In partnership with
Master Claude AI (Free Guide)
The professionals pulling ahead aren't working more. They're using Claude.
Our free guide will show you how to:
Configure Claude to be the perfect assistant
Master AI-powered content creation
Transform complex data into actionable strategies
Harness Claude’s full potential
Transform your workflow with AI and stay ahead of the curve with this comprehensive guide to using Claude at work.
| Zymbos Intelligence · Sunday 1 November 2026 | |||||
| |||||
|
A note on timing This issue reaches you a few days later than usual. The news covers the week to 27 October. | ||
Capable AI is getting cheaper faster than most contracts can follow. This week OpenAI halved the price of its new everyday model, Google Cloud started putting OpenAI's models on the same bill as everything else, Google went after the cost of idle AI agents, Barclays linked job cuts and savings partly to AI, and Anthropic's stock-market filing showed what the most advanced AI costs to build. AI is moving from a scarce product paid for by the unit to a feature folded into bills you already pay, and the costly mistake is a long contract signed at today's price. This issue maps the market; on 11 November, Issue 035 turns it into your 2027 AI budget. | ||
| ||
Pricing · US OpenAI halves the price of its everyday model OpenAI launched GPT-6.1 Terra on 26 October at half the price of the version it replaces, InfoWorld reports. Terra sits in the middle of OpenAI's range, built for everyday work such as coding, reading images and operating software on screen, with cheaper models below it and more capable, more expensive ones above. OpenAI charges for its models by the token, a chunk of text roughly three-quarters of a word long, so a lower token price cuts the cost of every task. In test results reported by InfoWorld, Terra scored higher than its predecessor on a set of professional tasks, while the cost of each completed task fell from about £2.68 ($3.58) to about £0.76 ($1.01). Analysts quoted by InfoWorld add two cautions. Test results are no substitute for trying the model on your own work. And the time people spend checking AI output can wipe out the saving: in one illustrative coding example, 5.4 extra seconds of human review per task was enough to cancel it. McGann's TakeA price list cut in half is a reason to reopen your numbers this month. The saving only becomes real when the cheaper model produces work you accept without extra checking. Measure the cost of one finished, approved task before and after you switch, and take that figure into your next renewal. Price note: ~£0.75 ($1) per million tokens in and ~£3.75 ($5) per million tokens out, at £1 = $1.3337. | ||
Procurement · Global OpenAI's models arrive on your Google Cloud bill OpenAI's models are coming to Google Cloud, the platform where businesses rent computing power and software from Google. According to Google Cloud, they are rolling out in private preview to selected business customers through Model Garden, Google's catalogue of AI models, with wider access to follow. The commercial detail matters more than the model names. Google says the spending draws down a customer's existing cloud commitments, the sums a business has already agreed to spend with Google over a set period, and appears on a single Google Cloud invoice. CNBC reports that the move puts OpenAI's models on all three of the biggest cloud platforms, although OpenAI still calls Microsoft its primary cloud partner. For buyers, there is now another route to the same models and another place to negotiate. There is also a new risk: the cost of AI can vanish into a cloud bill that finance approves as one line. McGann's TakeThis is bundling in its early form. The AI charge moves inside a bill you already pay, which is convenient, and convenience is how prices stop being questioned. If you buy this way, insist on usage reports by model so you can still see what the AI part costs. | ||
Infrastructure · Global Google goes after the cost of AI agents doing nothing An AI agent, software that carries out a multi-step task on its own, spends much of its working life waiting: for a person to approve something, for another system to reply, or for its next instruction. If the computer underneath stays switched on, the waiting is billed. Google's new Agent Compute Engine tackles that by putting idle agents to sleep inside small, sealed-off virtual computers and waking them with their work intact, InfoWorld reports. It is not yet available: SiliconANGLE reports a rollout in the coming months, so any saving is still Google's claim and no customer has reported results. Analysts quoted by InfoWorld list the catches. Each sleeping agent takes up storage, which costs money at scale. The service ties workloads more closely to Google. And an agent paused at the wrong moment can repeat an action, such as a payment, when it wakes. "Every external action has to be safe to repeat," as one analyst put it. McGann's TakeThe token price is the number vendors advertise. The full cost of an agent also covers waiting time, storage, retries and the people who check its work. Before you accept any agent's cost estimate, ask the supplier to price one complete job from start to finish, idle time included. |
Workforce · UK Barclays cuts 5,000 roles and calls AI "the icing on the cake" Barclays plans to cut about 5,000 roles this year, roughly 5% of its workforce, the BBC reports. Around 1,000 have already gone and the remaining 4,000 are due by the end of the year. About 60% of the cuts fall in support functions, including software developers, and the bank expects a further £1.1bn (~$1.5bn) of savings in 2027. The BBC reports the cuts as part of the bank's growing reliance on AI. Chief executive C. S. Venkatakrishnan was more measured, calling AI "the icing on the cake not the cake" in a wider overhaul of the bank's technology. The Unite union accused the bank of using AI as an excuse to shed loyal staff. The Associated Press places Barclays among several large employers now linking AI spending to the size of their workforce. The lesson for any organisation is about accounting. A saving should show what changed, whether systems, processes, people or software, before anyone decides how much of it AI earned. McGann's TakeA large savings figure with AI somewhere inside it is easy to announce and hard to audit. Before anyone in your organisation credits AI with a saving, ask for the bridge: what the work cost before, what it costs now, and what else changed. Until AI can be separated out, it has not earned a line in the business case. | ||
