Insights · AI and pricing

The meter moved

Jatinder Purewal · ACMA, CGMA · Executive MBA · 6th October 2026 · 17 min read

On 18 September I argued that AI was being repriced through the quantity inside the subscription rather than the price on it. Since then OpenAI has cut the allowance inside its $200 ChatGPT plan and put a $500 tier above it, Google is taking its better models out of its cheapest Gemini plans, and Microsoft will switch on a usage meter by default for new Copilot Business seats from 19 October, or 1 December through resellers. All three left the price alone, and on the meter itself rates fell. Here is what moved, why it is moving much faster than mobile pricing did, and the five things I would re-check before a 2027 budget is signed off.

A note on scope. This is a practitioner's read of the vendors' own pricing pages and launch posts, and of my own subscriptions, as at 6 October 2026, with judgements flagged as judgements. I pay for several of the products named and have no commercial relationship with any vendor. None of the sources here discloses what it costs to serve a subscription, so nothing here asserts one. It updates The meter under the seat, which carries the worked 400-seat model and a downloadable spreadsheet; this piece does not repeat them.

The short version, for anyone signing off a 2027 budget: re-price the meter and the seats on today's terms, measure cost per completed task rather than per token, put allowance terms into the contract and set your own ceiling on every meter before the vendor's default applies, write the routing policy, and bring agents inside the controls before they get credentials.

In this piece:

1. What moved, and when

2. The headline rates match; the cost of a task does not

3. The allowance shrank, and the meter is being switched on

4. The floor is low, and it moves both ways

5. Why this is moving faster than mobile did

6. The seat and the person are coming apart

7. Five things to re-check before the 2027 budget

8. In brief

Illustration, AI-generated: a brass utility meter mounted beneath the seat of a navy leather office chair

What moved, and when

On 29 September the price of ChatGPT's $200 Pro plan did not change. The allowance inside it did. OpenAI reopened the plan to new subscribers with what its help centre calls "a lower usage allowance than previously offered", kept the price at $200, and told eligible existing subscribers they keep the old allowance through 29 October and then move to the new one at the same price. Its stated reason is that its models are now more efficient.

On 23 September I added a dated note to my last piece on my website: for anyone buying seats, the allowance, not the rate, is still where the money moves. Six days later OpenAI moved it on its own $200 plan: it kept the price and cut the quantity. Whether each unit now does more is what a buyer has to measure.

It was one of a run of moves, each announced on a vendor's own pricing page, launch post, or partner or admin notice:

22 September. Anthropic released Claude Opus 5.5 at $4 and $20 per million input and output tokens, 20 per cent below Opus 5. OpenAI released GPT-6 Sol at $2 and $10 and GPT-6 Luna at $0.10 and $0.50, half the GPT-5.6 prices or less, and told VentureBeat the prices are permanent.
28 September. Anthropic released Sonnet 5.5 at an unchanged $2 and $10, estimating it costs up to 30 per cent less to run than Sonnet 5.
29 September. OpenAI released GPT-6.1 Sol at $2 and $10, added the $500 tier, cut the $200 allowance, and launched dots, agents that keep working on their own cloud computers when the user is away.
30 September. Google announced Gemini 4 Argon, its new frontier model, at an introductory $2 and $10 with $4 and $20 to follow. For now it is rolling out only to selected cyber defenders in its Fairwind Program and a group of trusted testers.
From 9 October. Google is taking Flash and Pro out of free Gemini, and Pro out of its AI Plus plan on dates it is giving subscribers by email. AI Pro and AI Ultra keep all three models.
From 19 October. Microsoft will switch usage-based billing on by default for new Microsoft 365 Copilot Business licences bought direct, and from 1 December for those bought through its resellers, in supported markets including the UK, with a default ceiling worth $40 a user a month against a $21 list-price seat.
And at the low end, since 10 September, DeepSeek's V4.1 Flash at 15 cents per million uncached input tokens off-peak and double at peak hours. OpenAI's own Luna, at 10 cents, sits lower still.

