What earns an entry. A test we ran, a number we can stand behind, or a source we checked. Every entry says how we know. The most useful ones are usually where we were wrong earlier and are correcting it in public.
Week of 31 July 2026
Read the abstract, not the chart someone screenshotted
Discipline
Kimi K3 shipped as open weights: 2.8 trillion parameters, 104 billion active, native vision, a million tokens of context, downloadable and self-hostable. Genuinely a bigger deal than another price cut. But the benchmark chart circulating this week is the vendor's own, from the vendor's technical report, on the vendor's eval suite, and their abstract says outright that K3 still trails the strongest proprietary models. Every layer of commentary we saw dropped that sentence. This is the whole discipline in one example: the primary source was more honest than the summary of it, and it took two minutes to check.
How we know: Read Moonshot's published K3 technical report abstract directly. Architecture and context figures are from that document. We have not run our own benchmarks and have no independent view on the model.
A frontier model sold out, which breaks a planning assumption
Market
Moonshot reportedly stopped accepting new Kimi K3 subscribers rather than throttling existing ones, and paying users saw the service speed up the same day. The interesting part is not the model, it is the shape: demand for frontier coding models is outrunning the compute to serve them. Most planning in this category quietly assumes capability rises while price falls, smoothly and forever. Sold out is not what that curve looks like. If you are building a workflow that depends on one cheap model being available on demand, that is a dependency worth writing down before it bites.
How we know: Reported by practitioners, not verified by us.
Revenue up, headcount down, automation in the middle
Market
The pattern showing up in software results is growth alongside a smaller team, with automation named as the bridge. Monday.com is the example being passed around: roughly 25 percent year-on-year revenue growth with reduced headcount and more automation, and the market liked it. Treat the specific figures as reported rather than checked. The reason it matters here is not the company, it is that the market is now rewarding the shape, which changes what your CFO expects from a sales team next year. Nobody is going to ask whether AI made your reps happier.
How we know: Second-hand, not verified against filings. We are noting the pattern being rewarded, not certifying anyone's numbers.
The models got cheaper again, and it changed nothing about the call
Practice
Opus 5 reads a million tokens of context at five dollars per million in and twenty-five out. Reading every call your team runs is now a rounding error rather than a project. That is a real change and it makes one specific thing possible: patterns that only exist at volume, like a rep who monologues for twenty minutes, become visible for the first time. What it does not do is ask the uncomfortable question for anyone. Cheaper models widen what you can see. They do not widen what you are willing to say out loud on a Thursday afternoon.
How we know: Pricing checked against Anthropic's published model documentation, 31 July 2026. The claim about what it does not fix is ours, from our own calls.
Week of 27 July 2026
Two percent is a good cold reply rate, and pretending otherwise costs you
Outreach
Around 2 percent is a good reply rate on cold outreach to US prospects right now. Teams that benchmark against the 8 to 10 percent numbers from the old playbook conclude their copy is broken, respond by sending more, and make the underlying problem worse. The lever is not volume. It is whether the first line could have been sent to 500 companies.
How we know: Our own cross-account data plus what we see in customer pipelines. Treat any vendor quoting double digits as quoting a best case.
Invert the question and the forecast gets honest
Closing
Asking a model to assess a deal returns a balanced summary of nothing. Asking it to argue the deal will not close this quarter, using only what is in the record, surfaces the gap you were avoiding. Follow it with a request for the one missing fact that would settle the argument and the exact question to ask.
How we know: Run both prompts on the same deal record and compare. The difference is not subtle.
Models need explicit permission to return nothing
Prompting
Without a line granting it, a model will manufacture a reason to follow up, a trigger event that does not exist, or an expansion case for an account that is perfectly happy. Adding say so instead of inventing one is the single highest value edit to most sales prompts.
How we know: Give the same account with genuinely no news to a prompt with and without the clause.
Why we bother
Two kinds of people write about this: veterans who have sold through several supposed revolutions and are sceptical of all of them, and operators who adopt anything new the week it ships. Both are half right. The veteran is right that the fundamentals hold. The operator is right that the edge goes to whoever adopts early, and that it closes fast once everyone has the same tools.
Holding both positions at once means you have to keep testing, because the only way to know which new thing is real is to run it against a live pipeline and write down what happened. That is what this page is.
Whatever proves durable gets folded into the stage pages and the prompt library, so those stay current instead of becoming an archive. The log is the working notes. The pages are the conclusions.
Claude for Sales is independent and not affiliated with Anthropic. Full disclosure on the about page.