May’s argument has held up better on the pressures facing AI businesses than on the crash I expected to follow them. Token prices are falling, even as bills make the cost of agent work harder to ignore. Customers have more choice about which models they use and where they run them. Features that once justified a separate product are appearing inside the main assistants. Those were central arguments in the Return on Intelligence series.
But I wrote too confidently about a crash as the turning point. Much of the change I described as belonging to the economy after a crash is already taking place. Companies can change their prices, customers can switch suppliers and power can move between businesses without one decisive moment when the whole market loses faith.
So this is a return to May’s map: what has held up, what I would change and what we still don’t know. The core question remains the same. If intelligence becomes cheaper and more widely available, who gets the benefit, and who can still make money selling it?
A follow-up to the May 2026 series, using developments available by 26 October.
Back to the convergence map
In Breaking Point, I brought together eight pressures that could become dangerous in combination: cheaper models, visible token costs, slow enterprise returns, doubts about infrastructure spending, weaker software pricing, vulnerable AI products, financing stress and a change in investor belief.
Several are now happening. Others are developing more slowly, or pulling in different directions.
01 / Revisit the convergence map
Several pressures from May are now visible
Choose a pressure to compare May's argument with what has happened since. Last checked on 26 October 2026.
Observed
The cost of agent work is more visible, even as token prices fall.
In May: Subsidised experimentation would give way to customers measuring the cost of useful work.
GitHub announced AI Credits based on token consumption, replacing premium requests for migrating plans. Annual subscribers had a separate transition. Source
OpenAI cut GPT-5.6 Luna's API prices by 80%, attributing the reduction to efficiency gains. Source
Opus 5.5 launched with input and output prices 20% below Opus 5 and cached-input prices 60% lower. Anthropic says it requires less compute to serve. Source
Vercel reports September AI Gateway spending about thirteen times December's, alongside a 47% fall in average price per token. The mix of models also changed. Source
Claude Enterprise lists a seat charge plus usage at API rates. Consumer usage limits are a different arrangement. Source
OpenAI documents shared Work/Codex allowances and credits, with token rates for credit-based usage; API billing remains separate. Source
Why it matters: Usage can grow faster than prices fall. Buyers need to compare the cost of completed work and their subscription allowances; suppliers need enough income after computing costs to fund the rest of the business.
The open question: How much do price cuts reflect cheaper computing, and how much is competition squeezing margins? The price alone does not tell us whether a supplier is losing money.
Read the whole convergence map
Model scarcity: Emerging
Customers have more credible alternatives to the leading closed models.
May: Smaller, cheaper, specialised and local models could reduce the premium for access to intelligence.
- July 2026: Kimi's K3 release expanded its agent and coding offering; its current documentation reports released weights. Source
- 10 September 2026: DeepSeek introduced V4.1-Flash with lower API prices and peak/off-peak rates. Source
- 6 October 2026: Mistral launched the Large 4 API preview, nicknamed Le Chonk, with weights promised for the end of October. Source
- 7 October announcement: Microsoft offered local-AI Surface hardware: laptop availability from 16 October, with the US RTX Spark Dev Box due to ship in November. Source
- 15 October 2026: Apple demonstrated Qwen and other open models with MLX on Mac Studio, including multi-machine inference for larger models. Source
A buyer can compare APIs with models on a laptop or dedicated infrastructure, using paid services for work that needs them. Some computing spend could move towards hardware and hosting rather than model-access fees.
How much work will customers actually move? A cheaper model still has to do the job well enough, and running it yourself requires suitable hardware.
Token economics: Observed
The cost of agent work is more visible, even as token prices fall.
May: Subsidised experimentation would give way to customers measuring the cost of useful work.
