What moved at the frontier, and what it means in production
The Frontier log is Applicat AI's running record of what changes for organisations running AI in production: model releases, pricing shifts, the moves the AI labs and cloud providers are making into deployment services, and the research that alters how we work. Each entry says what happened, and what we think it means for a system that is already live. We add to it as the frontier moves.
15 entries · last updated 22 September 2026
The current read
The upgrade is the risk now.
For most of this quarter the thing that moved was the contract: retention made a condition of use, training consent buried in a model string, published dates for price rises. In the week of 22 September the interface moved as well. Anthropic cut the price of its frontier model by a fifth and, in the same release, made three call shapes that worked the day before return an error. OpenAI put two cheaper tiers under its own frontier model, a hundredfold apart on input, and priced the entire request higher once it passes 272,000 input tokens. Neither change shows up in a benchmark. Both show up in a system that was working on Friday. The discipline this asks for is dull, and it is most of the job: pin your model strings, run your own cases before a new model reaches a user, and watch the bill as closely as the scores.
- PricingModels
OpenAI splits its frontier into three tiers a hundredfold apart, with a cliff at 272,000 tokens
What happened
GPT-6 Sol and GPT-6 Luna arrived on 22 September beneath GPT-6 Astra. All three carry the same 1.05 million token context window and 128,000 token maximum output. Astra is $10 per million input tokens and $50 on output, Sol is $2 and $10, and Luna is $0.10 and $0.50. Above 272,000 input tokens the whole request is billed at doubled input and cache rates and one and a half times output, not only the tokens past the threshold.
Source: OpenAI API pricingWhat it means in production
Two decisions that used to be architectural are now commercial. The first is which tier a call belongs on. A hundredfold spread on input makes routing by task worth real money, and the only way to learn which of your tasks survive on the cheapest tier is to run them there, on your own cases, and look at what breaks. The second is the cliff. A retrieval step that grows from 250,000 tokens to 280,000 does not cost twelve per cent more, it roughly doubles that request, so a system that quietly retrieves more as your corpus grows gets expensive without anyone changing a line of code. Cap what you send, alert on requests that cross the threshold, and let an evaluation pick the tier.
- ModelsOperations
Anthropic cuts the price of its frontier model and breaks three things in the same release
What happened
Claude Opus 5.5 shipped on 22 September at $4 per million input tokens and $20 on output, down from $5 and $25 for Opus 5, with a one million token context window, 128,000 maximum output tokens and adaptive thinking always on. Three call shapes that worked on Opus 5 now return a 400 on the new model. Thinking cannot be disabled. A tool_choice of any or tool is refused, in favour of auto with strict tool use. Computer use on the Claude API and on Google Cloud requires the computer_toolset_20260801 toolset, and the older computer_20251124 tool is rejected, though it keeps working on Amazon Bedrock.
Source: Anthropic, Claude Platform release notesWhat it means in production
A price cut is the easiest upgrade to wave through, and this one is not a drop-in. An agent that disables thinking to keep a step deterministic, or forces a tool call to guarantee a structured reply, stops working on the cheaper model, and it fails with an error rather than a worse answer, so it surfaces in production and not in an evaluation anyone ran for fun. The same release also behaves differently depending on where you run it, so a team on Bedrock and a team on the Claude API are not doing the same migration. Pin the model string you are live on, read the release notes before the benchmarks, and treat a cheaper model as a reason for more testing.
- MarketGovernanceIndependence
OpenAI ships a legal configuration of its top model, and you have to qualify for it
What happened
Astra for Law was announced on 17 September: GPT-6 Astra combined with a legal search index covering more than 230 million URLs of United States case law, statutes, regulations and administrative decisions, plus custom instructions for legal analysis and writing. It is offered to selected law firms through a Trusted Access programme in ChatGPT and Codex, with API access said to follow. Eligible firms get zero data retention on the API, and their ChatGPT Enterprise usage is excluded from human review by default. No price is published.
Source: OpenAIWhat it means in production
This is a lab moving past deployment services and into the vertical itself, bringing its own index and its own retention terms. For a buyer it repeats the pattern this log has tracked all quarter: the strongest capability is reached through a programme you are admitted to, and the data protections travel with the admission. If you work in a regulated sector, the question to put to any supplier is which of your protections depend on a programme you could be removed from. Note also what the arrangement does not include. The index belongs to OpenAI, the configuration belongs to OpenAI, and none of it moves with you when the ranking changes.
