The Pulse — July 28, 2026
The signals that entered our radar, organized with sources and context to understand what changed.
Kimi K3 releases open 2.8T-class weights
WHY IT ENTERED THE RADARKimi says K3 is a 2.8T-parameter, 16-of-896-expert MoE with native vision and a 1M-token context window; the weights were promised by July 27. Its claims—long-running coding, compiler/kernel work, research agents, and video work—make this a genuine “open frontier” moment, not merely another cheap model.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Open source just crossed 3 trillion parameters—what can you actually run or use?” Separate the headline parameter count from active parameters, hardware reality, API availability, and independently verified results.
Gemini 3.6 Flash: the agent economics story
WHY IT ENTERED THE RADARGoogle positions 3.6 Flash as a cheaper agent workhorse: $1.50/M input and $7.50/M output tokens, 17% fewer output tokens than 3.5 Flash, and reported gains on coding, computer use, and research benchmarks. Flash-Lite targets high-throughput subagents at 350 output tokens/sec.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The model that makes agents cheaper by writing less.” Show why reducing tool calls and output tokens can beat marginal benchmark gains for a production workflow.
A $500 RL fine-tune of a 9B model reportedly beats frontier models on one real task
WHY IT ENTERED THE RADAROn its catalog-review workflow, FermiSense reports a GRPO-tuned 9B open model scored 87.3% versus 76.9% for its best frontier configuration, at $0.50 per 1,000 listings. That is a powerful case for treating frontier models as a data-generation baseline—not the final production system.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“A tiny fine-tuned model beat the frontier—but don’t copy the headline.” Explain the conditions required: a narrow task, a reliable scorer/eval, proprietary examples, and enough volume to justify training.
Anthropic’s open-weights position: no ban, but mandatory capability testing
WHY IT ENTERED THE RADARAnthropic explicitly rejects a blanket ban on open weights. It advocates chip controls, action on industrial-scale distillation, and mandatory safety testing for sufficiently capable models—open and closed. This is a nuanced policy story amid the Kimi release.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Anthropic is not asking to ban open models. Here’s what it is asking for.” Use the three-policy framework and contrast it with the social-media framing.
Claude Code’s new default: Opus 5 with 1M context—and nested agent teams
WHY IT ENTERED THE RADARVersion 2.1.220 makes Opus 5 the default Opus model with a 1M context window and adds strict network allowlists, a DirectoryAdded hook, and forwarding for nested subagents. The product story is not just model IQ: it is safer, more observable multi-agent execution.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The real Claude Code update isn’t the model—it’s agent-team plumbing.” Demo the mental model: coordinator → subagents → nested specialists, then explain why permissions and network constraints matter.
Better coding agents still struggle to preserve a codebase over time
WHY IT ENTERED THE RADARIn a small 17-checkpoint SlopCodeBench subset, HumanLayer reports Opus 5 at 4/17 strict passes (24%); no model reached the end of any challenge with everything passing. The benchmark reveals a gap ordinary “one-shot issue” benchmarks hide: handling evolving requirements without accumulating regressions.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Why coding-agent demos lie (a little).” Contrast a spectacular first task with the less glamorous test: can the agent change the project ten times without breaking old behavior?
Open-agent orchestration is becoming a reusable workflow, not a bespoke build
WHY IT ENTERED THE RADARThe project packages a pragmatic pattern for dispatching any CLI agent in tmux with file-based completion signals and multi-turn iteration. AI Jason’s July 20 upload referenced it while arguing that persistent “sidekicks” and orchestration reduce context/token waste.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Your agent team needs a completion protocol, not more prompts.” Explain the unsexy reliability layer—observable workers, durable handoffs, and a coordinator that can retry.
OpenAI finds that AI use is already reshuffling tasks across job titles
WHY IT ENTERED THE RADARFrom 800,000+ U.S. ChatGPT messages, OpenAI reports that 43.5% of occupation-specific work messages concern tasks associated with another occupation. Customer-experience, design, HR, legal, and marketing workers show especially high “task crossover.”
SUGGESTED EDITORIAL ANGLEOpen original source ↗“AI may not replace your job title—it may steal your next handoff.” Frame it around one person doing the first 80% of a task that used to require legal, data, design, or engineering help.