What moved in AI pricing, September to December 2026: DeepSeek V4.1 Flash with peak and off-peak tokens; Claude Opus 5.5 at a 20 per cent lower rate; GPT-6 Sol and Luna at half the GPT-5.6 prices or less; Sonnet 5.5 at the same price; GPT-6.1 Sol, the $500 Pro tier and the cut to the $200 allowance; Gemini 4 Argon announced at an introductory $2 and $10; Google's free Gemini tier losing Flash and Pro from 9 October, with AI Plus losing Pro on dates emailed to subscribers; and Microsoft's usage meter switched on by default for new Copilot Business seats bought direct from 19 October and through resellers from 1 December

The headline rates match; the cost of a task does not

Two vendors now sell their mid-tier models at one headline rate, and a third has announced it. GPT-6 Sol came in at half the promotional rate of GPT-5.6 Sol, onto the $2 and $10 per million input and output tokens Anthropic already charged for Sonnet, GPT-6.1 Sol stayed there, and Google announced Gemini 4 Argon at the same rate for an introductory period, for a model not yet on general sale. Above them the top tier stays at $10 and $50, for GPT-6 Astra and Claude Fable 5.1. Below them OpenAI's own Luna costs $0.10 and $0.50, less than DeepSeek's V4.1 Flash at $0.15 and $0.60 off-peak for uncached input and output.

When the headline rate matches, the differences move into the terms around it. GPT-6.1 Sol shows how: OpenAI kept the $2 and $10 and halved cached input, to $0.10 per million tokens against $0.20 on GPT-6 Sol and on Sonnet 5.5. Anthropic says cached input makes up the majority of the cost of agentic and coding work. And the vendors' own claims are now about cost per task: Anthropic says Sonnet 5.5 needs fewer tokens for the same work, and OpenAI sells GPT-6.1 Sol under the heading "Near-Astra intelligence for a fifth of the price". For a buyer that changes the unit worth measuring. A price per million tokens tells you what a token costs. It does not tell you what a reconciliation, a variance commentary or a contract review costs. It is why task type and the model that served the request were two of the five fields I said to record from day one.

In my last piece I said the rate had been the smallest mover at the vendors I watched most closely. On the evidence of August it was. September changed that for anyone buying at the meter. The diary of dates I published with it has aged as well: the GPT-5.6 Sol promotion that OpenAI had committed to hold until at least 21 November was overtaken a month later by a newer model at half the price, which OpenAI told VentureBeat is permanent: a statement, not a contract.

Output price per million tokens at standard short-context rates on 6 October 2026, on a log scale: GPT-6 Astra and Claude Fable 5.1 at $50; Claude Opus 5.5 at $20; GPT-6.1 Sol and Claude Sonnet 5.5 at $10, and Gemini 4 Argon announced at $10 for an introductory period and $20 after it, not yet on general sale; DeepSeek V4.1 Flash at $1.20 peak and $0.60 off-peak; GPT-6 Luna at $0.50

The allowance shrank, and the meter is being switched on

Illustration, AI-generated: a closed laptop on a desk at night, wired into a brass meter whose needle has just started to rise

At OpenAI the subscription moved the other way: the same price, with less inside it. OpenAI now sells three Pro plans, at $100, $200 and $500 a month, all billed monthly. According to The Next Web and Digital Trends, the $200 plan's Work and Codex usage falls from 20 times the Plus allowance to 10: at once for new subscribers, and from 30 October for existing ones, who keep the old allowance until then. Half the allowance at the same price doubles the price of each unit of allowance; whether it doubles the cost of the work depends on how much more each unit now does. The $500 plan carries 25 times the Plus allowance and adds Ultrafast on GPT-6 Astra, which uses the included allowance at eight times the standard rate; Fast mode uses it at two and a half times. Speed is also a priced product in the API: Opus 5.5 in fast mode costs double its standard rate. Eligible existing subscribers are also reported to be getting a one-off grant of 62,500 credits, worth about $2,500, which expire at the end of the year.

Put the tiers side by side and the volume discount has gone. Take the Plus allowance, at $20, as one unit. On the multiples OpenAI and the press have given, Pro 100 buys five units for $100, the new Pro 200 ten for $200 and Pro 500 twenty-five for $500: about $20 a unit on every tier. The old Pro 200, at twenty units for $200, was the only tier that sold usage at $10 a unit, and it is the one that changed. Anyone who budgeted the bigger plan as the cheaper way to buy volume should re-run the numbers.

