Gemini's cheaper agents, Claude Tag's 65% PRs, and a new prompting rule

The short read

Models and deployment

Google DeepMind: three Flash models aimed at different bottlenecks

• What happened: Google announced Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. The 3.6 post says it uses fewer tokens than 3.5 Flash at the same cost; the accompanying announcement reports 17% fewer output tokens on the Artificial Analysis Index. 1 2
• Why it matters: The lineup separates three production concerns: general agent quality, high-throughput low-latency work, and security-specific vulnerability finding and patching. Flash-Lite is listed at 350 output tokens per second, with pricing of $0.30 per 1 million input tokens and $2.50 per 1 million output tokens. 2
• What to watch: Gemini 3.5 Flash Cyber will be available only to governments and trusted partners through CodeMender in a limited pilot, so the cyber model's rollout is being treated as a controlled deployment rather than a general API launch. Open the original announcement.

Developer workflow

Simon Willison: Claude Tag is already carrying a large share of product PRs

• What happened: In a transcript of his conversation with Cat Wu and Thariq Shihipar from Anthropic's Claude Code team, Willison reports that Claude Tag, Anthropic's collaborative Slack agent, lands 65% of the product-engineering PRs for the Claude Code team. The figure is for that team, not all of Anthropic. 3 4
• Why it matters: Claude Tag is described as proactive and multiplayer: it can monitor a Slack channel for bug reports, open a pull request, tag the engineer who last changed the relevant code, and retain team preferences. 4
• What to watch: This is a workflow adoption claim from an internal team, not an independent productivity study. The useful question is whether the same pattern works outside a tightly instrumented product group. Read the original post or read the full transcript.

Prompting and model behavior

Simon Willison and Ethan Mollick: stronger models change the prompting advice

• What happened: Willison says Claude Code's system prompt was recently reduced by 80% for its newest models. The accompanying discussion says examples and long lists of prohibitions can over-constrain Fable and Opus, while Mollick describes the change as a major shift in how people should think about getting work from more capable systems. 5 4 6
• Why it matters: The lesson is model-specific, not a universal ban on examples. Anthropic's team says it now uses different system prompts for different models and tests edge cases where an instruction is mostly true but harmful when applied literally. 4
• What to watch: Prompting advice is moving from "add more instructions" toward giving the model enough context, fewer hard constraints, and evals that catch misinterpretation. Mollick also notes that clear studies on prompting these stronger systems are still scarce. Read Willison's post and Mollick's post.

Human-computer interaction

Andrej Karpathy: use a voice ramble to give the model the missing context

• What happened: Karpathy describes a workflow built around a roughly ten-minute voice session: switch to voice, speak in an unstructured stream of consciousness, and sometimes turn it into a short interview with the model. 7
• Why it matters: His claim is practical rather than measured: a model can reconstruct a messy explanation into something clearer, which reduces the number of corrections needed to establish the user's intent. 7
• What to watch: The method shifts effort from composing a perfect prompt to supplying more raw context. It is a personal workflow report, so its value depends on how well a particular voice-recognition and model stack preserves the speaker's meaning. Read the original post.

Society and ethics

Ethan Mollick: model releases need shared safety testing standards

• What happened: Mollick proposed US-China cooperation on common testing and acceptance standards for new models, with the goal of making safety certification more transparent for both open and closed systems. 8
• Why it matters: The proposal treats release certification as a coordination problem. A shared testing vocabulary would at least make it easier to compare safety claims across model types, even if the parties disagree on policy or deployment.
• What to watch: This is a policy suggestion, not evidence of an agreement or an existing standard. The post gives no details about the Chinese government material it references, so the claim should be read as a call for cooperation rather than a report of a new bilateral process. Read the original post.