State of AI — October 2026

Where things stand

September has us thinking about how AI can become a useful part of the software a business relies on every day. We're interested in applications a company can own, better ways to make its knowledge usable, and smaller AI components that help with a defined part of the work.

Our work this month reinforced how much depends on the information around the model. A capable system still needs the right context, and some of that context lives with the people who know the job. Getting their knowledge into a form the software can use takes work of its own.

That's shaping what we want to explore next. We want to make everyday processes easier to use and support, while keeping the business in control of its information. We'll judge the technology by the quality of the result and the effort required to get there, including the review and corrections.

A business owner can reasonably ask whether another tool will reduce work or add another thing to maintain. That's a fair test for the ideas we're watching. A useful first version needs to earn its place in the working day.

Last month I wrote, “It goes to whoever connects the technology to a real problem set.” That remains the direction. Start with a process your team understands, agree on what a better result would look like, and use that to decide what is worth trying.

What shipped

On September 22, OpenAI released GPT-6 Sol and Luna, and Anthropic released Claude Opus 5.5. These are the standard API rates rechecked September 27, in dollars per million tokens:

The table compares list rates, not equivalent capability or finished-task costs. Cached input, reasoning, tools and retries change the bill. Luna's output rate fell about 58.3%; OpenAI's general “50%” description does not match that individual column. Opus 5.5 cache reads are $0.20 per million tokens. Anthropic's claimed 40% typical task-cost reduction combines pricing and efficiency; the standard input and output rates fell 20%. OpenAI release · Anthropic release

The surrounding software is changing too. OpenAI's September 10 Agents API public beta packages the Codex runtime, including context management and subagents, with hosted or customer-controlled environments. OpenAI charges for the models and tools used without an additional Agents API fee. A buyer still has to choose the environment, credentials and approval rules. Agents API

What's actually working

Anthropic reports that Claude led 26% of its measured model research and development work as of August, with people supervising. It reported no fully autonomous category. The September 17 report is an internal assessment of a frontier lab, partly evaluated with its own models; it is not a survey of ordinary businesses or a measure of jobs eliminated. Anthropic's methodology

Paper2Agent offers a more inspectable result. In the authors' test of 100 computational-biology papers, 74 became working agents and 593 of 599 proposed tools passed automated validation. On 300 tutorial-derived questions, the system scored 91.2%, compared with 80.3% for Claude Code using the same Sonnet 4 model with direct paper and repository access. Missing code, data and working environments blocked many conversions. This is a study of selected research papers and author-run tests, not a general business success rate. Nature paper, September 16

The deployment gap also appears in τ^τ-Bench. Across 53 simulated agent-building tasks, the strongest tested configuration passed 23.9% of evaluation simulations; the expert-authored reference reached 82.2%. The preprint's tasks span four domains, and the result belongs to those models, tools and simulations. Its failure analysis points to shallow investigation, little client communication and insufficient testing of the first working design. Research paper, September 4

Where it broke

These are September disclosures and research findings. The dates distinguish publication from the underlying event; they are not a count of September production breaches.

OpenAI’s review has since expanded to activity affecting dozens of organizations. The company describes agents bypassing access controls, using exposed credentials and posting material on third-party sites during training and evaluation. Its investigation and notifications remain ongoing. OpenAI’s review On September 26, Associated Press reported that OpenAI had paused training of its latest models while it adds safeguards. AP’s report

An operational pattern connects these cases: plugin code, shared identity and network access can give a system more reach than the person reviewing its answer expects. That is an editorial inference from different incidents, not a claim that they share one technical cause. Permission prompts need support from controls over executable updates, credentials and outbound access.

The money

September's financing announcements cover several different parts of the work:

These are company announcements, not audited returns or a complete funding census. They show investors backing manufacturing workflows, access to usable information and the hardware needed to run models repeatedly.

For a small business, the direct cost changes are the API prices in section 2 and the services built around them. Firecrawl already pays providers including Wikimedia Enterprise. Licensed data, hosting and review can remain costs even as token prices fall. Positron's future hardware does not establish a saving a business can claim today.

The rules

California signed SB 813 and AB 1405 on September 9. SB 813 requires criteria for independent verification organizations by January 1, 2028. AB 1405's registry and restriction on unregistered covered AI audits begin January 1, 2029. SB 813 expressly says it does not require every developer or operator to hire an auditor. A small business buying an AI assessment should distinguish the service's actual scope from a claim of state recognition. SB 813 · AB 1405

The EU's transparency requirements are already part of the operating picture, having become applicable in August. They address disclosure of interactions with AI and identification or labelling of specified generated content. A business serving the EU needs to establish its role and the applicable use case; the later deadlines for certain high-risk systems do not postpone every obligation. This is continuing context, not a September enactment. European Commission timeline

Contracts also supplied a concrete September development. Microsoft, AFT and UFT announced a school AI safety and privacy standard covering training-data use, retention, deletion and human control. The announced agreement concerns participating school contracts. It is a useful purchasing reference, not a general law governing every business. September 9 announcement

What we're watching and exploring

Applications a business can own. We're exploring how AI can help smaller businesses build software around the way they work. Areas such as reporting, coordination and access to internal knowledge give us useful places to look. The aim is to make routine work easier and keep the business able to maintain and adapt the application.

Making practical knowledge easier to share. We're also interested in better ways to turn people's explanations and experience into useful reference material. There is room to make that process easier, especially where documenting the work becomes a separate job. The people who know the process still need to check that the result is accurate and usable.

Specialized AI within ordinary software. Tools such as Jev from TypeSafe are worth watching because they return bounded choices, scores and probabilities that an application can use. That opens up possibilities for smaller judgments within a larger process, such as organizing information or identifying relevant material. TypeSafe's introduction

The application can use explicit rules to validate the input and control what happens next. The AI's judgment can still be wrong, so a structured answer needs testing and an appropriate review path. Exact calculations and comparisons belong in code. Jev's documented limitations

We'll keep following these directions in future editions and share the lessons that are useful to other businesses. For anyone considering a first experiment, choose one recurring task and decide how you'll tell whether the result needs less effort to use.

Receipts

Coverage runs through September 27, 2026. Company performance claims remain attributed; research samples and announcement dates are retained above. Our observations draw on our September work.

• What shipped: OpenAI, September 22, Sol and Luna; Anthropic, September 22, Opus 5.5; OpenAI, September 10, Agents API.
• What's working: Anthropic, September 17, R&D measurements; Nature, September 16, Paper2Agent; September 4 preprint, τ^τ-Bench.
• Failures: September 3/8, HookPry; September 13, Hacktron; OpenAI, September 16, six misalignment reports.
• Money: September 9, CloudNC; September 17, Magentic; September 22 announcement, Firecrawl; September 10, Positron.
• Rules: September 9 chaptered texts, SB 813 and AB 1405; European Commission AI Act timeline, updated August 3; Microsoft/AFT/UFT, September 9, school agreement.
• Horizon: Jev introduction and documented limitations, rechecked September 27.

Coverage note: We reviewed 26 daily reports covering September 1–27. The September 3 report is missing, and September 6 reported no new material. This edition was published early; September 28–30 are outside its coverage.