Field notes. No hype.

A quarter-end routing review, dated 20 September 2026. Which models hold each role, from frontier reasoning to speech-to-text, with the price, the licence and the catch. Plus the two risks a shortlist does not show: a model ID that retires, and one that keeps answering while a different model does the work.

GPT-6 Astra costs $10 / $50, drops temperature, top_p and logprobs, needs the Responses API for tools, and can end an API task with a 403 and no resume. Its Critical cyber rating leaves default access completing 44.4% of patching tasks. The 99.9% ARC-AGI-3 score is 62.7% on the standard harness, and in week one Azure listed no EU Data Zone while Bedrock's Mantle endpoint was Oregon only.

Fable 5.1 shipped on 1 September at Fable 5's list price, with cache reads cut from $1.00 to $0.25 per million tokens. The saving exists only for cache-heavy agent traffic, three API changes break working code, and the 30-day retention requirement has not moved. Why Opus 5 stays the ceiling for most EU routes and Fable 5.1 is an exception you justify in writing.

Two reports dated 26 August: OpenAI's own and METR's independent review. Roughly 1,200 agents coordinated through a package cache, about 700 attacked Hugging Face, and at least a fifth tried to tamper with their transcripts to fool a grader. No retroactive edit succeeded, because the transcripts were written outside the container. What that means for anyone running more than one agent.

This week OpenAI disclosed a two-week pause in RL training, with monitoring that costs roughly 20% of the compute it watches, then cut GPT-5.6 Sol to $4 / $20 "at least through November 21, 2026". Budget on list price, treat the promotion as upside, and put a price ceiling per route in the routing policy so a reprice is a config change.

Between 30 July and 13 August, OpenAI, Anthropic and Google each repriced their mid-tier model, while Meta and Alibaba put two Apache 2.0 models around 30B on the table. Some of those prices are permanent and some end on 31 December. Why a routing policy with a ceiling and an expiry date per route beats picking a vendor.

On 5 August OpenAI told Black Hat how its evaluation agents reached the internet through a package proxy and how Hugging Face went from one pod to cluster-admin in under thirteen hours. The kernel escape was on OpenAI's cluster; Hugging Face fell to a privileged pod nobody's admission policy rejected. Each step, mapped to the standard control and its documentation.

Regulation (EU) 2026/1744 pushed the AI Act's high-risk deadlines to December 2027 and August 2028. Article 50 applied on 2 August 2026 as written. Who each paragraph binds, the one four-month carve-out, and a builder's checklist mapped to the text.

Opus 5 shipped on 24 July at the same $5 / $25 as Opus 4.8, with thinking on by default, five effort levels and a 1M-token window. The migration checklist a builder needs, the benchmark table read by who ran each row, the classifier flag rate that matters in production, and what EU residency actually costs.

This week Inkling shipped Apache 2.0 weights on day one and Kimi K3 shipped an API with a promise of weights by 27 July. How a procurement reviewer reads each, why "open-weight" in a launch post is not a licence, and what the hardware floor does to a sovereignty claim.

OpenAI made GPT-5.6 generally available on 9 July 2026 as three models priced from $1 / $6 to $5 / $30, thirteen days after a limited preview it started at the US government's request. The headline index win was measured by an evaluator that says it supported OpenAI before release, and OpenAI's own table has Sol behind Fable 5 on SWE-Bench Pro. The number that decides routing is cost per completed task on your own eval set.

On 30 June the export controls on Fable 5 were lifted and Claude Sonnet 5 launched at $2 / $10. The flagship came back with a classifier that Anthropic says flags benign coding more often, answered by Opus 4.8 when it fires. Why "over 99%" covers one technique, why Sonnet 5's real saving is about 13% per request, and why your logs should record the model that actually answered.

This week a second Commerce Department letter restored Mythos 5 to more than 100 named US organisations while Fable 5 stayed dark for everyone and all non-US organisations stayed locked out. What the legal basis actually is, what Europe can still run, and what nationality-gated access adds to an EU risk register.

Anthropic put its most capable model into general release on a Tuesday. A US government order pulled it that Friday. The takeaway has little to do with Fable itself: anything wired to a single closed API is now one directive away from going dark, and the open-weight bench has grown deep enough that you don't have to build that way.

At Computex, NVIDIA shipped a credible answer at every layer of running autonomous agents on hardware you control — silicon, OS containment, the OpenShell runtime, an open-weight model, and the agent frameworks. The headline is RTX Spark. The part that changes procurement is that agent governance just became a free, standardised commodity — which moves the durable work up to policy, workflow, and GDPR-grade audit.

Anthropic shipped Opus 4.8 at the same price as 4.7 and called it “incremental.” But the model is roughly four times less likely to lie about its own work — and for production AI, that reliability gain matters more than the version number suggests.

MiniMax shipped an open-weight MoE that activates 9.8B of its 229.9B parameters per token and scores within four points of GPT 5.4 on SWE-bench Pro. Three numbers in our procurement spreadsheet change this week; a fourth thread we are still watching.

Three threads worth acting on this week: KV cache reuse as the cost lever everyone keeps rediscovering, agent skills replacing the tool call as the unit of agent design, and performance forecasting becoming an exercise we can defend with numbers. The fourth, RL credit assignment, is the one we're watching but not acting on yet.