Market Intelligence

The state of AI.

Each quarter we read the flagship AI reports so you do not have to: the frontier labs, the major consulting firms and independent research, distilled to the headlines that matter for building AI maturity. Every figure is quoted from its primary source and linked, so you can check it in seconds.

Updated quarterly · Independent and vendor-neutral · Fully cited

Verify before you cite. Every figure below is quoted from its primary source and linked. Entries are summarised for brevity, and sources may revise their reports after publication. Always open the linked source and verify a figure before you cite it or act on it.

Latest signal brief 26 October 2026

What has moved since Q3

The Signal Brief tracks the flagship AI releases between quarterly editions. Four developments since September are worth your attention, each one quoted and linked in full in the brief.

The monitoring layer is now attack surface

Independent reviewers tested a lab's own oversight, not just its models.

METR and Redwood Research found two client-controllable settings that could switch off OpenAI's internal agent monitoring, and a way agents could in principle bypass its sandbox and automated review. OpenAI says the production issues have been patched.

METR, Redwood Research, OpenAI

Agent incidents got a government playbook

Governments moved from describing agent risk to prescribing practice.

NIST, the NSA and CISA published ten practices for agent incidents in frontier training and evaluation, built on the SP 800-61 incident standard. The UK AI Security Institute set out eight priorities for technical best practice.

NIST, NSA, CISA, UK AISI

Three labs, three different leads

No single model is best at everything, and the gaps between the leaders are a few points.

Gemini 4 Argon scores 19.6% on a legal agent benchmark where rivals sit in single digits, while Claude Opus 5.5 leads agentic terminal work by nine points. Meta's Muse Wonder scores 52 on an independent intelligence index.

Google DeepMind, Anthropic, Meta

Builders and workers expect different futures

How much of their own work people expect AI to do now shapes how much unemployment they expect.

Anthropic staff expect AI to do 75% of their work by 2030, against 28% for US workers. In both groups, ten more points of expected exposure means 0.4 points more forecast unemployment.

Anthropic

Download the AI Signal Brief (PDF)

Latest edition

Q3 2026

The signals that moved this quarter

This quarter's flagship releases from the frontier labs, the major consulting firms, independent evaluators and the governance bodies. Agents moved from a capability story to a containment one, and the value question moved from adoption to measurement. Every figure is quoted from its primary source.

Covering July to September 2026 · Published 26 October 2026 · 13 sources

Read the Q3 2026 edition Download the full brief (PDF)

McKinsey 80% report individual productivity gains from AI, but only 37% report any EBIT impact, flat on last year.
BCG AI spending doubled to 3.3% of revenue in a year, and 80% of it now sits outside the IT budget.
KPMG 89% report measurable value from AI, but only 12% consistently weigh that value against what it costs.
OpenAI GPT-6 Astra became the first model its maker rates Critical in cybersecurity, and the cheaper GPT-6.1 Sol followed four weeks later.

All editions

Method

Each quarter, SP Optima sweeps a fixed watchlist of primary sources: the frontier labs, the major consulting firms, independent research and the governance bodies. We select the flagship reports published that quarter, verify every figure at its primary source, and quote it verbatim with a link. Where a figure cannot be verified at source, it is left out rather than repeated second-hand. The result is a short, cited read, not a link dump.

This is general market commentary, not advice. Sources carry their own methodologies and commercial incentives, noted in each caveat line. Always open the linked source and verify a figure before you cite it or act on it.