- 27 April → 1 June 2026: GitHub announced AI Credits based on token consumption, replacing premium requests for migrating plans. Annual subscribers had a separate transition. Source
- 30 July 2026: OpenAI cut GPT-5.6 Luna's API prices by 80%, attributing the reduction to efficiency gains. Source
- 22 September 2026: Opus 5.5 launched with input and output prices 20% below Opus 5 and cached-input prices 60% lower. Anthropic says it requires less compute to serve. Source
- 15 October 2026: Vercel reports September AI Gateway spending about thirteen times December's, alongside a 47% fall in average price per token. The mix of models also changed. Source
- 26 October pricing snapshot: Claude Enterprise lists a seat charge plus usage at API rates. Consumer usage limits are a different arrangement. Source
- 26 October pricing snapshot: OpenAI documents shared Work/Codex allowances and credits, with token rates for credit-based usage; API billing remains separate. Source
Usage can grow faster than prices fall. Buyers need to compare the cost of completed work and their subscription allowances; suppliers need enough income after computing costs to fund the rest of the business.
How much do price cuts reflect cheaper computing, and how much is competition squeezing margins? The price alone does not tell us whether a supplier is losing money.
Enterprise returns: Mixed
The gap between using AI and achieving major business gains is measurable.
May: Integration, permissions, data quality and organisational change could delay enterprise returns.
- 19 October 2026: Gartner reports that 13% of surveyed CIOs see significant or transformational AI value. The survey covered 2,002 CIOs and technology executives in May–July. Source
The revenue that suppliers expect from widespread transformation may arrive later than the investment made to serve it.
How quickly will smaller benefits turn into lasting business gains? The other survey respondents may already be getting useful, less dramatic results.
Infrastructure spending: Observed
Infrastructure spending is being questioned as well as celebrated.
May: The same capital spending could change from proof of demand into concern about mistimed supply.
- 23 October 2026: Reuters describes investor concern over the financing of AI spending and scrutiny of future buildout plans. Source
- April 2026 baseline: Amazon's shareholder letter described unserved demand and capacity constraints, a counterpoint to the overbuild argument. Source
- 21 October 2026: Dell reports more than 6,500 AI Factory customers, up from 5,000 the previous quarter. Source
Investors are asking when capacity will earn enough to justify its cost. Local and private inference also change which provider might receive the revenue, while shifting some investment risk to customers.
Will future capacity attract enough paying work? Today's shortages can coexist with too much investment in tomorrow's supply.
SaaS pricing power: Mixed
Software is easier to challenge, but established platforms can still grow.
May: Generated software, fewer seats and bundled AI could weaken generic workflow subscriptions.
- October 2026: ArtCraft's applications demonstrate an attempt to reproduce established creative workflows; PhotoCraft remains early alpha. Source
- 21 October 2026: SAP reported stronger cloud revenue and backlog as customers continued to buy its services. Source
- 28 September 2026: Microsoft researchers and collaborators describe software-factory work across tens of repositories, focused on improving data systems. Source
The threat depends on what the customer pays for: a replaceable feature, or records, integration and a dependable service.
Will customers replace subscriptions, or keep paying for reliable service and integration? Building an alternative is only the beginning.
Standalone AI products: Emerging
ChatGPT and Gemini now include work that once supported separate AI tools.
May: Narrow AI products could fail as platforms, incumbents and internal builds absorb their functions.
- July 2026: OpenAI's Work documentation describes research, files, apps and finished documents, spreadsheets, presentations and Sites inside its broader product. Source
- 19 August 2026: Google announced Deep Research in Gemini Live for all Gemini app users, subject to availability and usage limits. Source
- 20 October 2026: Google announced Gemini business tools for brand assets, connected apps and financial insights. Source
A separate product must offer enough additional value to justify another purchase when a similar function is already available in the customer's main assistant.
Which specialists offer enough extra value to keep their customers? Bundled features raise the pressure, but quality, trust and integration can still justify another purchase.
Financing stress: Unconfirmed
There is more scrutiny of the financing, but the stress May anticipated remains an open risk.
May: Leases, debt and guarantees could transmit trouble before the technology itself disappoints.
