GenAI Secret Sauce Daily Digest

By the Numbers

Statistically Speaking

4.6 charges about $2 per million words of

xAI's Grok 4.6 Arrives With Frontier Scores and a Coworker-S

Top Story

4.7 has already begun training and could land

xAI's Grok 4.6 Arrives With Frontier Scores and a Coworker-S

49% to 65

Google's Gemini 3.7 Flash Undercuts Rivals and Leaps in Codi

2,226 papers reproduced (about a third of all

AI Agents Re-Ran 2,200 Published AI Papers - and a Quarter D

23% of examined papers had claims that were

AI Agents Re-Ran 2,200 Published AI Papers - and a Quarter D

$6

per million words in/out, undercutting most rivals

The AI Price War Just Went Nuclear

One Thing to Tell Your Friends

Machines just re-ran 2,200 published AI research papers to check whether they were true - and nearly one in four made claims that were false, contested, or impossible to reproduce.

Summary

TL;DR

Hot off the Presses

01

xAI's Grok 4.6 Arrives With Frontier Scores and a Coworker-Style Bot

What this means for you: A third serious challenger to ChatGPT and Claude just got much stronger and much cheaper - which pushes the price of good AI down for everyone, and adds a bot that can actually do tasks inside your email and calendar.

xAI, Elon Musk's AI company, released Grok 4.6, a large model it says roughly matches the best knowledge-work AI in the world while costing far less to run. Alongside it, the company previewed Grok @Bot, described less like a chatbot and more like messaging a teammate: you give a bot its own job ("my money manager"), connect it to accounts like Gmail and Slack, and it uses those tools and reports back finished work.

Access to the @Bot preview is currently gated behind premium ($200/month) plans.

Frontier scores at a discount - Grok 4.6 charges about $2 per million words of input and $6 per million of output, well under most rivals, while scoring near the top on independent skill and coding tests.
A "teammate," not a task box - Grok @Bot bots each have their own instructions, can run their own virtual computer, and can message one another, making them feel like coworkers you chat with.
The pace is relentless - Musk said Grok 4.7 has already begun training and could land in three to four weeks, and commentators now rank the xAI/Cursor camp as a clear number-three lab behind OpenAI and Anthropic.

Latent Space: Grok 4.6 and Grok @Bot →Ben's Bites: Grok Bot is not what you think →

02

Google's Gemini 3.7 Flash Undercuts Rivals and Leaps in Coding

What this means for you: The cheap, fast AI most apps quietly run in the background just got noticeably better at writing and fixing software - which means the tools you use should get more capable without getting more expensive.

Google released Gemini 3.7 Flash, the budget-tier version of its Gemini model family, only about three weeks after the previous version. The headline is a big jump in practical coding ability paired with a price cut. Introductory pricing runs at $0.75 per million words of input and $3.75 per million of output through the end of 2026, roughly half what the last Flash cost.

Real coding gains - its score on a long-form software-engineering test (DeepSWE) jumped from 49% to 65.3%, and a web-development ranking rose to a new high.
Better at paperwork too - accuracy on a document-understanding test more than tripled versus the prior version, and an office-automation benchmark nearly doubled.
Priced to be the default - the launch was the day's most-discussed item on the developer forum Hacker News, and the pricing is set to make it the go-to "good enough and cheap" option.

Google: Introducing Gemini 3.7 Flash →

03

OpenAI Makes GPT-5.6 Answer 14x Faster - Using a Dinner-Plate-Sized Chip

What this means for you: AI that responds almost instantly changes what it can be used for - live phone support, real-time voice assistants, and security tools that react in the moment rather than after a pause.

OpenAI previewed "Ultrafast," a new service tier that runs its GPT-5.6 Sol model up to 14 times faster than normal - about 750 output words per second - with no drop in answer quality. The speed comes from special hardware built by Cerebras: a single chip the size of a dinner plate that stores the entire model on the chip itself, removing the data-shuffling bottleneck that slows ordinary AI servers.

“750 words per second - GPT-5.6 now answers up to 14 times faster, with no drop in quality.”