- PricingModels
Gemini 3.8 Flash ships with a published date for its own price rise
What happened
Gemini 3.8 Flash launched with two prices. Google set it at $0.75 per million input tokens and $3.75 per million output through 31 December 2026, rising to $1.50 and $7.50 on 1 January 2027. Alongside it came Gemini 3.8 Flash Cyber, a vulnerability detection and patching model gated through the Fairwind programme to governments, critical infrastructure operators and software maintainers, with no published price.
Source: GoogleWhat it means in production
A doubling with a date attached is more useful than a quiet increase, and it belongs in the business case where finance can see it. A workload whose economics only work at the introductory rate does not have economics, it has a grace period. Model the unit cost at the list rate, keep the architecture able to move the workload when the rate changes, and put the date in the diary. The Cyber variant is the more interesting half. Its capability is rationed by eligibility, which is a procurement question most organisations are not set up to ask.
- GovernancePricing
Meta prices consent to train at a twelve-fold discount, and puts the switch in a config file
What happened
Muse Spark 1.3 arrived from Meta with a second model string beside it, muse-spark-1.3-contributor, priced at $0.10 per million input tokens against $1.25 for the standard model and $0.20 against $4.25 on output, in exchange for permission to use prompts and completions to train future Meta models. The standard model does not train on customer data. Choosing between the two means changing a model identifier in configuration; no contract term governs it.
Source: Tech TimesWhat it means in production
The discount matters less than where the decision now sits. When no-training is a contractual protection, changing it takes procurement and legal. When it is a model name in a config file, any engineer with commit access can move a workload onto a training tier, and no data loss prevention tool, gateway or posture manager will see it, because nothing else about the request looks different. So if you allow this family of models at all, pin the permitted model strings at the gateway and alert on everything else. The model identifier is now a data boundary. Govern it like one.
- ModelsGovernance
Anthropic cuts cache reads by three quarters, and requires thirty-day retention to use the model
What happened
Anthropic shipped the same underlying model twice, as Claude Fable 5.1 and as Claude Mythos 5.1 under different safeguards, with Mythos available only to organisations vetted through its cyber and life sciences verification programmes and initially only in the United States. Cache reads fell from $1.00 to $0.25 per million tokens while input and output held at $10 and $50. The migration guide lists a run of breaking changes, among them the removal of forced tool choice, the removal of the option to disable thinking, and a requirement for thirty-day data retention that returns a 400 error for organisations on zero-data-retention terms.
Source: Anthropic migration guideWhat it means in production
Read the retention requirement first. Organisations that negotiated zero data retention, which is common in financial services and healthcare, cannot use this model without an explicit exception, and that is a procurement conversation with a longer lead time than the upgrade itself. The rest is the ordinary cost of living at the frontier: a real reduction on cached reads, paid for in integration work. That work is cheap if your prompts, tool definitions and model strings are version-controlled in one place, and expensive if they are scattered through application code, which is the actual argument for keeping the model layer behind a gateway you own.
- Open weightsGovernance
Z.ai holds an open-weight model back for two weeks over its own cyber capability
What happened
GLM-5.3 was announced on 14 August. Z.ai then held the weights back for around two weeks, giving access first to vetted security partners before releasing them publicly at the end of the month. The company reported 84.5 per cent on CyberGym against 77.2 for GLM-5.2, and 54.4 per cent on ExploitBench against 24.4, and said the gains that help defenders find weaknesses earlier carry risk for attackers once weights are public.
Source: BetaNewsWhat it means in production
Open weights are no longer a clean answer to data residency. A staged release means the version you can self-host may trail the hosted one, and the gap will be widest exactly where capability is most sensitive, which is often the capability that made self-hosting attractive in the first place. So plan a regulated workload around the date the weights actually land. And be clear internally that self-hosting moves the safety question inside your perimeter; it does not remove it.
- RegulationGovernance
The AI Act's transparency obligations become enforceable
What happened
Enforcement of the AI Act began, run by the European Commission's AI Office and national authorities, with transparency obligations taking effect the same day: interactive systems must tell people when they are dealing with AI, deepfakes must be labelled, and AI-generated or altered content must carry machine-readable marks. The Commission published a list of more than 180 organisations that had signed the accompanying code of practice.
Source: European CommissionWhat it means in production
For most production systems this is a small change that becomes an expensive one if nobody owns it until an auditor asks. Put disclosure in the interface and in the logs, where it will survive the next redesign. The machine-readable marking requirement is the one that reaches into architecture: the mark has to be applied at the point content is generated, which for most estates means touching several systems. Budget a quarter for that.