- 26 October 2026: Anthropic's IPO filing warns about capacity payments regardless of usage, alongside improved operating results. Source
- October 2026: Brace for Impact maps financial connections and possible transmission of losses beyond AI suppliers. Source
- 26 October 2026: Reuters links market unease to the capital needed for AI investment. Source
Useful products do not remove payment deadlines. Some businesses may need to refinance before customer income covers their commitments.
Who has to pay if a customer or project runs out of money? That depends on the borrowing, leases and guarantees behind each deal.
Belief stress: Emerging
Investors are questioning AI spending despite continuing enthusiasm for AI.
May: Investors would start asking who earns the profit rather than rewarding AI exposure alone.
- 16 October 2026: The Ezra Klein Show brought the bubble argument and a more optimistic case into a direct debate. Source
- 26 October 2026: Reuters reports selling in technology stocks and concern over the eventual payoff from AI spending. Source
The change in the story has begun: spending more is no longer automatically reassuring. That is evidence of strain in belief, even while AI remains useful.
Will the doubts last, or will optimism return? A selloff can reverse, and AI is only one of the forces moving markets.
What May got right
The bill matters more when agents do more
May argued that enthusiasm about usage would eventually meet a harder question: does the work justify its cost? Copilot’s move to token-based AI Credits makes that question harder to avoid. GitHub announced the change just before the series, on 27 April, and began the transition on 1 June. Annual subscribers have separate transition terms.
Under the old premium-request system, a quick question and a lengthy agent session could consume the same allowance. The new approach charges according to the tokens used. GitHub said it could no longer keep absorbing the growing cost of agent work.
Claude Enterprise also separates the seat charge from usage, while OpenAI’s Work and Codex plans combine allowances and credits. The arrangements differ, but the purchasing decision is becoming familiar: what work can this budget actually buy?
The subscription and the model price can move differently. OpenAI’s new Pro 200 terms reduce the included allowance while keeping its monthly price at $200, even as its newer models become cheaper to use. A lower API price does not guarantee a more generous subscription.
An agent asked to fix a bug might finish quickly, or spend much longer reading files, trying changes and rerunning tests. Someone still has to check the result. The useful comparison is the cost of getting the bug fixed, including retries and review, rather than the price of the first answer. GitHub’s own engineering work describes why saving tokens in one step can cost more later if the agent then has to recover missing information.
Cheaper tokens can therefore sit alongside a larger monthly bill. That is not necessarily a bad deal. It is a deal customers will have to judge against the work completed.
Intelligence is becoming easier to source
May’s power map explicitly included local, private, open and specialised models. That part of the argument looks stronger.
Kimi’s K3, DeepSeek’s newer releases and Mistral Large 4, or “Le Chonk”, widen the choice. They are at different stages: Mistral’s October release is an API preview, with downloadable weights still promised. A customer does not need every alternative to beat the best model at everything. It may be enough to use a cheaper one for routine work and reserve the expensive model for difficult tasks.
Local AI adds another option. Qwen models and tools such as LM Studio let developers try that work on their own machines. Apple’s unified memory lets the CPU and GPU share the same pool, helping larger models fit. Its M5 Pro and Max hardware predates May; October’s MLX work shows further progress, including larger models spread across Mac Studios. The size and memory of the machine still matter.
Microsoft’s new Surface hardware follows the same direction. Surface Laptop Ultra became available in October; the US Surface RTX Spark Dev Box is due to ship in November. The appeal is straightforward: buy a machine and run suitable work on it, instead of paying a remote provider for every step.
That replaces some API costs with hardware, electricity and maintenance. It also leaves the owner paying for a machine when it is idle. A mixture of local and cloud models may be the better fit. Even Apple Intelligence combines on-device processing with cloud services, including Google Cloud infrastructure. Running an assistant on your laptop does not mean all its work stays there.
Enterprises have more options than buying another cloud service
The same choice is reaching larger organisations. They can own GPU servers in a colocation facility, paying another company for the space, power and cooling. Or they can rent dedicated capacity from a specialist host.