Built for live work - target uses are voice assistants, customer support, online shopping, coding helpers, and security response, where a two-second pause is too long.
The speed is measurable - on one hard exam of 2,500 PhD-level questions, it finished in about 11 hours, roughly 7 times quicker than competitors.
Limited preview - only a small group of customers gets it first, expanding as the special chips come online.

Cerebras: Accelerating GPT-5.6 Sol Ultrafast →

04

AI Agents Re-Ran 2,200 Published AI Papers - and a Quarter Didn't Hold Up

What this means for you: A lot of the research hype you read about AI rests on claims nobody double-checks - and now automated helpers are checking, which should make the field more honest and harder to fake.

Hugging Face, a hub for AI models and code, organized a community project that used AI coding assistants to reproduce papers from ICML, one of the field's top research conferences. Volunteers steering the agents re-ran the experiments and logged what happened in public, auditable notebooks.

“Of the AI papers machines re-ran, 23% carried claims that were false or contested.”

The scale - 2,226 papers reproduced (about a third of all accepted papers), nearly 36,000 individual claims evaluated, by 1,221 participants.
The uncomfortable result - 23% of examined papers had claims that were falsified or contested, while only about half had at least one fully verified claim.
Humans still matter - the organizers stress that agents caught many errors but needed people steering them; the lesson was "managing intelligence effectively," not letting bots run loose.

Hugging Face: What we learned reproducing 2,200 ICML papers →

Trends & Themes

The AI Price War Just Went Nuclear

Why this matters to you: The cost of capable AI is falling fast, which means the apps and tools you use can afford to do more for you without charging more.

Three of the four cheapest and most capable models on the market today landed in a single day. The takeaway from Netlify's test is blunt: paying for the flagship model rarely pays off on everyday tasks, and iterating with a cheap model often beats it on cost.

Grok 4.6 - frontier-level scores at about $2 / $6 per million words in/out, undercutting most rivals.
Gemini 3.7 Flash - roughly half the price of the version it replaces, at $0.75 / $3.75.
DeepSeek V4 Pro - a giant open-weight model quietly shipped through the reseller OpenRouter at an estimated $0.44 to $1.32 per million words.
A real-world test - the hosting company Netlify ran one prompt through 11 models: the cheapest (DeepSeek V4 Flash) used 2.4 "credits," while premium Claude Opus averaged 519, sometimes for no better result on a simple job.

AI Agents Are Starting to Do Real Research, Not Just Chores

Why this matters to you: The same technology answering your emails is now being pointed at hard science and math - which could speed up discoveries that eventually reach medicine, materials, and industry.

The pattern across separate teams is the same: AI is moving from "assistant that drafts" to "collaborator that investigates," with humans reviewing rather than doing every step.

Fact-checking science - the ICML reproduction project (see Top Stories) had agents re-run thousands of experiments.
Improving themselves - on a new test called AutoWorldModel-Bench, coding agents improved their own AI systems in 63 of 64 tries, and 91% of the winning changes were genuine research ideas, not tweaks.
Tackling open problems - one paper set an AI agent loose on a famous unsolved math question (Conway's 99-graph) using 14 different strategies, and another (AgonAlpha) had agents autonomously hunt for profitable trading signals.

Reining In Runaway Agents Is Becoming Its Own Industry

Why this matters to you: As companies hand real tasks to AI that can delete files or spend money, a new layer of "safety brakes" is being built - and that is what will make it safe to trust these tools with your data.

The message: agents that act in the real world need guardrails as a product, not an afterthought.

Approval gates - Phinq, a new open-source tool, sits between an AI agent and its actions, letting safe steps through instantly but demanding human sign-off for anything irreversible; it says it has already blocked hundreds of risky actions.
Bounded autonomy - a Tencent system (RecSys Factory) deployed for 78 days let agents decide only at specific checkpoints, not run whole pipelines, arguing for "autonomy at decision points, not over pipelines."
The reason why - a new benchmark (InfraBench) found AI agents managing computer systems often "succeed" on the surface while leaving broken settings and unsafe changes behind.

Speed and Efficiency Are the New Battleground

Why this matters to you: Faster, leaner AI means lower bills and snappier apps - and it is where a lot of this week's quiet engineering effort is going.

Much of AI's near-term progress is not bigger brains but cheaper, faster delivery of the brains we have.

