Mistral opens Large 4 weights, OpenAI updates its Model Spec, Step Functions workflows become agent tools
Today's thread is letting agents act while hard-coding the boundaries: OpenAI's new Model Spec turns autonomy scope and untrusted data into rules, and AWS lets agents trigger while deterministic workflows execute. On the open side, Mistral released a near-frontier 1T model and IBM showed model updates that only swap adapters.
A natively multimodal, hybrid instruct-and-reasoning MoE with about 1 trillion total parameters and 160+ languages; reasoning effort is low, medium or high, defaulting to high.
Licensed under the Mistral Ecosystem Open-Weight license, a modified MIT: free to use, modify, fine-tune and commercialize, but businesses running Model-as-a-Service need a separate commercial license.
48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task; 41% on Terminal-Bench 4.0, 13 points above the preview.
API pricing per million tokens: $0.70 input, $0.07 cached input, $2.10 output; it refuses malicious cyber requests 96.6% of the time on average, which Mistral says is the highest it measured among open models.
Builder's takeThis is the closest the open camp has come to the closed frontier this month, but 1T parameters is not something a small team self-hosts casually, so for most products the realistic path is the API first: it sits in the same price band as Terra, and one comparison run on PandaClaws-style long-form generation will tell you whether switching pays. Watch the license: if your product itself resells model calls, that is the Model-as-a-Service case that needs a separate agreement.
Five core values: be faithful to human intent, minimize harm, be honest and transparent, facilitate human agency, and act with care.
Instructions follow a chain of command (Root, System, Developer, User, Guideline, no default authority), with higher-authority and more recent instructions taking precedence.
Multi-step real-world actions must stay within an agreed-upon scope of autonomy, with a clear scope and ending condition and approval for anything out of scope.
Tool outputs, quoted text and web content are treated as information, not instructions, by default; side-effecting actions such as sending email or deleting files need extra care, and the model should support monitoring, correction and shutdown.
Builder's takeThe most useful part is the agent sections, which read like a ready-made design checklist: an agreed scope, an ending condition, approval for anything out of bounds, and tool output is not instructions. I would put it next to Amazon Quick's smart-approval list from earlier this week and walk every step of a PandaClaws-style pipeline that fetches pages and publishes content, looking for places where web content still flows straight into instructions.
It runs through Amazon Bedrock AgentCore Gateway, which exposes the workflows to agents over the Model Context Protocol (MCP).
Agents can start runs, for example a support agent launching a refund workflow with eligibility checks, manager approval and a credit step, and can look up runs, for example reporting which overnight runs failed and at which step.
Available in every AWS Region where AgentCore Gateway is offered, billed at standard Step Functions and AgentCore Gateway rates.
Builder's takeThis is the pattern I keep recommending: the agent decides only whether to act and with what input, while the side-effecting steps (approvals, charges, data changes) stay in a deterministic workflow. If you already orchestrate on AWS with Step Functions, wire up the read-only run lookup for ops questions first, then consider letting agents start workflows, with the approval step inside the workflow rather than enforced by the prompt.
GPT-6.1 Terra added to the bundled, Vertex AI and Amazon Bedrock model catalogs.
Clearer safety-block explanations in the TUI, with selectable safer follow-up prompts.
New codex doctor --network-only for focused configuration, connectivity and WebSocket diagnostics.
Fixes: approval and permission settings are preserved when resuming or forking threads, and reasoning-token accounting carries across resume and fork to avoid premature compaction.
Builder's takeThe useful part for teams is the Vertex AI and Bedrock routes: if your billing and data compliance already sit with one cloud, you can now run Terra in Codex from that cloud account without a separate OpenAI account. The fix for approval settings being lost on resume also matters: if you run long tasks via resume, upgrade and recheck your approval configuration.
AnySpeechRecognizer implements Apple's SpeechAnalyzer API, defaulting to Apple Speech but able to swap in NVIDIA Parakeet or Alibaba Qwen3-ASR running locally on Core ML, plus Sortformer diarization for up to 4 speakers.
On LibriSpeech (M4 Max), Parakeet Unified file transcription hits 1.43% WER on test-clean versus 2.12% for Apple's engine; Parakeet transcribes an hour of audio in about 20 to 40 seconds using about 100 MB of memory.
Every engine tested had a word error rate above 23% on AMI meeting recordings.
AnyDecisionModel reads answers from model probabilities instead of parsing generated text, across 6 backends including MLX, Core AI, llama.cpp and Ollama; it needs Swift 6.2 and macOS 26 / iOS 26 or later.
Builder's takeFor on-device transcription products like AI Interview, this is a cheap place to experiment: same API, swap the engine and compare. But take the meeting-audio result seriously, every engine above 23% WER: multi-speaker, far-field interviews are much harder than read speech, so evaluate on your own recordings. Reading answers from probabilities instead of parsing text is also worth borrowing; scoring tasks get much more stable.
Granite Switch is a modular architecture that mounts cache-efficient specialized adapters on one model, serving several capabilities from a single endpoint; this release builds on granite-switch-4.1-30b.
Only a new speech recognition model, granite-speech-5.0-470m-turboctc (10 languages, built-in punctuation and capitalization), and an updated Decision adapter were trained; the base model was not changed.
Fully backward compatible with 4.1 with no migration; IBM plans tutorials and an Adapter SDK so developers and third parties can build and version their own adapters.
Builder's takeThere are no scores this time; what is worth watching is the release model: base unchanged, adapters swapped, which means most of your production eval baseline does not need rerunning. If you already mount LoRA-style adapters per feature, version and regression-test each adapter on its own instead of upgrading the whole model every time.
Researched and drafted with AI assistance; editorial standards and views set by Darius.
Every Monday, the week in AI in one email
Every Monday morning: the 6 AI stories that mattered most in the past seven days, with my take, in your inbox.
Your email is only used to send the weekly and is never shared. Every email has a one-click unsubscribe link. See the latest issue →