This market existed before May. The recent change is the expanding range of offerings. Equinix’s September Inference Exchange is in preview, while HPE’s October announcements add options for running and managing private AI. Dell reports more than 6,500 AI Factory customers, up from 5,000 the previous quarter.
For a company with enough recurring work, dedicated equipment may make sense. For one with occasional bursts of demand, it may be an expensive way to keep GPUs idle. Control over sensitive data can also justify a choice that is not the cheapest.
The consequence for the map is that growing AI use need not send all the revenue to the largest clouds. Some can go to equipment suppliers, specialist hosts and operators of the facilities. Some of the financial risk moves to the customers buying the machines.
A useful feature is harder to defend as a whole company
May also argued that a thin layer around a model would be vulnerable when the main platforms added the same function.
OpenAI’s Work tools now cover research, documents, spreadsheets, presentations and websites. Google’s Gemini announcements extend research and small-business tools. Before buying a separate product, a customer can try what is already in the assistant they use.
Specialists can still win on quality, privacy, integration or service. They need a better reason to exist than having been first to put a useful model behind a convenient interface.
Software factories create another route to competition. They link specifications, coding, testing, review and operation into a repeatable process. A “lights-on” version keeps people approving changes. A “dark” factory attempts more of the process without human code review.
A September paper from Microsoft and collaborators describes work across tens of repositories, focused on improving data systems. An October study of review agents also shows why checking the result remains difficult, particularly for complex changes and private company code.
Repeated agent work could make dedicated computing worth considering. But the two choices are separate: a company can run supervised agents on its own GPUs, or use cloud APIs in a highly automated factory.
ArtCraft’s attempts to reproduce Adobe workflows show why cheaper construction matters too. PhotoCraft is still early alpha, so it has a long way to go before professionals can depend on it. Yet an alternative can put pressure on prices well before it replaces every feature.
What I would change
I was too certain about the crash
May’s language often treated a crash as the gateway to the mature AI economy. I would change that.
The pressures I described can lead to several outcomes: lower prices, slower growth, mergers, changes in billing, some business failures and gains passed on to customers. They can arrive at different speeds. A company can lose its advantage without the entire market breaking at once.
This does not remove the risk of a sharp correction. It changes how much weight I would put on that one outcome. Seeing several pressures develop is a reason to revisit the valuations, not a reason to assume the ending has already been decided.
Low prices need not be a temporary subsidy
May anticipated cheaper inference, but also asked what would happen when usage was “priced honestly”. I would be more careful with that phrase. A low price does not, by itself, show that a supplier is selling at a loss.
Max Woolf’s October account of the price cuts makes this worth revisiting. OpenAI cut GPT-5.6 Luna’s API prices by 80% in July. In September, Anthropic launched Opus 5.5 with input and output prices 20% below Opus 5, and cached-input prices 60% lower. Caching lets an agent reuse material it has already processed, such as files and conversation history. That discount can matter more than the headline price on a long job.
Anthropic says the newer model requires less computing power to serve. Competition can push prices down while efficiency leaves room for a supplier to earn money on each task. That still has to pay for training, staff and the rest of the business.
Nor do cheaper tokens guarantee a smaller market. Vercel reports that September spending through its AI Gateway was about thirteen times December’s, while the average price per token fell 47%. This is one gateway’s traffic, and the average changes as customers switch models. Even so, it shows why falling prices and growing revenue can happen together.
Suppliers can improve their economics as they grow
Anthropic’s IPO filing, published on 26 October, gives another reason to resist a predetermined ending. It reported about $11.6 billion in revenue and $1 billion in operating profit for the second quarter. Competition and expensive computing have not prevented it from improving its operating results.
It also reports $265.6 billion of purchase commitments and $321.4 billion of leases not yet started at the end of June, with some lease arrangements still conditional. These are payments spread over future years. The question is whether customer income will grow fast enough to pay those bills while funding the next generation of models.