Google is doing the same thing at the bottom of its range for personal accounts. From 9 October its free tier keeps only Flash-Lite, and its cheapest paid plan, AI Plus, loses Pro on a date Google is emailing to each subscriber. AI Pro and AI Ultra keep all three models. Since May its limits have been compute-based, refreshing every five hours up to a weekly cap, and the new effort settings use more of the limit at the higher levels. The lever is the same one OpenAI pulled: what the price buys, not the price.

If you buy Microsoft 365 Copilot, this is the change to put in the diary. New Microsoft 365 Copilot Business licences come with pay-as-you-go billing switched on by default, for usage-based features such as Copilot Cowork and the Work IQ APIs: from 19 October if bought direct from Microsoft, and from 1 December if bought through its Cloud Solution Provider resellers, a date already moved once from 2 November. It starts in supported markets, the UK among them; eleven others, including France and Germany, follow later. Copilot Business is the version for organisations of up to 300 users, bought on top of a Microsoft 365 plan. The default ceiling is set at 4,000 Copilot Credits per user per month, which at one cent a credit is $40 a user a month on top of a seat listed at $21 in the US ($18 on a promotion that runs to the end of 2026; in the UK, £16.10 before VAT, or £13.80 on the promotion). Microsoft's own example is 100 users drawing up to 400,000 credits a month between them. Resellers set their own prices, so budget from your reseller's quote. Administrators can change the ceiling, and the charge follows actual usage, so $40 is a limit, not a forecast. Existing subscriptions keep their terms, and a spending policy already in place stays. But Microsoft's notice to administrators says that for an existing Copilot customer, buying one new Copilot Business seat switches usage-based billing on across all its Microsoft 365 Copilot subscriptions, and the policy it creates also covers future eligible services. For a business that has budgeted AI as a fixed seat, this is the meter under the seat, written into the default.

Microsoft 365 Copilot Business, for new licences bought direct from 19 October 2026 or through resellers from 1 December: the seat at a $21 US list price per user per month, and a usage meter switched on by default with a ceiling of 4,000 Copilot Credits per user per month, which is $40 at one cent a credit; for 100 users, $2,100 a month of seats and up to $4,000 a month of metered usage; the ceiling can be changed and the charge follows actual usage; for an existing Copilot customer, one new seat switches the meter on across its Copilot subscriptions

OpenAI cut the price on the meter and the allowance in the seat in the same fortnight, and gave efficiency as the reason for both. Anthropic went the other way in the seat: alongside Opus 5.5 it raised the five-hour limits on its paid plans. My judgement is that the vendor, not the buyer, sets the default for how an efficiency gain is shared between price and quantity, and can set it differently on the meter and in the seat, in either direction. Whether a smaller allowance on a more efficient model buys the same work as before is a claim the buyer can only test one way: by measuring work done per unit of allowance, before and after the change.

I see this from the subscriber's side as well. The limits on my own coding plan moved twice in eight days. On 14 September Anthropic's summer promotion ended and its weekly limits settled 25 per cent above where they were before it: a permanent rise of a quarter, and a cut of about a sixth against the summer. On 22 September it raised the five-hour limits on its paid plans alongside Opus 5.5. The price of the plan did not move either time. After eleven years of running my own business on a marketplace that changes its terms without asking, I read the allowance page before the price page. It is the same habit, applied to a new supplier.

OpenAI's September 2026 changes, two ways: on the meter, GPT-6 Sol and GPT-6.1 Sol at $2 and $10 against GPT-5.6 Sol's promotional $4 and $20; in the seat, ChatGPT Pro 200 kept its $200 price with a lower allowance, reported as half the Work and Codex usage, and Pro 500 was added above it

The floor is low, and it moves both ways

Below the frontier the floor is low, but it moves both ways, and not only because of open weights. On uncached input, the lowest rate in this piece is OpenAI's closed Luna; on cached input, DeepSeek's off-peak rate is lower. DeepSeek brought in peak and off-peak pricing with its V4 models in August, raising developer prices by up to 1,100 per cent depending on the model, the token type and the hour, as Quartz reported. Its V4.1 Flash, released on 10 September at lower prices, costs $0.15 per million uncached input tokens and $0.60 per million output tokens off-peak, and double at peak. Even so, its off-peak output price is more than double the flat $0.28 that V4 Flash charged before August. And Google has already dated a doubling of its Gemini 3.8 Flash prices for 1 January 2027.

