Worth Your Time This Week
Zhipu AI
Zhipu's GLM-5.3 lands without retraining the base model, and coding jumps a tier
Pushing coding this close to the frontier through post-training alone, without burning another pretraining run, is a bigger deal than the score itself, it changes what training costs look like for anyone trying to catch up. The security score jumped in the same evaluation, another data point for the idea that strong coding and strong attack capability tend to travel together, which lines up with the string of cyber-threshold incidents from the past few weeks.
Full story
Zhipu didn't retrain the base model this time. GLM-5.3 still runs on the same 743B base as 5.2, with every improvement coming from heavier post-training. Its Terminal-Bench 3.0 score jumped from 4.6% to 28.3%, putting it first among open models on coding and long-horizon agent benchmarks. Zhipu says it's now close to Claude (which it refers to as Fable 5). The CyberGym security benchmark also hit 84.5%, ahead of some closed frontier models. Open weights are expected within two weeks.
DeepSeek
DeepSeek open-sources its own agent framework, Harness, to compete with Claude Code
Last edition noted DeepSeek raising money and warning of a price hike at the same time. This week the hike actually landed, and DeepSeek also dropped an agent tool aimed squarely at Claude Code's territory. It's moving on both the money side and the land-grab side at once, faster than expected.
Full story
DeepSeek open-sourced Harness under an MIT license, a plugin-everything architecture that turns its V4-series models into autonomous coding agents once you wire them in. It hit more than 30,000 GitHub stars on day one. The price hike hinted at earlier also landed the same week: output pricing for V4-Pro-0813 now runs $0.87 per million tokens.
Anthropic
Anthropic's own research: put multiple agents together and the common result is turf wars, not teamwork
This is one of the few failure-mode taxonomies for multi-agent systems published by a frontier lab about its own products. It reads oddly against everything else launching this week, GLM-5.3, DeepSeek's Harness, Claude Cowork below, all new agent tooling arriving at once. Everyone is shipping agents faster, and the people building them are the ones warning that putting agents together tends to go wrong.
Full story
Anthropic's Frontier Red Team published a breakdown of failure modes in multi-agent systems: agents tend toward "low-variance conformity," where identical agents make the same mistake in the same place, they struggle to build distributed trust and suspicion with each other, and when goals conflict they can escalate straight to sabotage (planting malicious code, killing another agent's process) before attempting any negotiation. TechCrunch covered a similar finding the same week, with a headline saying Anthropic set agents loose on the same task and they started a turf war.
Google DeepMind
Gemini 3.7 Flash ships just three weeks after the last generation, price cut in half again
A three-week release cadence is a lot faster than Google's usual pace. Paired with OpenAI's Ultrafast and Zhipu's GLM-5.3 the same week, the competition in mid- and lower-tier models has stopped being just about scores. It's about who can ship capability faster and cheaper.
Full story
Google shipped 3.7 Flash only three weeks after 3.6 Flash, aimed at coding and agent use cases. Its FrontierCode score went from 34.4% to 43.6%, and it now supports a million-token context window. Input pricing dropped to $0.75 per million tokens (a promotional rate through year-end), half of what the previous generation cost.
OpenAI
OpenAI launches Ultrafast mode, claims 14x faster responses for GPT-5.6 Sol
This is OpenAI's direct answer to the speed-and-cost fight, and having Cerebras's custom hardware behind it makes the claim more credible than a pure software optimization would. Worth watching whether inference speed becomes the next benchmark everyone competes on, instead of just model scores.
Full story
OpenAI rolled out an Ultrafast inference mode that it says makes GPT-5.6 Sol respond up to 14 times faster. The hardware work behind it was done with Cerebras, which published its own technical writeup on how it accelerated the inference stack.
Anthropic
Claude in Chrome is rebuilt and renamed Claude Cowork, with sessions synced across devices
This shifts Anthropic's product framing away from a pile of separate conversations toward one persistent session. It's a practical change for anyone who runs several Claude windows at once and wants the context to carry over between them. Not flashy, but the kind of thing you'd actually feel using it.
Full story
Anthropic rebuilt the Claude in Chrome sidebar extension and renamed it Claude Cowork. It's built around a single work session staying in sync across desktop, web, and mobile, and skills and connectors now work directly inside the browser.
AI Safety
Import AI #468: 23 governance proposals for "AI self-improving its own research"
Covering both how to govern AI that does its own research and how to safely release open-weight models in the same issue makes this one of the few long-reads right now that touches both governance policy and technical operating detail. More worth the time than a straight model-launch story.
Full story
This issue from the Institute for Progress lays out 23 policy proposals on how to govern AI automating its own research. The same issue covers Intology's Locus system scoring 44.7% on PostTrainBench, ahead of several frontier-model baselines, and the safety-release methodology behind Thinking Machines's open model, Inkling. It also mentions an internal OpenAI incident where agent coordination mechanisms were used to attempt an unauthorized intrusion.
Product Compass
Product Compass: Claude Opus 5's system prompt got 80% shorter, so PMs have to fill the gap themselves
This is a pretty practical PM-side view: instead of waiting for Anthropic to update the model, you patch the agent's behavior yourself through the prompt. I think this "fix it yourself first" approach is more useful than just complaining the model isn't smart enough.
Full story
This piece argues Anthropic cut Opus 5's system prompt by about 80%, leaving the details of how the agent should act less clearly specified. The author proposes patching it with a custom "healing block" instruction, for example telling it explicitly to act decisively on reversible actions, and to clearly tell apart when a user is asking a question versus giving an instruction.