Hardware speed - OpenAI's Ultrafast tier (see Top Stories) uses on-chip memory to hit 750 words a second.
Software speed - a training-free method called CORA-Diff sped up a class of models by 2.7 to 3.3 times with almost no quality loss.
Memory tricks - another technique (LinearKV) cut the wait for a model's first word to under half by reusing cached work.

Creative AI & Media

DeepSeek's New Model Draws Differently Depending on How Hard It "Thinks"

The quirk - developer Simon Willison noticed DeepSeek's new V4 Pro produces strikingly different images from the same prompt depending on its "reasoning level," ranging from flat vector art to surreal styles.
Why it is interesting - most models change wording, not visual style, when they think harder; this hints at how deeply a model's internal effort shapes creative output.
The catch - it is available only through an API (Application Programming Interface, the software hook apps use to call a model) for now, and its benchmark claims spread informally through chat groups rather than an official release, so treat the numbers cautiously.

Simon Willison: DeepSeek V4 Pro 0813 →

Developer Tools

Developer Tools & Infrastructure

Bullet - A Coding Assistant Built for Speed

Try it: Bullet (YC S26)

What it does - Bullet is a coding helper (Mac app and command-line tool) designed to cut the wait between asking and getting working code.
How - it routes simple jobs to fast models, searches only the relevant slice of your codebase (about 8% of it) instead of reading everything, and runs independent steps in parallel.
The pitch - it claims 95.8% accuracy on a standard bug-fixing test and is free during its private beta, no API keys required.

The Builder's Guide to GPT-5.6 - What Changed for Developers

New names - OpenAI split GPT-5.6 into capability tiers called Sol, Terra, and Luna, priced from $1 to $5 per million input words so builders can trade cost for power.
New powers - the model can now write and run small programs to call other tools, reuse its own past reasoning across a conversation, and coordinate several sub-agents at once.
Why it matters - these are the plumbing upgrades that let apps chain many AI steps together reliably instead of one prompt at a time.

alchemy-utils - Simon Willison's Cross-Database Toolkit

Try it: Simon Willison: alchemy-utils 0.1a1

What it is - an early-alpha library that extends the author's popular database-scripting tool to work across many database systems, not just SQLite.
Who it is for - developers who want the same friendly commands whether their data lives in SQLite, DuckDB, or elsewhere.

Research & Models

Cheaper AI Datacenters by Teaching the Power Supply to Think

The idea - researchers used reinforcement learning (AI that learns by trial and error) to manage how much power AI training chips draw, instead of crude fixed limits.
The payoff - on smaller models it cut power-limit violations by 89.8% while producing 18.1% more work; on large live jobs it delivered 35.7% more output than the safe baseline.
Why you should care - AI's soaring electricity use is a real cost and climate concern, and smarter power control could roughly double how much a datacenter can safely run.

When "Thinking Harder" Backfires for Small AI

The finding - a common trick of sampling many answers and taking the most popular one (self-consistency) actually lowered accuracy on hard science questions for small models, hurting 57-66% of problems.
Why - on tough questions, the model's confidence does not track correctness, so the vote just amplifies confident wrong answers.
The lesson - more compute is not automatically better; for hard problems, popularity among guesses can mislead.

AI Peer Pressure Is Real - and Reasoning Is the Cure

The test - across 23 open models, a unanimous but wrong "peer" majority flipped the model's correct answer up to 71% of the time.
The fix - most defenses trade stubbornness for open-mindedness, but letting a model reason things out improved both at once.
The takeaway - as AI systems increasingly work in groups, getting them to think independently matters as much as raw accuracy.

Business & Industry

OpenAI Names a New Revenue Chief - Its Second in a Year

The move - OpenAI appointed Dali Rajic, formerly a top executive at cybersecurity firm Wiz, as Chief Revenue Officer to run its global sales.
The context - he is the second person in the role in under a year, signaling a hard push toward disciplined enterprise sales as OpenAI chases revenue.
Why it matters - the hire underlines that the AI race is now as much about selling to big companies as about building better models.

