposite mistake is just as easy though: signing a three year deal that fixes a per seat price on top of a model you are not allowed to swap. The model is becoming a commodity part. Build your stack so you can change it the way you would change a hard drive, and the price cuts land on your P&L instead of your supplier's.

The hardware has gone the other way, which is the joke running underneath all of this. A few months ago you could pick up one of the Nvidia mining cards for a couple of hundred pounds. Once the full 64GB of VRAM became usable, the same card went past two thousand. Cheap tokens, expensive silicon.

A one in three failure rate is not a rounding error

The second thing happening is that agents have become genuinely useful, and that is where the good news stops being simple.

Stanford's 2026 AI Index has the numbers. On OSWorld, which tests an agent doing real tasks on a real computer, success went from 12% to around 66% in a year. On SWE-bench Verified, which tests code fixes, it went from 60% to near 100%. OpenAI reports GPT-6 Astra at 72.6% on OSWorld 2.0 at roughly 40 minutes per task, against 65.7% at roughly 75 minutes for the model before it. Google says Argon inside Antigravity pushed agentic code submissions up by more than 35% whilst cutting rollbacks by nearly a third.

Which sounds like the problem is solved. It is not, and the same report says so plainly: those agents still fail roughly one attempt in three on structured benchmarks.

Sit with that for a minute, because it is the number that matters to a business. A process that is right two times in three is not a junior member of staff. It is a process with a one in three defect rate, and you would not put that in front of a customer without a check on it.

The tokens for the failed attempt cost a fraction of a penny. The failed attempt itself costs whatever you would have paid if a person had made the same mistake, plus the cost of the person who has to catch it before it leaves the building. That second number has not moved at all.

Cheaper tokens do not make a cheaper mistake, and the mistake is where the money goes.

The cost of being wrong has not fallen

You can see the market working this out in real time. Documented AI incidents rose to 362 in the AI Index, up from 233 in 2024. The EU AI Act's transparency obligations arrive in August 2026 and require machine readable marking of generated content, with the high risk obligations following in December 2027. OpenAI's [textGrain](https://openai.com/index/eu-text-provenance/) watermark catches around 80% of 200 token passages and around 95% of 400 token passages at a 1% false positive rate, which sounds impressive until you read the rest: replacing 10% of the words with synonyms drops detection to 66%, and replacing 25% drops it to 17%. That is why access to the detector is limited to vetted researchers.

The lesson is not that detection is useless. It is that provenance has to be recorded when the content is made, not guessed at afterwards. If your plan for the August obligations is to run everything through a detector at the end, you are buying a 17% answer to a question a regulator will ask you in writing.

And if you think this is an EU problem, remember how procurement works. Your EU customers' compliance teams will send the questionnaire to their UK suppliers long before anyone at the [EU Commission](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) comes knocking.

## The bottleneck is not the model

Organisational adoption sits at 88% in the AI Index, and yet the picture underneath is deeply uneven. On Eurostat's data, only 19.95% of EU enterprises used AI technologies in 2025, up 6.47 points on the year. That ranges from 55.03% of large enterprises down to 5.21% in Romania. I would like the equivalent figure for the UK and I do not have one I trust, so I will not pretend otherwise.

Here is the part that tells you where the constraint actually is. Anthropic committed $100 million to train 10,000 Frontier Deployed Engineers by the end of 2027, starting with cohorts from Accenture, Bain, Capgemini, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk. When a model vendor pays to build deployment engineers inside the big integrators, that is the market admitting the scarce resource is not access to a model. It is somebody who can sit with your finance team, understand the process, and wrap the thing properly.

Which brings me back to the position I started with. The price collapse is real and you should take the money. But do not let a falling unit price talk you into skipping the wrapping, because the wrapping is the product and it always was. The tools you give the model, the guardrails on those tools, the instructions that tell it when to stop, and the check that catches the one in three: none of that gets cheaper when the tokens do.

What I am watching next

Two things. Whether the price cuts keep coming at this pace through the winter, and whether the verification tooling catches up with the agents, because right now the gap between what an agent can attempt and what you can safely let it finish is where most of the value and nearly all of the risk sits.

For what it is worth, I will be doing what I always do, which is trying the new releases in the home lab before they go anywhere near something that matters. It is a lot cheaper to find out there.

If you are building a business case this quarter, price the checking alongside the model. Most people do not, and it is usually the line that decides whether the project pays for itself. Tell me what you find, or get in touch if you want a second pair of eyes on how you have wrapped it.