DeepSeek's peak runs from 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, outside Chinese public holidays: the Beijing working day. In London that is the small hours and the start of the working day, until 11:00 in summer and 10:00 once the clocks go back on 25 October. On a weekday outside those holidays, an agent scheduled for 02:00 London time pays the peak rate.

Pricing the same unit differently by time of day is one of the oldest devices in the mobile tariff book. When I was at O2 in the early 2000s, building customer value by tariff, the same minute was priced differently by time of day. It is back, on an AI price list.

Alibaba, Moonshot and Zhipu also release open-weight models that a company with the hardware can run under each model's licence. The Next Web's coverage of the 22 September cuts named Alibaba, DeepSeek and Moonshot as the pressure behind them, writing that Chinese open-weight models "now do much of this work for nothing". My judgement is that this pressure does not go away: when capable models can be run in-house, the price of routine work is pulled towards the cost of running it.

For a finance leader the question is not political. It is where the data is processed, how long it is kept, whether it trains the model, and under which law. DeepSeek's own privacy policy says it collects, processes and stores personal data in the People's Republic of China; a copy of the same open weights hosted in the UK or the EU goes most of the way to answering the jurisdiction question, at that host's price, though remote access and subprocessors still need checking. In January 2025 Italy's data protection authority ordered DeepSeek to stop processing Italian users' personal data. Routing routine work to the floor has two halves, the saving and the approval of where the data goes, and both need pricing before the routing policy is written.

Why this is moving faster than mobile did

Illustration, AI-generated: an early-2000s mobile phone with a keypad and a modern smartphone on a desk, with a small brass meter standing between them

Here is the part that changes the picture. In the UK mobile examples below, five pricing devices are spread across about two decades. AI vendors have used all five in 2026 alone, across consumer plans, business seats and the API, and consumer AI subscriptions are only about four years old. These are examples, not first appearances: DeepSeek first discounted its off-peak hours in February 2025. The reasons that follow are why I expect AI pricing to keep moving faster than mobile's did.

An allowance inside the price, with a rate above it: inclusive minutes on UK contract tariffs by 2000; GitHub Copilot's credits inside the seat, with paid usage above them, from June 2026.
The same unit priced by time of day: peak and off-peak minutes in the early 2000s; DeepSeek's peak and off-peak tokens from August 2026.
An advertiser paying for the free tier: Blyk, an ad-funded UK mobile network, from 2007 until it closed its UK service in 2009; ads on ChatGPT's Free and Go plans, in the US from February 2026 and in the UK from June.
The same price with less inside it: O2 ended unlimited data for new smartphone customers in June 2010; OpenAI cut the allowance inside its $200 plan in September 2026.
A regulator asking whether the customer could see what they were paying for: Ofcom's findings on out-of-contract customers in 2019; the Competition and Markets Authority's consumer-protection case, opened in July 2026, on how Microsoft presented changes, including Copilot, to its Microsoft 365 Personal and Family plans, with no finding yet.

The reasons are structural. Mobile prices were tied to physical things: handsets recovered over contracts of up to two years, networks that took years to build, tariffs sold through shops. An AI price can be changed by editing a web page, and the product changes every few weeks: GPT-6.1 Sol arrived one week after GPT-6 Sol. An API customer can switch models far faster than a mobile customer could leave a contract, though re-testing, data terms and committed spend still take time.

The same five pricing devices in two industries: an allowance inside the price, time-of-day pricing, an advertiser paying for the free tier, the same price with less inside it, and a regulator asking whether the customer could see the price. The UK mobile examples run from 2000 to 2019; the AI examples are all from 2026, across consumer plans, business seats and the API; examples, not first appearances

The seat and the person are coming apart

Illustration, AI-generated: a laptop left open on a desk at night, beside a key in a leather tray and an unused access badge on its lanyard

The unit at the centre of most AI budgets, the seat, assumes a person. On 29 September OpenAI launched dots, always-on agents that work for one person and keep working when that person is away, and began piloting specialist dots for companies, each set up with "its own identity, credentials, and access to the systems it needs".

