Three announcements in July were read as one piece of news: AI was getting cheaper. On the 7th, Accenture Edge and Google Cloud launched a line of pre-configured agents for mid-sized companies. On the 21st, Microsoft signed a multi-billion dollar commitment funding Mistral's compute infrastructure in Europe. On the 24th, Anthropic shipped Claude Opus 5 at the previous generation's price. By mid-September, public price sheets show increases on some models, cuts on others and, at a single provider, two billing changes in one month.
AI adoption in the enterprise is the organisational work that turns technical access to artificial intelligence tools into real professional use: scoped use cases, integrated company data, access governance and change management measured over time, work whose cost answers to its own calendar and staffing constraints.
Model prices no longer move in one direction
The 2024 and 2025 pattern fitted in a sentence: at equal capability, the price fell year over year. The Token Price Index published by BenchLM, which tracks 40 models across three segments, no longer shows a uniform decline. In its 15 September 2026 update (methodology v1.1), the frontier segment is up 72% year over year and the budget segment up 7.4%, while mid-tier is down 7.7%: over twelve months, the top and bottom of the market got dearer while the middle got cheaper.
What Anthropic's price list shows
Anthropic's catalogue shows no per-token price increase across its summer releases. Claude Opus 5, released on 24 July 2026, holds the Opus 4.8 price: 5 dollars per million input tokens, 25 on output. Claude Fable 5.1, released on 1 September 2026, keeps Claude Fable 5's 10 and 50 dollars unchanged, and cuts cache reads to a quarter of the previous rate. On 10 August 2026, Anthropic cancelled the Claude Sonnet 5 increase it had announced for 1 September and made the introductory rate the standard one, 2 dollars on input and 10 on output (Anthropic public price list, 10 September 2026, unchanged on 16 September).
BenchLM's 15 September update carries those rates as they are, with Claude Fable 5.1 entering the frontier segment at 20.00 dollars per million blended and Claude Sonnet 5 at 4.00 dollars in mid-tier. Each model counts at a blended rate (three quarters input price, one quarter output price), and a segment also moves when a model enters or leaves the index. A high-end model joining can therefore push the frontier segment up without any existing rate rising, which is why purchasing decisions are better taken from the vendor's own price list.
Launch promotions expiring at the entry level
Outside the basket BenchLM tracks, the entry level carries a different risk: some listed rates are launch promotions, and they expire. Z.ai's promotional pricing for GLM-5.3-Flash ended on 9 September 2026 and the list rate doubled (Z.ai public price list, 10 September 2026). Upstage's discount on Solar Pro 4 steps down in stages: 90% until 11 September 2026, 70% from then until 10 October, and full price after that, at 0.30 dollar per million input tokens and 1.20 on output (Upstage public price list, checked on 16 September 2026). Between the first and second step, the same volume therefore billed three times higher, and the last step in October will multiply the bill by just over three again.
DeepSeek and billing rules that change mid-month
DeepSeek illustrates the other form of instability, with no promotion involved. On 16 August 2026 the provider raised its V4 price sheet, off-peak hours included: on output, V4-Pro went from 0.87 dollar per million tokens to 1.98 off-peak and 3.96 at peak, V4-Flash from 0.28 to 0.66 and 1.32 (Engadget, 14 August, TechTimes, 17 August 2026, DeepSeek public price list checked on 16 September 2026). Peak hours, billed at twice the off-peak rate, run from 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, which in Paris summer time means 03:00 to 06:00 and 08:00 to noon. On 10 September 2026 a new V4.1 Flash price sheet took effect, at 0.15 dollar per million input tokens and 0.60 on output, off-peak. Its documentation then announced that, from 14 September, calls to V4-Pro would be routed to V4.1 Flash and billed at the Flash rate. Checked on 16 September 2026, the same page says the opposite: in response to user demand, V4-Pro stays available after 14 September, with billing unchanged. In one month, the catalogue changed its billing rules twice and dropped a third change it had already announced.
For a finance function, the token price has become an unstable input: it climbs at the top, climbs back at the entry level the day a promotion lapses, and a single catalogue's billing rules can change several times in a month.
Price cuts arrive without notice too
Price cuts have not gone away either, and they arrive with the same notice, which is to say none. On 30 July 2026, OpenAI cut the price of Luna, the cheapest model in its GPT-5.6 range, by 80% (OpenAI announcement, 30 July 2026). The cancelled Sonnet 5 increase is the same mechanism running the other way. What changed since 2025 is predictability: a rate can now shift either way, on a specific model identifier, between two budget cycles.
July's announcements on models and pre-configured agents
July's two other announcements complete the picture. According to Microsoft's 21 July 2026 press release, Mistral Medium 3.5 and Mistral OCR 4 were available in Foundry from that day, and Medium 3.5 in Copilot Studio, which widens the choice of models available from a single cloud provider. On 7 July 2026, Accenture Edge and Google Cloud aimed their pre-configured agents at companies with revenues between 300 million and 3 billion dollars.
How an architecture absorbs a price change
A price cut on an existing model identifier flows through to the invoice by itself, without a line of code changing. An expiring promotion, time-of-day pricing or a more capable model released under a new name all call for a decision, and what that decision costs depends mostly on how the application was built. Switching models can happen without switching vendors: at Anthropic, for instance, this summer's releases leave a choice between Opus 4.8 and Opus 5 at the same price, or between Sonnet 5 and Fable 5.1 depending on how hard the task is.