Customers can change their spending faster than Anthropic can change its bills. Much of its revenue depends on usage, while its computing agreements run for years. Cheaper open models and private hosting give customers more alternatives. Demand for AI could keep growing even as competition leaves suppliers with less money to meet their commitments.
Easier model choice does not automatically mean less concentrated power
The May map recognised that devices, permissions, data and customer relationships would matter. I would now give more attention to the companies trying to bring those advantages together.
Revolut’s AIR, which began its UK rollout before May, starts inside a banking app that already holds financial records and provides card controls. Grok Bot’s August launch offers agents that work across connected apps using a cloud computer. Apple’s assistant starts from the operating system. Each has a different route to becoming the place where the user asks for things to happen.
That position could be valuable even if the models underneath become easier to replace. The assistant can influence which service gets considered, which information it receives and which model does the work. Moving your model is also easier than moving an assistant’s accumulated memory and connections.
Nor are local hardware and cloud services necessarily opposing businesses. Microsoft sells both. Apple puts an interface on the device while using cloud infrastructure for some requests. More places to run AI can still leave a small number of companies controlling how people reach it.
Software businesses will change at different speeds
May distinguished replaceable workflow software from systems of record such as finance, HR and ERP. That remains a useful way to read the software market.
SAP’s October cloud growth is a reminder of how much customers still pay for dependable business systems. Replacing a feature is one thing. Moving company records, integrations and responsibility for keeping the business running is another.
I would be more cautious about the speed of the transition. Generating an application can become cheap long before maintaining it becomes easy. The same applies to software factories: producing more code does not remove the need to decide what to build, check it and support it.
What remains unresolved
Will the spending earn enough?
October’s market reporting shows investors questioning the financing and eventual payoff of AI spending. The selling reported on 26 October is one sign of that unease. It is still too early to call it a lasting change in belief.
“Overbuild” is also more complicated than empty data centres today. A provider can be short of capacity now and still have committed too much money to future demand. Amazon’s April shareholder letter described shortages. The question is whether the capacity being financed will attract enough paying work at prices that cover its costs.
Local and private AI make that harder to predict. Demand for intelligence can grow while a particular provider gets a smaller share of it. The chips and electricity are still needed, but the income may turn up somewhere other than the business that made the largest commitment.
How quickly will customers turn AI into business gains?
Gartner’s October survey found that 13% of the CIOs and technology executives surveyed reported significant or transformational AI value. Others may be getting smaller benefits. The finding points to the distance between trying the technology and changing how a business operates.
That delay matters to companies investing in expectation of future customers. It also matters to those customers when agent bills arrive before the hoped-for savings. The software can improve faster than an organisation’s data, permissions and working practices.
Who carries the losses if income arrives late?
The report Brace for Impact follows the financing behind the buildout: debt, leases, guarantees and private credit. Its central concern is how trouble at AI companies could spread to lenders, investors and people who never chose to make an AI bet.
A useful technology does not suspend payment deadlines. A business can have a promising future and still run out of money before it gets there. The size of that risk depends on who owes what, when payments fall due and who has agreed to cover a failure.
OpenAI’s financial picture is harder to inspect. Its June announcement confirmed a confidential IPO submission, without publishing the accounts. Its partnership announcements give a partial view: the February Amazon deal expanded its AWS agreement by $100 billion over eight years, alongside Amazon’s planned $50 billion investment.
An investor can also be a supplier and a sales channel. That makes it important to ask how much of the growth is ultimately paid for by customers outside those relationships.
02 / Follow the money
One technology, several financial outcomes
Choose a scenario to see how the benefits, costs and risks could spread.
Models become cheaper and customers automate more work.
- 01
Customers
Lower costs can widen adoption and increase the work attempted.
- 02
Model suppliers
Lower prices need enough extra paying customers to cover the cost of running and improving the models.
- 03
Infrastructure owners
Busy machines need paying customers if their owners are to cover equipment, power and operating costs.