Markets · US Anthropic's filing shows what the most advanced AI costs to build Anthropic, the company behind the Claude models, filed on 26 October to list its shares on the Nasdaq stock exchange, the Associated Press reports. The filing gives the first public view of a leading AI developer's accounts. Revenue for the first half of 2026 was £12.2bn ($16.3bn), according to the BBC. The headline loss of £60.5bn ($80.7bn) is mostly a one-off accounting charge of £54.3bn ($72.4bn), because loans that convert into shares rose in value along with the company. Even without that charge, the company made a loss. CNBC picks out the risks the company itself lists. Two stand out for buyers: a shortage of electrical power to run its computers, and the chance that customers choose cheaper or open-source models over its most advanced ones. In other words, one of the leading model makers lists cheaper alternatives as a threat to its business. McGann's TakeThis is the best evidence for the other side of this week's argument. The most advanced AI is expensive to build, and its makers need customers who keep paying for it. Expect them to ask for long commitments; their own filing names cheaper models as a threat, and a short contract keeps that option yours. | ||
| ||
This Week's Analysis The Scarcity Trap Fifteen years ago, a mobile phone bill counted every megabyte, and heavy users dreaded the end of the month. Today most people never think about data; it comes inside the plan. AI is on the same road, from scarce and metered to bundled and invisible. You will sign AI contracts somewhere along that road, and where you sign decides what you pay. My claim: the professionals who time this shift will save real money, and the trap is a long commitment signed at today's scarcity price. Three reasons. First, the price of capable AI is falling fast. OpenAI's Terra (story 1) costs half what its predecessor did, and the cost of a finished task in published tests fell by about two-thirds. Each new release resets the going rate. A contract priced on last year's rate card locks in a number the market has already left behind. Second, bundling is how sellers protect themselves when the unit price falls. When the price of what you sell keeps dropping, you wrap it inside something bigger. Google Cloud now lets businesses pay for OpenAI's models out of money already committed to Google (story 2). Uber has given ChatGPT Work, an OpenAI agent that works across email, calendars and documents, to about 30,000 staff without saying what it costs, TechCrunch reports. The AI price is still there in both cases. It is harder to see, and harder to challenge at renewal. "Pay today's price for today's work. Do not promise to pay it in 2028." Third, contract length is the one lever the buyer controls. You cannot set the price curve. You can choose monthly over annual and annual over three years, and you can ask for a price review, usage reports and the right to switch models at each renewal. Suppliers have their own reasons to want long deals. Anthropic's filing (story 5) lists customers switching to cheaper models among the risks to its business. A long contract protects the seller from that switch. A short one keeps it open to you. The strongest counter-argument is that the very best AI may stay expensive for years. Building it costs a fortune, as story 5 shows. In the same published tests that flattered Terra, Anthropic's most capable model scored slightly higher at nearly nine times the cost per task. If that extra accuracy matters to your work, waiting for the bundle could leave you a year behind the people who paid. That argument is right about the model and wrong about the contract. Pay for the premium model where you can show it changes the result: the contract clause that carries the risk, the analysis a client pays for. Use cheaper models, or the AI already inside software you own, for routine work. Then keep every commitment short enough to renegotiate when the next price cut lands. The bundle is coming. Whether it saves you money depends on what you signed before it arrived. |
| |||||||||||||||||
OpenRouter AI model marketplace · No monthly fee · Pay as you go
One account, hundreds of AI models, and you pay only for what you use. What it is OpenRouter is a website that gives you access to AI models from many different companies through one account. Instead of paying a monthly subscription to each, you buy credits and pay for what you use, at prices published model by model. The site also has a chat page, so you can try models in your browser without writing any code. That makes it a live window on what AI actually costs: this week's Terra is listed there at OpenAI's own price, next to hundreds of rivals. It is also the most technical tool we have reviewed in a while. Prices are quoted per million tokens, you choose between models and the companies that run them, and the bill depends on how much you use. If you want one simple assistant at a fixed monthly price, this is the wrong tool. Who it suits People who already use AI every day and want to compare models on their own work, small teams who would rather pay for use than buy a seat for everyone, and anyone who wants to see what AI really costs per task. Ratings Scores are from a desk review of the live OpenRouter site, pricing page and model listings on 30 October 2026. We have not run a hands-on reliability test.
The chat page opens in the browser without signing in. The rest of the site assumes you know what tokens and providers are, and 600+ models is a lot of choice for a newcomer.
