The AI Transformation Brief—September 23, 2026
The AI Transformation Brief
// Today’s Signal
The enterprise AI market is moving from new model releases to AI work that is authorized, measured, and reviewable. OpenAI and Anthropic are cutting the cost of capable AI problem solving and business work that can continue for hours. Palo Alto Networks is turning the most advanced AI models into a continuous security operation. Okta is building a cross-vendor architecture for finding, limiting, monitoring, and stopping AI software workers. Databricks is pulling model choice, budgets, tool permissions, and records showing what happened into the data platform. The converging signal is direct: intelligence is becoming easier to replace, while the system that authorizes, measures, and proves machine work is becoming the strategic asset. The next AI budget should fund control over outcomes, not another unconnected license for one AI model.
// Top Stories
OpenAI introduced GPT-6 Sol and GPT-6 Luna as faster, lower-cost models built from the same advances behind GPT-6 Astra.
OpenAI lists GPT-6 Sol at $2 per 1 million input tokens, which are small units of text an AI model processes and $10 per 1 million output tokens, which are small units of text an AI model generates, down from $4 and $20 for GPT-5.6 Sol promotional pricing. GPT-6 Luna is listed at $0.10 per 1 million input tokens, which are small units of text an AI model processes and $0.50 per 1 million output tokens, which are small units of text an AI model generates, also a 50% reduction from its predecessor's promotional pricing. OpenAI The pricing change arrives with usable enterprise performance claims. OpenAI reports GPT-6 Sol at 33.2% on AutomationBench 1.0.6 at a stated $0.27 cost per task, versus 26.9% for Claude Opus 5 at 11.1 times the cost per task. On DeepSWE 1.1, OpenAI reports 68.8% for GPT-6 Sol at maximum effort, within 1.1 percentage points of Claude Fable 5's 69.9% at extra-high effort and at approximately 80% lower cost per task. The company notes that the Claude fallback comparison excludes some fallback cost and that its evaluations have important caveats. OpenAI
The story is not a new model tier. It is a new procurement floor. When a model reaches top-tier coding and computer-use performance at a fraction of the task cost, routing becomes a business decision rather than a research preference. The value moves to the system that chooses the right AI model for each task, preserves context, measures quality, and proves when the cheaper path is safe. Put a test of choosing the right AI model for each task into every material AI workflow this quarter. Compare verified completion, rework, latency, and cost, not the price of small units of text alone. OpenAI
Anthropic announced Claude Opus 5.5 on September 22, 2026, as the first model in its Claude 5.5 family.
The company lists pricing for software connections at $4 per 1 million input tokens, which are small units of text an AI model processes and $20 per 1 million output tokens, which are small units of text an AI model generates, compared with $5 and $25 for Opus 5, and says the model is available through Anthropic, Amazon Web Services, Google Cloud, and Microsoft Azure. Anthropic reports 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1, 1,846 Elo on GDPval-AA v2.1, and 81.8% partial on OSWorld 2.0. Anthropic Anthropic says Opus 5.5 completed a 680,000-line code migration in less than a day and a 200,000-line codebase audit and fix in under three hours in its examples. It also says the model is about 85% less likely than Opus 5 or Claude Mythos 5.1 to attempt containment-boundary circumvention, while warning that benchmark margins are less reliable at these capability levels and that pre-deployment failure detection remains unsolved. Anthropic
The enterprise consequence is a lower cost for business work that can continue for hours, not a better chat answer. The multi-cloud availability weakens the argument that a single provider must own the entire path from model to business workflow. The scarce asset becomes the way the company organizes work that can continue for hours: planned human review points, approvals, safe tool access, tests showing whether the result is good enough, and a clean handoff when the model reaches a boundary. Assign one real coding or knowledge workflow that can continue for hours to the model, but require a human approval point and a measurable, tested way to reverse a bad action before expanding authority. Anthropic
Palo Alto Networks announced Unit 42 Continuous Frontier AI Defense on September 22 as an annual-subscription service available worldwide.
It combines Unit 42 security experts with Anthropic Mythos models, OpenAI GPT cyber models, and models whose trained files are available for use to discover security weaknesses, check whether a weakness can be used in a real attack, simulate routes an attacker could take, and prioritize fixes, and recommend code-level fixes across applications, identities, cloud infrastructure, source repositories, APIs, and network assets. Palo Alto Networks The company reports that one Unit 42 investigation used more than 50 recognized methods attackers use in under 10 hours, 97% faster than a skilled red team. Palo Alto also reports that continuous Mythos-based scanning produced more than one year of traditional penetration-testing results in three weeks, found 3.2 times more high and critical security weaknesses per product, and reduced average time to fix a problem by 51%. It cautions that no single model found more than 40% of vulnerabilities and that leading cyber models had less than 10% overlap in what they found. Palo Alto Networks
Security testing is shifting from an annual inspection to a continuously running production capability. The important design choice is not which AI model finds the most issues by itself. It is whether multiple models, expert review, and systems that fix problems can produce a verified answer before the attack path changes again. That makes coverage, evidence, and response speed a company-wide way-of-working problem. Create a live list of security problems with an accountable owner, a severity rule, and a time bound from discovery to proof that the problem was fixed. Palo Alto Networks
Okta announced the Blueprint Alliance on September 22 with AWS, CrowdStrike, Databricks, Docker, Google Cloud, Lovable, Proofpoint, Salesforce, ServiceNow, Wiz, and Zscaler.