- 04
Capital providers
Investors and lenders need the business to generate enough money to pay them.
Evidence to watch: The cost of finished work, repeat customers and suppliers' profit margins.
Demand could expand enough to offset lower prices. Cheaper models alone do not establish a crisis.
Read the alternative paths
Revenue arrives late
Customers adopt AI, but their payments fall short of the bills coming due.
- Model suppliers: Long-term capacity bills can fall due before customer revenue catches up.
- Cloud and project companies: A late payment, renegotiated lease or non-renewal can weaken expected project cash flows.
- Lenders and guarantors: Who loses money depends on the security behind the loans and who promised to cover missed payments.
- Households and communities: Pensions, insurance, employment and local infrastructure can become channels of harm.
The outcome depends on how the projects are financed and which businesses run into trouble.
The buildout earns its keep
Customers renew because measured gains exceed the full cost of adoption.
- Operating businesses: Reliable outcomes support repeat spending and free resources for new work.
- Suppliers: Healthy contracts fund serving costs and continued product development.
- Infrastructure owners: Enough paying work keeps the machines busy and covers borrowing and new investment.
- Market structure: Value may still concentrate around distribution, records, identity and access to capital.
A productive boom can contain overpriced firms and failed projects. Growing demand does not make every company a good investment.
The debate between Paul Kedrosky and Azeem Azhar on Ezra Klein’s show captures the disagreement about whether the spending can be justified. Ed Zitron’s writing presses on the dependence on continued investment, while Azhar’s earlier argument gives more weight to productive capacity and the quality of its financing.
One useful discipline from Zitron’s discussion of annualised revenue is simply to ask what the number measures. A run rate extrapolated from recent sales is different from money already earned over a year. Neither an impressive model nor an impressive revenue headline answers the whole business question.
Where that leaves the map
I would keep the central idea from May: cheaper intelligence changes what is worth owning. I would put less certainty around the sequence of events and more attention on the businesses adapting across several parts of the market at once.
An agent that can offer a refund still needs the right customer record and permission to issue it. Microsoft’s approach to agent security puts controls around those actions. The engineering around coding agents similarly matters because a capable model still needs tools, feedback and checks to get useful work done.
Those can be lasting advantages. So can the customer relationship, especially if the main assistant becomes the route through which other services are bought and used.
03 / Who keeps the value?
What moved, and what needs a second look?
Choose a layer to see what I would keep from May and what I would change.
Production and verification
In May
Orchestration, evaluation and execution were grouped together.
Where I stand now
Software factories join planning, coding, testing and operation. Hosting choice and removal of human review are separate decisions; verification remains a constraint.
Researchers describe working software factories, but still find problems with automated review of complex company code. Read the evidence
Read the whole ownership map
Compute and energy
May: Chips, memory and energy remain essential.
Revision: Compute can sit on a laptop, in customer-owned servers at a colocation site, with a specialist host or in a hyperscaler. Ownership changes who bears utilisation and financing risk.
Models
May: Frontier, local, open and specialised intelligence.
Revision: Route work across local open models, private inference and paid APIs by quality, cost and privacy. Model choice and choice of personal assistant are separate decisions.
Identity and authority
May: Permissions and governance form a layer.
Revision: Every agent needs permission: whose account is it using, what can it access and which actions need approval?
Records and context
May: Trusted records, knowledge of the business and customer relationships matter.
Revision: I would keep this part of May's map. A new model still needs reliable records and an understanding of how the business works.
Production and verification
May: Orchestration, evaluation and execution were grouped together.
Revision: Software factories join planning, coding, testing and operation. Hosting choice and removal of human review are separate decisions; verification remains a constraint.
Distribution and personal agents
May: Distribution and ownership of the workflow were possible sources of durable power.
Revision: The assistant can become the customer's starting point, carrying memory and permissions across services. More model choice can coexist with dependence on one intermediary.
