STEVEN STEINBECK
← ALL DISPATCHESDISPATCH 58

WEEK OF JAN 31 – FEB 6, 2026

Twenty minutes apart: Opus 4.6 and GPT-5.3-Codex

Anthropic shipped a million-token Opus and agent teams, OpenAI answered with a coding model that helped debug its own training, and the agent-harness discipline got its name.

Two frontier coding models landed twenty minutes apart, which is now the normal shape of a release week. What is worth reading is not who won the score table but what both shipped: better context management and cheaper long agent runs.

By Friday a practitioner essay had named the thing everyone was already doing — engineer the harness, not the prompt — and within days a frontier lab published its own version of the same argument.

THU · Feb 5, 2026Frontier modelsLong context

Claude Opus 4.6 brings a million-token context and agent teams

Anthropic upgraded its top model, targeting what it calls context rot — quality decay over long conversations. Opus 4.6 is the first Opus-class model with a 1M-token context window, supports up to 128k output tokens, and adds context compaction on the API to summarize and replace older context mid-task, plus adaptive thinking with effort controls.

In Claude Code it introduces agent teams as a research preview: multiple agents working in parallel on independent, read-heavy work such as codebase review. Anthropic reports the highest score on Terminal-Bench 2.0 and roughly +144 Elo over GPT-5.2 on GDPval-AA.

Claude Opus 4.6 brings a million-token context and agent teams
cdn.sanity.io

WHY IT MATTERS

For developers this is the long-context-and-orchestration release. Compaction plus 1M context directly addresses the context-management pain of long agent runs, and agent teams move harness design from one loop to a small supervised fleet.

AnthropicSystem cardGadgets360
THU · Feb 5, 2026Frontier modelsCoding agents

GPT-5.3-Codex, the first model OpenAI says helped build itself

Roughly twenty minutes after Anthropic's post, OpenAI released GPT-5.3-Codex, merging GPT-5.2-Codex's coding with GPT-5.2's reasoning. It runs 25% faster and uses less than half the tokens of its predecessor.

OpenAI says early versions debugged their own training runs, managed deployment and diagnosed evaluations — its first model instrumental in creating itself. Reported scores: Terminal-Bench 2.0 at 77.3% versus 64.0%, and OSWorld-Verified at 64.7% versus 38.2%. It is also the first OpenAI model rated High capability for cybersecurity under its Preparedness Framework, with API access held back pending safety review.

WHY IT MATTERS

Two things matter beyond the score table: token efficiency, which makes long agent runs cheaper, and mid-task steerability without losing context. The High cyber rating plus a delayed API rollout is the template for how frontier coding models will ship from here.

OpenAIGadgets360
THU · Feb 5, 2026Agent platformsEnterprise

OpenAI launches Frontier, a control plane for agent fleets

Also on February 5, OpenAI announced Frontier: a platform for building, deploying and managing agents across enterprise systems, framed around giving agents the same things people need at work — shared business context, onboarding, feedback-driven evaluation, and scoped identity and permissions.

It is positioned as vendor-neutral, supporting third-party agents, and runs locally, in a customer cloud or hosted, with SOC 2 Type II and ISO 27001/27017/27018/27701 compliance. Early adopters include HP, Intuit, Oracle, State Farm, Thermo Fisher and Uber.

OpenAI launches Frontier, a control plane for agent fleets
TechCrunch

WHY IT MATTERS

This is agents-as-coworkers productized: identity, permissions, evals and audit trails wrapped around agent execution. If you are building internal agents, it is a direct competitor to a self-built harness plus orchestration layer.

OpenAITechCrunch
THU · Feb 5, 2026HarnessesPractice

Mitchell Hashimoto names the discipline: engineer the harness

HashiCorp co-founder Mitchell Hashimoto published “My AI Adoption Journey,” a six-step account of adopting coding agents. Step five is the one that spread: every time an agent makes a mistake, engineer a permanent fix into its environment — a documentation file or a programmatic tool — rather than re-prompting, so the harness gets stronger as the agent works.

He pairs it with practical habits and a caveat that skill formation in junior engineers deeply worries him. The framing — agent equals model plus harness — was picked up within weeks by Thoughtworks' Birgitta Böckeler and echoed in OpenAI's own harness-engineering write-up.

WHY IT MATTERS

It gave practitioners vocabulary for what they were already doing, and the core claim is testable: with the same weights, harness changes move results more than model swaps in many workloads. The reactive rule — one observed failure, one permanent environmental fix — is cheap to adopt in any repo.

Mitchell HashimotoMartin Fowler on harness engineering
WED · Feb 4, 2026Open weightsVoice

Mistral ships a self-hostable realtime ASR model

Mistral released two speech-to-text models. Voxtral Realtime is 4B parameters under Apache 2.0 with weights on Hugging Face, using a purpose-built streaming architecture rather than chunking an offline model, with configurable latency down to sub-200ms. At 480ms delay it stays within 1–2% WER of the batch model.

Voxtral Mini Transcribe V2 is the closed batch model at $0.003/min with speaker diarization, word-level timestamps, context biasing and 13 languages, claiming roughly 3x the speed of ElevenLabs Scribe v2 at about a fifth of the price.

WHY IT MATTERS

A 4B open-weight streaming ASR model you can self-host on a single 16GB GPU changes what voice agents cost to run — realtime transcription stops being a paid API dependency and the audio stays inside your infrastructure.

MistralSimon Willison
MON · Feb 2, 2026FundingDeployment

Waymo raises $16B at a $126B valuation

Alphabet's Waymo closed a $16 billion round at $126 billion post-money — more than double its October 2024 valuation — led by Dragoneer, DST Global and Sequoia, with Alphabet remaining majority investor.

Waymo says it does more than 400,000 rides a week across six US metros, tripled annual volume to 15 million rides in 2025, and points to 127 million fully autonomous miles with a 90% reduction in serious-injury crashes. The capital funds groundwork for over 20 additional cities in 2026, including Tokyo and London.

WHY IT MATTERS

Robotaxis remain the clearest at-scale deployment of AI in the physical world, and this is the largest bet yet that the unit economics hold. Regulatory and city-by-city expansion, not model quality, is the binding constraint.

WaymoCNBCThe Verge