Each dot has its own cloud computer and runs on GPT-6 Astra, OpenAI's frontier model. The first is included in Pro and Business Premium. Pro subscribers in the UK, the European Economic Area and Switzerland do not get dots at launch, but Business Premium is available across all supported regions, so in the UK the agent arrives through the business plan rather than the personal one. For the first month OpenAI is not counting dots' usage against plan allowances, and it will publish usage terms for each plan after that. On its normal terms, conversations with a dot do not count toward usage limits but tasks it starts in Codex or ChatGPT Work do, so consumption will no longer follow working hours.

The specialist dots build on early testing inside OpenAI across procurement, invoice processing, email marketing, customer support and commercial contracting. An agent with credentials in procurement and invoice processing is, for a finance leader, a controls question before it is a productivity one. Most of the answers already exist. They were written for people.

Delegated authority. What can it commit the company to, and up to what value? OpenAI's Custom Rules let a user allow an action, require approval or block it, but OpenAI itself says a dot can make mistakes following its rules. The limits belong in the real matrix and in the access the connected systems grant.
Segregation of duties. The agent that raises a purchase order should not be the one that approves the invoice against it, or the one that can change a supplier's bank details.
Ownership and access. Every agent needs a named owner, and credentials that never sleep need reviewing on the same cycle as anyone else's.
Evidence. Resetting a dot deletes its conversations and memories, so keep the record an auditor will ask for where the company controls it.
Cost. An always-on agent can draw on the allowance at any hour. Budget it as a variable cost, with an owner, a hard spend limit and a way to stop it.

At Parexel I signed purchase orders and contracts under a delegated-authority matrix that sat inside the SOX control environment of a NASDAQ-listed business. Nobody would have given a new member of staff authority to commit spend without a line in that matrix, and the same rule should apply to an agent. In my own business today, forecasting and margin analysis run through agent-run pipelines, and nothing they produce informs a decision until it has been checked back to source and cross-checked across models. That is a control, and it was built as one. The discipline is the same in both places: know what each actor may do, and keep the evidence that it did only that.

Agents with credentials, read from the finance seat: specialist dots with their own identity and access need a named owner and access reviews; Custom Rules need limits enforced in the connected systems and a line in the delegated-authority matrix; work in procurement and invoice processing needs segregation of duties; the activity record needs keeping where the company controls it; an always-on agent needs a budget line, a spend limit and a way to stop it

Five things to re-check before the 2027 budget

For most December year-end businesses the 2027 budget is being built now, and an AI line drafted over the summer was drafted at summer prices. The ChatGPT Pro and Gemini changes are signals for most finance teams rather than bills. Microsoft's can become one without anyone deciding to spend: new Copilot Business seats arrive with the meter on, from 19 October direct and 1 December through resellers, and for an existing Copilot customer one new seat switches it on across the tenant. Five re-checks:

1. Re-price both halves on today's terms, and diary what is already dated: Google's Gemini changes from 9 October (for AI Plus, on the dates in Google's emails), Microsoft's default meter from 19 October for direct purchases, the smaller ChatGPT Pro 200 allowance after 29 October, the same Microsoft default through resellers from 1 December, Gemini 3.8 Flash at double the price from 1 January 2027, and Argon's introductory price ending on a date Google has not given. The prices are in dollars, so show the exchange-rate exposure in a sterling budget, and treat credits bought in advance as prepayments that can expire.

2. Measure cost per completed task, not per token, counting only work that was accepted after review and including the rework and checking time. With headline rates matched, the differences sit in tokens per task, cached-input rates and speed premiums, and task type and the model that served the request are the fields that make them visible.

3. Put the allowance terms in the contract: notice before an allowance change, grandfathering for the term, compensation when the allowance falls, protection on what a credit buys, and a ceiling you set yourself on every meter. Microsoft's default for new Copilot Business seats is 4,000 credits per user per month, in a spending policy that also covers future eligible services; decide your own number, and your own scope, before the meter switches on; your administrators can read Microsoft's notice, MC1476200, in your own admin centre. Give every meter an owner who sets its ceiling, reconciles its usage each month and decides what happens when a cap stops work. A self-serve monthly plan gives you the vendor's standard terms and little more, so negotiate these into an enterprise agreement before signature.