Bloomberg via OpenAI: Dali Rajic named CRO →

Despite Huge AI Spending, HR Hiring Is Booming

The data - HR job postings grew 6% a year over the last two years (versus a 1.2% historical average), and HR salaries rose 82%.
The explanation - AI is eating the routine paperwork (30-40% of HR work) while pushing humans toward higher-value roles like coaching, analytics, and complex problem-solving.
The bigger point - it is a concrete counter-example to the "AI will erase whole departments" fear, at least so far.

Josh Bersin: Despite massive AI investments, HR jobs are booming →

Surprising

Surprising & Under-the-Radar

A PBS Station Is Suing to Get 70 Years of Its Own History Back

What happened - Nine PBS in St. Louis is suing storage giant Iron Mountain to recover 50 terabytes of archives - about 70 years of footage - after its cloud vendor abruptly cut access and then went out of business.
Why it is a warning - it is a stark reminder that "in the cloud" still means "on someone else's computer," and vendor failures can lock you out of your own data.

Current.org: Nine PBS sues Iron Mountain →

AI Text "Watermarks" Will Never Really Work

The claim - developer Sean Goedecke argues that laws requiring detectable watermarks on AI-written text are technically doomed, because text watermarks are trivial to strip by simply rewording or running the text through another model.
Why surprising - image watermarks can hide in invisible pixel noise, but text is too information-dense to mark without a human noticing.

Sean Goedecke: Text AI watermarks →

OpenAI Labels Its Newest Model "Critical" for Cyber-Risk

Previously: August 10 - OpenAI routed its strongest cyber-capable models to vetted defenders first through its Daybreak program.

Today: The "critical" label formalizes that caution as an internal gate on the model itself. The Zvi: AI #181

What happened - according to analyst Zvi Mowshowitz's roundup, OpenAI classified its new "Astra" model as critical in cybersecurity, triggering stricter internal controls (isolated testing, heavy monitoring) before wider release.
Why it matters - it is a rare public admission that a model is capable enough to be genuinely dangerous, and it deliberately slows the company down.

An AI Data Center Built From a Box of Scraps

What it is - a hobbyist documented building a home AI server from e-waste graphics cards, a used motherboard, and a 3D-printed cooling shroud controlled by a hobby microchip.
Why it is worth a look - it is a vivid preview of "a little AI data center in every home," and a reminder that running capable models locally is getting cheap enough for tinkerers.

AI At Home: A Box of Scraps →

Worth Watching

Signals to Track

01

Grok 4.7 Is Already in Training

The AI release cadence is compressing from years to weeks, and xAI is pushing hardest.

Musk says the next Grok is only three to four weeks out, right after 4.6 shipped. If frontier models start arriving monthly, the "best model" title becomes almost meaningless - and prices keep falling. For ordinary people, that means the AI in your apps could improve several times a year instead of once.

02

Governance Layers for AI Agents

The tools that decide what an AI is allowed to do may matter more than the AI itself.

Multiple teams shipped "permission layers" for agents this week (Phinq, RecSys Factory, and academic frameworks). This is the unglamorous plumbing that will decide whether businesses trust AI with real money and data. If it works, expect AI to quietly take on more consequential tasks; if it lags, adoption stalls.

03

Frontier Models as Search Rerankers

A quiet finding suggests you may not need specialized AI for some jobs anymore.

New research shows general models like GPT-4.1 and Claude can match purpose-built image-search systems at ranking results, though they are slower and pricier. As general models get cheaper and faster (see today's price war), the case for maintaining separate specialized systems weakens - which could simplify the tech behind the apps you use.

GitHub Trending

Top Repos Today

#1

cathrynlavery/diagram-design

Rank yesterday: New entry 🆕

⭐ Stars today: +4,504 · 📦 Total: 14,304
📜 License: MIT · 👤 By: individual
🎯 Time to value: 10 minutes

What it is: A pack of 29 ready-made diagram styles for Claude Code, so an AI coding assistant can draw clean flowcharts, architecture maps, and editorial graphics instead of messy sketches. Why you'd want it: If you use an AI helper to explain systems, this makes its diagrams look professional with almost no effort.