600+ models from 80+ providers on paid plans, each with its price shown provider by provider. Terra appears at OpenAI's own rate, with cheaper and dearer routes listed alongside.
Budgets and spend controls on Standard, Business and Enterprise. Advanced notifications only on Business and Enterprise. Paying in advance with credits also limits what you can spend.
Automatic fallback to other providers is built in, and failed or fallback attempts are not charged for model use. Contractual service levels only on Enterprise. No independent uptime test.
No monthly fee and no lock-in, but a 5.5% fee on every credit purchase. Strong for light or varied use; a heavy daily user may still be better off with a flat subscription. Verdict The clearest live view of what AI costs, model by model, and a sensible way to pay for light or varied use without another subscription. Too fiddly to be the only AI tool for someone who wants one simple assistant. Run one real task through it, check the bill, and compare it with what your subscription costs you each month. Pricing verified on the live OpenRouter pricing page, 30 Oct 2026. Free tier: £0 ($0); 25+ free models, capped at 50 requests a day, or 1,000 a day once you have bought ~£7.50 ($10) of credits. Paid tier (Standard): no monthly fee; each model at its provider's listed price, plus a 5.5% fee when you buy credits. Team tier (Business): no monthly fee; the same model prices, an 8% fee on credit purchases, advanced notifications and more workspaces. Enterprise: priced by contract. Example: GPT-6.1 Terra is listed at ~£0.75 ($1) per million tokens in and ~£3.75 ($5) per million tokens out, so a 2,000-word document in and a 1,000-word summary out costs under 1p (~1¢), more if the model reasons at length before answering. Model prices follow the providers and can change at any time. Sterling figures use the 27 October rate of £1 = $1.3337. | |||||||||||||||||
| |||||||||||||||||
The Stack Cost Audit Cost per result · Works in Claude or ChatGPT Most people can tell you what their AI tools cost each month. Very few can say what each one costs per piece of finished work, and that is the number that matters when prices are falling. This prompt builds that view. Use it before any renewal, after adding a new AI tool, or when two tools seem to do the same job. Gather last month's invoices and receipts, any usage figures, and a line on what each tool is for. Remove card numbers, account IDs and client names before you paste anything in. Good output is a single table: each tool, its true monthly cost with annual plans divided by twelve, what it delivers and the cost per result. Then the three overlap questions, a keep, test, downgrade or cancel call for each tool, and a list of the figures you still need to find. It should mark missing data as missing and never fill a gap with a guess. It works in Claude or ChatGPT. You are my AI spending auditor. Work only from the figures I give you below. Do not guess prices, usage or savings. Where a figure is missing, write MISSING and carry on. STEP 1. TRUE MONTHLY COST For each AI tool, show: - what I pay and how often (monthly, annual or usage-based) - the true monthly cost: annual plans divided by 12, plus usage charges, top-ups and platform fees - who uses it, and how often - the one result it is meant to deliver (for example: client reports drafted, meetings summarised) - how many of those results it delivered last month, and the cost per result - the renewal date and notice period STEP 2. THE THREE OVERLAP QUESTIONS 1. Which tools do the same job for the same people? 2. Which jobs are already covered by software I pay for anyway, such as my office suite, cloud account or another platform? 3. Which tool am I paying a premium for, and can I show it produces a better result than a cheaper option? STEP 3. DECISIONS Give one table row per tool: KEEP, TEST A CHEAPER OPTION, DOWNGRADE or CANCEL, with the reason and the evidence. Keep verified savings and estimated savings in separate columns. Name the one contract that ties me in longest, and the date I can next change it. End with a list of the missing figures I need to find. MY DATA: [Paste your subscriptions, invoices, usage figures, who uses each tool and what it is for. Remove card numbers, account IDs and client names first.] Know your cost per outcome, not just your cost per month. |
| ||
Closing Perspective The cheapest AI you will ever buy The cheapest AI you will ever buy is the one already inside something you pay for. That is where this market is heading, and it changes what a sound buying decision looks like. For the past few years the sensible move was to pay for the best model you could get, because the gap between models was wide and the tools were new. That still holds where the best model makes a difference you can measure. Everywhere else, prices are falling and sellers are folding AI into bigger bundles. A three-year contract signed today bets that neither will happen. My prediction: by 30 June 2027, Microsoft, the biggest name in office software, will move at least one feature that today needs the paid Microsoft 365 Copilot add-on into its standard Microsoft 365 business plans, with no separate AI charge. It counts even if the base plan price rises. It does not count if the feature arrives only as a trial or a time-limited offer. You will be able to check it on Microsoft's own plan pages. Until then, find one number. Reply with your true monthly AI spend, to the nearest £10 or $10, with annual plans divided by twelve and usage charges included. Most readers will have to go and look. That is the point. John McGann Founder, Zymbos AI |
Zymbos Intelligence zymbos.ai You're receiving this because you subscribed at zymbos.ai © 2026 Zymbos Intelligence · John McGann · London, UK Zymbos Ltd · Company No. 16198848 · Teddington, England |