State of AI for development: September 2026

Three things from September that we think change decisions, and what we are telling clients to do about each.

Decision-only models arrived at a fraction of chat-model cost

TypeSafe launched Jev, a model that cannot write text. It picks from options you define, scores something or returns a calibrated probability, at 40 to 400 times less than a chat model, and it does not refuse. It passed a trillion tokens a day within a week, around a hundred open clones appeared in a fortnight and OpenAI shipped its own Decisions API. Early tests put Jev as a judge around 277 times cheaper than GPT-6 and within a few points on some benchmarks, and a cascade that escalated only low-confidence calls kept 99 per cent accuracy at 57 per cent of the cost.

What we are telling clients: much of the AI inside real software is a decision, such as routing a ticket, judging whether a document is relevant or choosing whether to escalate. Until now that meant training a classifier, which needs labelled data and someone who knows how, or asking a chat model, which is slow, expensive and occasionally refuses. Calibrated decisions at near-zero cost change what is worth automating, including judging every agent trace. Walk through where your systems ask a chat model to make a choice and trial a decision model there against your own labelled examples. Sceptics beat Jev with a small embedding model and logistic regression on one standard classification set, so test the benchmarks on your own data before trusting them.

Daily AI use rose while training fell

PwC’s survey of 49,364 workers found daily AI use up from 14 per cent to 22 per cent in a year, while the share who say they have the training they need fell from 59 per cent to 51 per cent. More than half sit in what PwC calls the engine room, with neither scarce skills nor much AI experience, and around a third say they are regularly overwhelmed. BCG found AI spend has doubled to 3.3 per cent of revenue, with more than 80 per cent now outside IT and only 5 per cent of firms holding a full set of controls. KPMG found 12 per cent consistently measure AI value against cost.

What we are telling clients: BCG’s figure explains PwC’s. When AI was an IT programme, someone owned the rollout, the training and the controls. Now business units buy tools for their own teams and nobody in particular owns enablement, so licences get bought and usage climbs while the training that would make that usage good falls into the gap. Give AI enablement a named owner wherever the spend sits. Run coaching in waves, with small cohorts working on their real work for a few weeks before the next wave starts. It is slower than a licence rollout or an e-learning module, but it reaches the people who will not teach themselves, and each wave produces coaches for the next. Track training alongside licence spend.

Accumulated agent guidance is ageing faster than teams prune it

In January we said instruction files were the highest-return thing a team could do. This month Thariq Shihipar from the Claude Code team suggested starting new projects without a CLAUDE.md and adding only the failure modes you see repeatedly, because the lists are model-specific and pile up into over-constraint. Anthropic’s Opus 5.5 guidance drops old prompt tricks because they now clash with how the model was trained. On one agentic coding benchmark Opus 5.5 at extra-high effort scored lower than at medium, for 2.8 times the cost, mostly through scope creep. A skills paper found retrieved skills can lift the overall score while hurting the specific tasks where a skill fired.

What we are telling clients: every instruction, skill and prompt trick was written to fix something a particular model got wrong, and the model now changes every few weeks. A person will ignore advice they know is out of date, but a model follows it, so a team that has spent a year adding to its CLAUDE.md files and skill libraries may be making its agents worse. We agree with Anthropic and would prune hard. Treat guidance as expiring by default: anything written for a previous model or more than a couple of months old is out of date unless someone has checked it still helps. When a new model lands, go back to minimal instructions and medium effort and re-add only what you see it get wrong. Claude Code now reads AGENTS.md when there is no CLAUDE.md, so one file can serve the tools that support it, and Anthropic has added evals for skills so you can test with and without before keeping them.

---

This digest is the public edition. Advisory clients get the full monthly review, a tailored edition for their context and the conversation that goes with it.