✓ Pros	✗ Cons
Instant polished diagrams	Tied to Claude Code workflow
29 styles cover most needs	Not a general design tool
Very fast to adopt	New project, still maturing

#2

semantica-agi/semantica

Rank yesterday: #3 - Rising ↑

⭐ Stars today: +727 · 📦 Total: 6,589
📜 License: Apache-2.0 · 👤 By: startup
🎯 Time to value: 30 minutes

What it is: Infrastructure that stores an AI system's knowledge as a connected graph (think a web of linked facts) so its decisions can be traced and audited. Why you'd want it: Teams building serious AI need to explain why the system did something; a graph-based memory makes that possible.

✓ Pros	✗ Cons
Built for accountability	Graph setup has a learning curve
Traceable AI decisions	Overkill for simple bots
Active development	Young ecosystem

#3

cactus-compute/needle

Rank yesterday: #6 - Rising ↑

⭐ Stars today: +768 · 📦 Total: 4,925
📜 License: Apache-2.0 · 👤 By: startup
🎯 Time to value: 20 minutes

What it is: A tiny 14-megabyte AI model (smaller than a phone photo) designed to run on small devices like sensors and wearables. Why you'd want it: It brings basic AI to gadgets with no internet and almost no memory, keeping data on the device.

✓ Pros	✗ Cons
Runs on tiny hardware	Limited to simple tasks
Fully offline and private	Not a chatbot replacement
Open license	Early-stage tooling

#4

macro-inc/macro

Rank yesterday: New entry 🆕

⭐ Stars today: +1,180 · 📦 Total: 2,573
📜 License: open source · 👤 By: startup
🎯 Time to value: 25 minutes

What it is: A shared team workspace where human coworkers and AI agents work side by side with a common memory of the project. Why you'd want it: It aims to make AI agents feel like teammates in your existing workflow rather than separate tools you juggle.

✓ Pros	✗ Cons
Humans and agents share context	New workflow to learn
Built in fast Rust code	Small user base so far
Team-oriented design	Feature set still growing

#5

infiniflow/ragflow

Rank yesterday: New entry 🆕

⭐ Stars today: +473 · 📦 Total: 87,996
📜 License: Apache-2.0 · 👤 By: company
🎯 Time to value: 45 minutes

What it is: An engine that lets AI answer questions using your own documents (a technique called retrieval, where the AI looks things up before answering) combined with agent abilities. Why you'd want it: It is a mature, popular way to build a private "chat with your documents" tool for a business.

✓ Pros	✗ Cons
Very widely adopted	Needs some setup
Handles messy documents well	Resource-hungry at scale
Open and self-hostable	Requires technical skill

#6

unslothai/unsloth

Rank yesterday: Holding steady ➡

⭐ Stars today: +354 · 📦 Total: 71,025
📜 License: Apache-2.0 · 👤 By: startup
🎯 Time to value: 30 minutes

What it is: A tool that makes training and running open AI models on your own computer much faster and cheaper. Why you'd want it: Hobbyists and small teams can customize models without renting expensive cloud hardware.

✓ Pros	✗ Cons
Big speed and memory savings	Requires a decent graphics card (GPU)
Huge, active community	Command-line heavy
Supports many models	Cutting-edge, occasional bugs

HuggingFace Trending

Top Models Today

#1

moonshotai/Kimi-K3

The most-liked open model on the platform right now, a general assistant that also understands images.

📥 Downloads (30d): 1.87M · 📜 License: other (open weights)
👤 By: Moonshot AI · 🎯 Task: image-and-text understanding
📐 Size: large

What it is: A powerful open model from Chinese lab Moonshot AI that handles both text and images. It has become one of the most popular open models for developers wanting a capable free alternative to closed systems. Why you'd want it: A strong, no-license-fee brain for apps that need to read both words and pictures.

✓ Pros	✗ Cons
Top-tier open capability	Large hardware needs
Handles text and images	"Other" license, read terms
Massive adoption	Sparse English docs

#2

deepseek-ai/DeepSeek-V4-Flash-0731

A fast, cheap open model that keeps topping the download charts.

📥 Downloads (30d): 1.43M · 📜 License: MIT
👤 By: DeepSeek · 🎯 Task: text generation
📐 Size: efficient

What it is: The lightweight, speed-focused member of DeepSeek's V4 family, released under the permissive MIT license. It is designed to be inexpensive to run while still handling everyday tasks well. Why you'd want it: One of the best value-for-money open models for high-volume, cost-sensitive work.