- TalentDelivery
Forward-deployed engineers become the industry's scarcest talent
What happened
Demand for forward-deployed engineers is projected to surge by the end of 2026, according to TechCrunch: the share of companies planning to hire them jumped from under ten per cent early in the year to around seventy per cent by the second quarter, against a pool of roughly 17,000 in the United States of whom only around 2,000 have elite-level expertise.
Source: TechCrunchWhat it means in production
Capacity is now the constraint on getting AI into production, and it is a hiring constraint. Organisations that cannot hire this profile will buy it as a service, they will judge the firms they buy it from on the depth of the bench, and the methods that survive will be the ones that hand capability to the client's own people as they go.
- MarketIndependence
Ode with Anthropic launches as a standalone services firm
What happened
Ode with Anthropic launched as an enterprise AI services firm under chief executive Chris Taylor, introduced by Anthropic, Blackstone and Hellman & Friedman and backed by a consortium including Goldman Sachs, General Atlantic, Apollo, GIC and Sequoia. It is aimed at mid-sized organisations across financial services, healthcare, retail, manufacturing and software.
Source: Business WireWhat it means in production
A second frontier AI lab now sells deployment as well as models. That collapses two procurement questions into one: you cannot ask who deploys the system without also asking whose models they have an interest in selling. We expect model-neutral evaluation to become a standard procurement requirement.
- MarketCloud providers
Microsoft commits $2.5 billion and 6,000 people to the Frontier Company
What happened
With 6,000 industry and engineering experts, the Microsoft Frontier Company opened as an operating business that deploys AI inside enterprise operations, positioned as the largest outcome-driven engineering organisation in the industry. Early customers include London Stock Exchange Group and Unilever.
Source: TechCrunchWhat it means in production
For organisations already committed to Azure and Microsoft 365, deployment capacity has now arrived inside the vendor relationship, carrying the platform dependency that implies. For everyone else, and for mid-sized enterprises in particular, the gap is unchanged, and model-agnostic architecture is the hedge either way.
- MarketSpending
Gartner puts 2026 AI spending at $2.59 trillion, with $585 billion on services
What happened
Gartner's forecast puts worldwide AI spending at $2.59 trillion in 2026, up 47 per cent, with AI services at around $585 billion. Gartner described 2026 as the inflection year in which enterprises expand embedded generative AI and new agents across multiple workflows, while noting limited appetite for disruptive change in favour of tactical initiatives.
Source: GartnerWhat it means in production
The money is moving from experimentation into services and operations. What buyers say they want is tactical, well bounded and attached to a number, which favours anyone whose method starts from a narrow value case and an evaluation set. The premium sits with firms that can run many small systems well.
- MarketIndependence
OpenAI launches the Deployment Company and acquires Tomoro
What happened
The OpenAI Deployment Company was set up as a majority-owned business, with $4 billion of initial investment from 19 firms led by TPG and including Bain & Company, Capgemini and McKinsey. On the same day OpenAI agreed to acquire Tomoro, a London and Edinburgh applied AI consultancy with around 150 forward-deployed engineers and clients including Tesco, Virgin Atlantic and Supercell.
Source: OpenAIWhat it means in production
The model company is now a services company. Traditional consultancies and integrators saw their shares fall on the news. Capacity to deploy OpenAI models went up that day, and so did the value of having someone independent run the head-to-head.
- MarketConsolidation
Accenture completes its acquisition of Faculty
What happened
Faculty, the London applied AI company founded in 2014, became part of Accenture on completion of the acquisition, taking more than 400 data scientists and AI engineers and the Frontier decision intelligence product with it. Faculty's founder Marc Warner became Accenture's chief technology officer.
Source: Accenture newsroomWhat it means in production
The independent applied AI bench in Europe got smaller on the day the market needed it most. Consolidation into a global consultancy brings scale and platform alliances, and it moves the firm's incentives along with them. Outside the largest accounts, independent capacity is now scarce.
- ResearchDelivery
MIT NANDA: 95 per cent of enterprise generative AI pilots show no P&L impact
What happened
The GenAI Divide: State of AI in Business 2025, published by MIT's NANDA initiative, found that around 95 per cent of enterprise generative AI pilots produced no measurable impact on profit and loss. It also found that deployments built with external specialists succeeded about twice as often as internal builds, and that the tools that stalled could not retain feedback or adapt to context.
Source: MIT NANDA report (PDF)What it means in production
This is the report that named the problem applied AI has to solve. Its findings are the basis of the Applicat Method: narrow processes, proofs built in production conditions, evaluation over opinion, and an operating model for the years after go-live.
FAQ
About the log
Bring the frontier into production.
Tell us about the process you want to change. We reply within one business day.