Products and organisations
May: Generated workflows and AI-operated companies.
Revision: An emerging operating model whose economics must include review, integration, maintenance and responsibility.
The questions I would keep asking are practical. Are customers getting enough useful work for the money? Can suppliers cover their costs? Can people change models and assistants without losing their records and working arrangements? And can the businesses financing the infrastructure afford to wait?
May was right to separate the success of AI from the fortunes of the companies selling it. I would keep that argument, and let go of the certainty that it needs a crash to become visible.
Sources
Links and publication dates. Financial disclosures and pricing checked on 27 October 2026.
- Max Woolf, 25 October 2026: OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?.
- OpenAI, 30 July 2026: GPT-5.6 efficiency and price reductions.
- Anthropic, 22 September 2026: Opus 5.5 pricing and serving efficiency.
- OpenAI, terms checked 27 October 2026: Pro allowances and transition dates.
- Vercel, 15 October 2026: AI Gateway Production Index.
- OpenAI, 8 June 2026: Confidential submission of draft S-1.
- OpenAI and Amazon, 27 February 2026: Partnership and investment announcement.
- Anthropic, 26 October 2026: IPO filing.
- Qwen and LM Studio: Qwen3.8-27B model card and LM Studio’s local packages.
- Apple: M5 Pro and Max MacBook Pro, 3 March 2026 and MLX local and distributed inference, 15 October.
- Microsoft, 7 October 2026: Surface Laptop Ultra and Surface RTX Spark Dev Box.
- Apple, 8 June 2026: Foundation-model architecture and Private Cloud Compute expansion.
- Equinix, 2 September 2026: Inference Exchange announcement and preview status. See also its 2024 private DGX offering.
- Dell, 21 October 2026: Enterprise agentic AI and AI Factory customer count.
- HPE, 21 October 2026: Private Cloud AI and agent infrastructure updates.
- Pavlenko and collaborators, 28 September 2026: Towards an AI Software Factory for Data Systems.
- Yang and collaborators, 9 October 2026: (When) Should a Software Factory be Lights-off?.
- Revolut, 9 April 2026: AIR launch.
- Grok Bot, 11 August 2026: Launch and access conditions.
- GitHub, 27 April 2026: Moving to usage-based billing.
- Anthropic: Current pricing.
- OpenAI: Work and Codex pricing.
- Moonshot/Kimi: Dated announcement index. See also K3 documentation and K3.1 Pro.
- DeepSeek, 10 September 2026: V4.1-Flash announcement. See the current price card.
- Mistral, 6 October 2026: Mistral Large 4 / Le Chonk.
- OpenAI, July 2026 release entries: What’s new.
- Google, 19 August 2026: Gemini learning and research tools.
- Google, 20 October 2026: Gemini tools for small businesses.
- Gartner, 19 October 2026: Survey on realised AI value.
- Reuters, 23 October 2026: Wall Street closes lower to end rocky week.
- Reuters, 26 October 2026: AI concerns and market selling.
- Amazon, 9 April 2026: Jassy’s 2025 shareholder letter.
- Reuters, 21 October 2026: SAP’s cloud growth and outlook.
- The May baseline: Breaking Point and The New Power Map.
- PhotoCraft: README at the inspected commit.
- ArtCraft coverage: Lumien, 8 October 2026.
- OpenAI, 11 February 2026: Harness engineering.
- GitHub, 2 September 2026: How we make AI coding more cost efficient without sacrificing task quality.
- Microsoft, 19 October 2026: Agents are your newest insiders.
- The Ezra Klein Show, 16 October 2026: Is There an A.I. Bubble? A Debate.. Episode listing and publisher summary.
- Azeem Azhar, 18 October 2025: The imminence of the AI bust.
- Ed Zitron, 20 October 2026: AI Had Better Work Out.
- Ed Zitron, 13 October 2026: Annualized Enron Rate.
- Scherer and Jenkins, October 2026: Brace for Impact.
Return to the series preface.