4. Write the routing policy: which work may run on the low-cost floor, on whose infrastructure and under which law; which stays on frontier models; when paying for speed is worth it; and which jobs can wait for off-peak hours.

5. Bring agents inside the control framework before they get credentials: a named owner, a line in the delegated-authority matrix with the limits enforced in the connected systems, segregation of duties, access reviews, records kept where the company controls them, a data protection impact assessment where the processing is likely to be high-risk, and a hard spend limit.

My judgement is that this is still far from a settled place. Adoption is still growing and the fight for share goes on. The largest labs are also heading for the public markets: Reuters reports that Anthropic could list days before the US midterm elections in November, and OpenAI's chief executive has said it will not list this year. Public investors will look for margins as well as growth. I expect more changes of this kind, in both directions, over the medium term, one to three years, before pricing settles. Budget for the meter to move again, not for today's price list to hold.

Mobile gave finance teams about twenty years to learn how a meter works. AI is giving them a budget cycle.

In brief

What changed in AI subscription pricing in September and October 2026? OpenAI, Google and Microsoft each moved to change what a subscription includes rather than its price. OpenAI cut the allowance inside its $200 ChatGPT Pro plan and added a $500 tier. Google is removing Flash and Pro from free Gemini from 9 October, and Pro from AI Plus on dates it emails to subscribers. Microsoft will switch usage-based billing on by default for new Copilot Business seats, from 19 October if bought direct and 1 December through resellers, in the UK and most other markets. On the API, OpenAI's new mid-tier model came in at half its predecessor's promotional rate, matching Anthropic at $2 and $10 per million tokens.

What is changing in Microsoft 365 Copilot billing? New Microsoft 365 Copilot Business licences in supported markets, including the UK, come with pay-as-you-go billing switched on by default: from 19 October 2026 if bought direct from Microsoft, and from 1 December 2026 if bought through its Cloud Solution Provider resellers. The default ceiling is 4,000 Copilot Credits per user per month, which administrators can change; at $0.01 a credit, that is $40 a user a month. Existing subscriptions keep their terms and any spending policy already set, but Microsoft's notice to administrators says that for an existing Copilot customer, buying one new Copilot Business seat switches usage-based billing on across all its Copilot subscriptions.

Did ChatGPT's $200 Pro plan go up to $500? No. Pro 500 is a new tier at $500 a month, with the largest allowance and Ultrafast on GPT-6 Astra. Pro 200 stays at $200 with a lower usage allowance for new subscribers; eligible existing subscribers keep the old allowance through 29 October 2026, then move to the new one at the same price. Pro 100 is $100.

How are Chinese open-weight models affecting AI prices? They add pressure at the low end, though the floor moves both ways: DeepSeek raised its prices in August when it introduced peak-hour pricing, and its V4.1 Flash now costs $0.15 per million uncached input tokens off-peak, double at peak. The Next Web's coverage of the 22 September price cuts named open-weight competition from Alibaba, DeepSeek and Moonshot as the pressure behind them. For a finance leader the deciding question is where the data is processed and under which law: DeepSeek's own service processes and stores personal data in China.

If your 2027 plan carries an AI line drafted at summer prices, it is worth re-running before sign-off. The one-page model from the first piece is downloadable here, with the inputs unlocked, and it is a conversation I have directly: get in touch.

Discuss this piece on LinkedIn · Also as a LinkedIn Article · More Insights

References

All pricing pages accessed 6 October 2026; prices change on the vendors' schedules.