The open reference architecture asks four operational questions: where AI software workers are, what they can do, what they are doing, and how the organization responds. The members say the architecture covers discovery and digital identity, access limited to one specific task, a record showing who or what delegated an action, permission checks while software is running, monitoring, containment, limits on how often a system can act, canceling the software credentials that allow action, stopping an active software session, and a controlled way to restore access. Okta Okta cites a Gartner prediction that an average global Fortune 500 enterprise will use more than 150,000 AI software workers by 2028, while only 13% of organizations believe they have appropriate rules for AI software workers. The alliance says members will test cross-vendor signal sharing and publish integrations using MCP, a standard that lets AI tools connect to other software, plus OCSF, SSF, and CAEP. No pricing or adoption results are provided, so the immediate signal is the creation of shared rules rather than proven market scale. Okta
The market is standardizing the questions that determine whether machine work can be trusted. Digital identity is becoming a shared enterprise asset across model providers, data platforms, business applications, and infrastructure vendors. That creates room for a neutral coordination layer, but it also creates a pile of rules nobody can consistently enforce if each vendor invents its own signals and response controls. Make AI software worker inventory, task-scoped access, action logs, revocation, and recovery mandatory procurement requirements. Do not approve AI software that cannot answer all four questions in production. Okta
Databricks' Unity Gateway documentation, updated September 22, describes an enterprise AI governance service built on Unity Catalog.
It governs models, AI software workers, MCP servers, tools, what happens while the software is running, traffic, safety rules, and usage from one management layer. Foundation Model APIs provide pay-per-token access, charging for the small units of text processed to GPT, Claude, and Gemini without requiring the customer to run the model infrastructure. AI assets are registered as data and AI resources with permissions, with access granted or revoked through catalog privileges. Databricks Unity Gateway supports limits on how often a system can act, traffic splitting, fallbacks, Smart Routing for coding tasks, budgets, per-user thresholds, and hard caps across model and MCP services. Its monitoring layer records requests, tokens, latency, spend, performance, activity, and audit data in system tables and Unity Catalog Delta tables. Databricks says external providers can be used with using the customer's own provider credentials, and that tool filtering and company rules can restrict MCP access. Databricks
Choosing an AI model is becoming a managed business rule inside the data stack. The durable advantage is not access to GPT, Claude, or Gemini. It is the ability to route work, enforce limits, preserve evidence, and move between providers without losing the company's rules. This is where the enterprise can keep model competition while retaining control over cost and authority. Put model choice, spending limits, tool permissions, and records showing what happened under one owner, and test a provider switch on a real workflow before relying on one model again. Databricks
// Shelly Palmer Pulse
Shelly Palmer reports that OpenAI says an internal model has resolved more than 100 long-standing open problems across most areas of mathematics and formally verified its Navier-Stokes result in Lean in 17 hours.
Palmer highlights OpenAI's independent nine-member Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study in Princeton, and the concern from 27 Fields Medalists that rushed AI announcements can outrun proper writeups, attribution, and scientific standards. The enterprise lesson is direct: when AI can produce work that is machine-checkable, the scarce capability shifts to verification, attribution, and deciding how the work enters a trusted institution. Shelly Palmer
// What It Means For Your Business
The strategic asset is moving from access to models toward control over verified work completed by software.
Model capability is becoming easier to substitute while digital identity, choosing which model handles each task, safety rules, and proof remain harder to replace.
Expect the market to reorganize around the management layer beneath the interface, with model vendors, data platforms, security companies, and identity providers competing to define the authority path. Databricks Okta
Trust will become a customer-visible product attribute when AI systems produce or act across company boundaries.
Position services supported by AI around verifiable outcomes, clear clear rules about what software may access, and a named answer to who is accountable when an automated action fails; publish enough evidence for customers and partners to evaluate the service. Shelly Palmer Okta
Redefine productive work as a completed, reviewed, and linked to the person or software that produced it.
Build one workflow with checkpoints, safe retries, a cost limit, a quality test, an escalation route, and a tested stop path before granting the system authority to change a customer, financial, security, or compliance record. Anthropic Palo Alto Networks
Require every AI investment case to show cost per verified completion, rework, failure recovery, and the capacity or revenue unlocked.
Model prices will fall, but unmanaged retries, weak controls, and unmeasured remediation can erase the savings. OpenAI Databricks
Separate models from the shared systems that authenticate data, route requests, control tools, enforce limits, record events, and manage recovery.
Standardize digital identity, company rules, model choice, and ways to preserve proof of what happened, then test a provider switch against one high-impact workflow before making a model a dependency. Databricks Okta
Treat AI permission to act as a decision to give software authority, not a software purchase.
Require management to report which digital identities used by AI software can act in live business systems, what they can change, how quickly they can be stopped, and whether a material action can be reconstructed from an evidence trail. Okta Palo Alto Networks
The most consequential shift is that model capability is becoming a cheaper input to AI work that is authorized, measured, and reviewable. The assumption it breaks is that the best model will own the economics of enterprise AI. The decision it forces is whether the company will own the model-choice, digital-identity, evidence, and recovery layer that turns intelligence into accountable work or let each vendor define that boundary separately.
Are you still buying AI as a feature, or are you building the system that makes machine work accountable when the model, provider, and interface all change?
The Transformation Brief is written daily by Les Ottolenghi. Delivered every morning at 6:00 AM MT, a 7-minute read on the AI shifts that matter to operators and boards.
**Build the system that lets AI tools work together. Price the outcomes. Redesign the organization.**
Read On
The AI Transformation Brief—October 14, 2026
Banks, card networks, and regulators are deciding who answers for AI that acts for customers, and companies that wait will inherit their ...
Read the brief →The AI Transformation Brief—October 11, 2026
Gatekeeping decides who benefits from AI, and this weekend shows four different gates. Microsoft's CEO argues the limits on what a model ...
Read the brief →The AI Transformation Brief—October 10, 2026
Permission is becoming the operating constraint on AI. Anthropic is cutting live-internet access for internal AI tests until its monitoring ...
Read the brief →