Measuring cost per completed task
The price per million tokens says little about what a processed case costs. A model that is cheaper per token can use more of them to finish the same task, because it reasons at greater length or makes more attempts and tool calls. The way text is split into tokens also varies between generations: Anthropic's pricing documentation, checked on 16 September 2026, states that its models from the 4.7 generation onward use a newer tokenizer that produces more tokens for the same text. The unit a finance team needs is cost per task: a reconciled invoice or a closed ticket, measured on real traffic.
Keeping the model identifier in configuration
When the model identifier lives in a single configuration file, moving off an expiring promotional rate means replaying an evaluation set built from real cases, then changing one line. When it is scattered through business code, the same switch becomes a project that needs its own budget, and it waits for the next budget cycle while the invoice keeps running. Our analysis of AI model lock-in covers that abstraction layer in detail and how to spot tight coupling.
Using the billing modes vendors document
A public price sheet has more lines than input and output rates. Anthropic charges less for cache reads and for requests processed in batches (pricing documentation checked on 16 September 2026), and DeepSeek bills its off-peak hours at half its peak rates. Those mechanisms reward an architecture designed around them: stable instructions and reference documents placed at the start of each request, where the cache can reuse them from one call to the next, and non-urgent jobs grouped to run in batches or in the cheapest time slots. A bank reconciliation agent working on the previous day's statements can, for example, be scheduled into a discounted slot. These savings are decided at the architecture stage, and they are hard to recover once the agent is in production.
Why don't model prices decide AI adoption?
Because the price of a model and a company's cost of adoption rest on unrelated bases. The first is negotiated in dollars per million tokens and now moves both ways. The second is measured in redefined roles and teams trained for professional use.
What the research on AI pilots measures
S&P Global (2025) finds that 46% of AI proofs of concept are abandoned before reaching production. MIT (2025) puts at 95% the share of generative AI pilots with no measurable impact on the P&L. The 2026 edition of Deloitte's State of AI in the Enterprise report finds that only 21% of companies have mature governance of their agents. Gartner (2025) expects more than 40% of agentic AI projects to be cancelled by the end of 2027. Gartner attributes those cancellations to escalating costs, unclear business value or inadequate risk controls (press release, 25 June 2025). Those costs cover an agent project as a whole, of which the per-token rate is only one component, and the MIT and Deloitte measures relate to how work is organised.
The organisational workstreams behind adoption cost
- Redefining roles: deciding what an agent handles alone, what a human validates, what stays off-limits.
- Integrating company data: connecting the systems used in production.
- Training teams for professional use: moving from an isolated personal prompt to a scoped and repeatable use case.
- Installing governance: monitoring agents continuously once they are running.
These four workstreams take the same amount of leadership, IT and business-team time, whatever the list price of the underlying model.
Managing model purchasing when price sheets move
Model spend is now best handled as a variable cost line, much like energy or cloud, with a named owner and a register reviewed on a fixed schedule. Procurement, IT and finance should work from the same register, otherwise each of them finds out about a price change when the invoice lands.
Keeping a register of models and their deadlines
For each use case in production, the register records the model identifier being called, the price sheet in force and its date, when any promotion ends, the deprecation date announced by the vendor, and who decides on a switch. These dates are almost always public in advance: Upstage displayed its discount steps in its September price sheet, and DeepSeek had put a date on its routing change. Someone still has to catch them in time, by rereading the price sheets of the models in production every month.
Budgeting at list price and tracking cost per task
An AI budget is built on the catalogue price without promotions, and any running discount is treated as a temporary gain. Tracking then happens per use case, with an alert threshold on cost per task: a drift points to a price sheet that changed or an agent that consumes more, and both deserve a review. Claude Cockpit, our service for steering agents in production, tracks usage and cost for each agent alongside its performance indicators.
The most common misreadings
- Comparing prices per million tokens without relating them to a measured task.
- Reading a move in an aggregated index as a price increase without checking whether it comes from a rate change or from a new model entering the basket.
- Mistaking a promotional rate for the normal price, then discovering the gap on the expiry date.
- Leaving model switches without a named owner, so the choice made at prototype stage is never reopened.
The regulatory calendar has already moved
On 20 July 2026, France's data protection authority, the CNIL, published a joint exploratory note with the French AI and Digital Council on agentic AI, systems that act on a user's behalf. On 2 August 2026, the transparency obligations under Article 50 of the EU AI Act became applicable, with a single deferral set by the AI omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026: providers of generative AI systems placed on the market before 2 August have until 2 December 2026 for the content marking required by Article 50(2) (European Commission FAQ checked on 16 September 2026).
These obligations fall under Article 99(4), which caps fines at 15 million euros or 3% of global annual turnover. The 35 million euro or 7% ceiling, widely quoted in the press, sits in Article 99(3) and covers the prohibited practices of Article 5, a separate regime.
A company deploying cheaper agents without governance in place accumulates the same risk with more agents in circulation. In our view, the model register also serves Article 50, which assumes you know which agents interact with users. The cost of governing was untouched by any of the summer's price moves.
What should a company do now?
When a technology gets cheaper, the reflex is budgetary: cut the envelope and treat the AI question as settled once the tool is installed. September's market makes that reflex expensive. An AI agent in production sized on promotional pricing becomes an unpredictable line item the day the promotion lapses, and it remains, without proper scoping, a system nobody supervises.
The budget freed up by a cheaper model is worth more elsewhere: on mapping use cases and on governing agent access. Those items have the advantage of being predictable. Their cost can be planned over a financial year, which no model price sheet still allows.
Before adding one more agent to the stack, a company benefits from knowing where it stands on these four workstreams. Koneetiv's AI maturity assessment places it on six dimensions in a few minutes, including technical foundations and governance, the two that decide whether an agent deployment holds up when the model price changes direction.