OpenAI Help Center, "About ChatGPT Pro tiers", help.openai.com/en/articles/9793128 - Pro 100, Pro 200 and Pro 500 at $100, $200 and $500 a month, billed monthly; the lower Pro 200 allowance for new subscriptions at an unchanged $200, "to reflect our increasingly efficient models"; the previous allowance kept through 29 October 2026 for subscriptions active at the eligibility cutoff or in the seven days before it. OpenAI, Codex documentation, "Speed", learn.chatgpt.com - Fast mode drawing included limits at 2.5 times and GPT-6 Astra Ultrafast at 8 times the standard rate. OpenAI (29 September 2026), "DevDay 2026 Recap" - Pro 500 at 25 times the Plus allowance. The Next Web (29 September 2026), "OpenAI halves Pro 200 usage and launches a $500 ChatGPT plan at DevDay", and Digital Trends (September 2026), "OpenAI wants $500 a month for ChatGPT now, and the $200 plan is paying for it" - the reported size of the cut, from 20 times to 10 times the Plus allowance, and the one-time credit. TechCrunch (9 April 2026), "ChatGPT finally offers $100/month Pro plan" - Pro 100 launched with five times the Codex usage of Plus, against 20 times on the $200 plan.
OpenAI (29 September 2026), "Introducing dots", openai.com/index/introducing-dots, and OpenAI Help Center, "Getting started with your dot", help.openai.com/en/articles/20001530 - always-on agents with their own cloud computer; the first dot included in Pro and Business Premium, with Pro excluding the UK, the EEA and Switzerland at launch; tasks in Codex and ChatGPT Work counting toward usage limits, and conversations with a dot not counting; Custom Rules to allow an action, require approval or block it; Activity View; resetting a dot; "Your dot can make mistakes, including when following your rules"; specialist dots with their own identity, credentials and access, built on early internal testing across procurement, invoice processing, email marketing, customer support and commercial contracting. ChatGPT release notes (29 September 2026) - dots usage not counted toward plan allowances for the first month.
OpenAI (22 September 2026), "Introducing GPT-6 Sol and Luna", and API pricing, developers.openai.com/api/docs/pricing - GPT-6 Sol and GPT-6.1 Sol at $2 and $10 per million input and output tokens, with cached input at $0.20 on GPT-6 Sol and $0.10 on GPT-6.1 Sol; GPT-6 Luna at $0.10 and $0.50; GPT-6 Astra at $10 and $50; GPT-5.6 Sol's promotional $4 and $20, available at least through 21 November 2026; short-context rates, with long-context and faster processing priced higher. OpenAI (29 September 2026), "GPT-6.1 Sol" - "Near-Astra intelligence for a fifth of the price". VentureBeat (22 September 2026), "OpenAI releases GPT-6 Sol and Luna models, slashing API costs 50% or more" - the permanence of the new prices. The Next Web (29 September 2026), "OpenAI releases GPT-6.1 Sol at a fifth of GPT-6 Astra's token prices".
Anthropic (22 September 2026), "Claude Opus 5.5", anthropic.com - $4 and $20 against Opus 5 at $5 and $25; "costs 40% less to run than Opus 5", on Anthropic's tests at default settings on typical workloads; cache reads, which make up the majority of agentic and coding work costs; the increase in five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans. Anthropic, pricing documentation, platform.claude.com/docs/en/about-claude/pricing - Claude Fable 5.1 at $10 and $50; fast mode for Opus 5.5 at $8 and $40; Sonnet 5.5 cache reads at $0.20 per million tokens. Anthropic, "Claude Sonnet", anthropic.com/claude/sonnet - Sonnet 5.5 at $2 and $10, "up to an estimated 30% less to run than Sonnet 5 for typical workloads billed by token"; GitHub Changelog (28 September 2026), "Claude Sonnet 5.5 in GitHub Copilot", for the release date. Anthropic support, "Claude Code May to August 2026 weekly limits promotion" - weekly limits from 14 September 2026 at 25 per cent above their level before the promotion.
Google (30 September 2026), "Gemini 4 Argon: our next era of frontier intelligence", blog.google - rolling out to a set of trusted cyber defenders through the Fairwind Program, with an initial cohort of trusted testers; to launch at an introductory $2 and $10, then $4 and $20. Google (2 September 2026), Gemini 3.8 Flash launch post - introductory pricing of $0.75 and $3.75 to 31 December 2026, then $1.50 and $7.50 from 1 January 2027.
Google, Gemini Apps Help, "Changes to Gemini model access and limits", support.google.com/gemini/answer/17004136 - model access by plan for personal accounts, from 9 October 2026 without a plan and, for AI Plus, from dates emailed to subscribers (no plan: Flash-Lite; AI Plus: Flash-Lite and Flash; AI Pro and AI Ultra: Flash-Lite, Flash and Pro); effort levels that use more of the limit; compute-based limits since 17 May 2026, refreshing every five hours up to a weekly limit. Android Authority (October 2026), "Google is downgrading Gemini for users on budget plans".
Microsoft Partner Center, "October 2026 announcements", learn.microsoft.com/partner-center/announcements/2026-october - usage-based billing on by default for new Microsoft 365 Copilot Business licences bought through CSP from 1 December 2026 (previously 2 November), default spending limit of 4,000 Copilot Credits per user per month, adjustable by admins; rolling out in supported markets, initially not in Australia, Belgium, Brazil, France, Germany, India, Italy, Korea, the Netherlands, Poland or Spain. Microsoft 365 Message Center, MC1476200, "Simplifying AI access: introducing usage-based billing" (21 September 2026, updated 1 October 2026), as republished by Modern Workspace Pro (mwpro.co.uk) and PupuWeb - the same default for new licences bought directly from Microsoft from 19 October 2026; for existing Microsoft 365 Copilot customers, a new Copilot Business seat deploys usage-based billing across all Microsoft 365 Copilot subscriptions in the tenant; a spending policy covering future eligible services; existing terms and existing spending policies unchanged. Microsoft, "Copilot Credits Guide" (October 2026 edition) - pay-as-you-go at $0.01 per Copilot Credit. Microsoft, "Microsoft 365 Copilot pricing" - Copilot Business at $21 per user per month on an annual plan; Pax8 (18 November 2025), "Your SMB AI opportunity is here: new Copilot offers that are perfect for SMBs" - the $21 list price and the $18 promotional price from 1 December 2025 to 31 December 2026. Microsoft UK, "Microsoft 365 Copilot Business", microsoft.com/en-gb - from £16.10, now from £13.80, per user per month paid yearly, before VAT.
MakeUseOf (30 September 2026), "OpenAI is giving away $2,500 worth of free credits, but only to some users" - 62,500 credits for existing Pro 200 subscribers, reported as worth $2,500 and expiring at the end of 2026, from OpenAI's email to the subscribers affected.
DeepSeek, "Models and pricing", api-docs.deepseek.com/quick_start/pricing; "DeepSeek-V4-Pro GA Release" (13 August 2026); "DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient" (10 September 2026) - peak and off-peak pricing from 16 August 2026; V4.1 Flash at $0.15 and $0.60 off-peak for uncached input and output, $0.30 and $1.20 at peak; peak hours 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Quartz (August 2026), "DeepSeek is raising AI developer access prices by up to 1,100% starting Sunday". DeepSeek, Privacy Policy - personal data collected, processed and stored in the People's Republic of China.
Reuters (5 September 2026), "Exclusive: Anthropic IPO launch shifts toward mid-October, sources say" - marketing from mid-October at the earliest and a listing days before the US midterm elections in November. Fortune (5 October 2026), "Can superintelligence ever be controlled? Sam Altman on AI safety, doomsday scenarios, and OpenAI's delayed IPO" - no OpenAI listing in 2026.
The Next Web (22 September 2026), "OpenAI cuts GPT-6 prices in half with Sol and Luna" - open-weight competition from Alibaba, DeepSeek and Moonshot as the pressure. Open-weight releases: Alibaba's Qwen, Moonshot's Kimi and Zhipu's GLM model repositories on GitHub.
Euronews (31 January 2025) - the Garante's order of 30 January 2025 and the requests for information in Ireland and Belgium.
Oftel (2001), "Effective Competition Review: Mobile" - the inclusive-minute structure of UK contract tariffs; TechCrunch (24 September 2007), "Blyk launches ad-supported mobile network", and IT Pro (2009), "Blyk killed off in the UK"; The Register (10 June 2010), "O2 to step back from unlimited mobile data deals"; Ofcom (22 July 2019), "Helping consumers to get better deals in communications markets: mobile handsets"; Competition and Markets Authority (case opened 27 July 2026), "Microsoft: consumer protection enforcement case"; GitHub Changelog (1 June 2026), "Updates to GitHub Copilot billing and plans"; OpenAI (February 2026), "Testing ads in ChatGPT", and ChatGPT release notes (4 June 2026), "Rolling out ads in the UK".
Purewal, J. (2026), The meter under the seat - the worked 400-seat model, the five fields and the contract mechanics this piece builds on; and What AI actually changes in FP&A.

The views here are my own. They do not represent the position of any current, former or future employer or client, and nothing here draws on confidential information.