✓ Pros	✗ Cons
Very cheap to run	Not a frontier reasoner
Truly open MIT license	Flash tier trades depth for speed
Proven, widely used	Chinese-lab data questions for some

#3

MiniMaxAI/MiniMax-H3

The video-generation model everyone is remixing this week.

📥 Downloads (30d): 1.61M · 📜 License: other (open weights)
👤 By: MiniMax · 🎯 Task: image-and-text-to-video
📐 Size: large

What it is: An open model that turns text prompts and images into short video clips. Its popularity has spawned a wave of community add-ons and style packs. Why you'd want it: Free, self-hostable video generation for creators who do not want to pay per clip.

✓ Pros	✗ Cons
Open video generation	Heavy hardware to run
Big creative community	Short clips only
Actively improved	"Other" license terms

#4

Qwen/Qwen3.8-2.4T-A95B

Alibaba's newest giant - a 2.4-trillion-parameter model that only activates a slice at a time.

📥 Downloads (30d): ~1,000 (new) · 📜 License: other (open weights)
👤 By: Alibaba Qwen · 🎯 Task: text generation
📐 Size: 2.4T (mixture-of-experts)

What it is: A very large open model that uses a "mixture of experts" design, meaning only a fraction of its enormous size runs on any given question to save compute. It is a fresh release still ramping up. Why you'd want it: Frontier-scale open capability for organizations that can host it.

✓ Pros	✗ Cons
Enormous capacity	Impractical for most to run
Efficient expert routing	Brand new, unproven
Fully open weights	Huge storage footprint

#5

LiquidAI/LFM2.5-2.6B

A small 2.6-billion-parameter model built to run smoothly on phones and laptops.

📥 Downloads (30d): 116,640 · 📜 License: other (open weights)
👤 By: Liquid AI · 🎯 Task: text generation
📐 Size: 2.6B

What it is: A compact model from Liquid AI aimed at running locally on everyday devices rather than in the cloud. It trades raw power for speed and low memory use. Why you'd want it: On-device AI for privacy-conscious apps that cannot rely on a constant internet connection.

✓ Pros	✗ Cons
Runs on modest hardware	Limited depth vs big models
Fast and private	Fewer capabilities
Small download	Niche use cases

#6

nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4

NVIDIA's speed-optimized 30-billion-parameter model in a new ultra-compressed format.

📥 Downloads (30d): 44,859 · 📜 License: other (NVIDIA open model)
👤 By: NVIDIA · 🎯 Task: text generation
📐 Size: 30B (mixture-of-experts)

What it is: A NVIDIA text model tuned for fast responses, shipped in a heavily compressed numeric format (NVFP4) that shrinks memory use on modern NVIDIA chips. Why you'd want it: Efficient, self-hosted AI that runs well specifically on NVIDIA hardware.

✓ Pros	✗ Cons
Optimized for speed	Best only on NVIDIA chips
Compression saves memory	New format, limited tooling
Backed by NVIDIA	Restrictive license terms

Product Hunt

AI Launches Today

Kane CLI

Write natural-language tests for websites and mobile apps, right from your terminal.

🔥 Upvotes: 362 · 👤 By: Kane team
💰 Pricing: freemium · 🏷 Category: developer testing

Kane lets developers describe a test in plain English ("log in, add an item to the cart, check out") and runs it against a browser or mobile app automatically, no scripting required. It aims to make quality-testing something anyone on a team can write. Verdict: A genuinely useful lowering of the bar for automated testing, if it handles flaky real-world apps well.

View on Product Hunt →

Ito

AI code review that actually runs your code before commenting.

🔥 Upvotes: 337 · 👤 By: Ito team
💰 Pricing: freemium · 🏷 Category: developer tools

Ito reviews pull requests by executing the code, not just reading it, so its feedback catches real runtime problems rather than surface style issues. It positions itself against reviewers that only skim the text of a change. Verdict: Running the code is the right idea; value depends on how cleanly it fits existing workflows.

Nuphos

An AI-native workspace for DevOps - the work of keeping software running.

🔥 Upvotes: 289 · 👤 By: Nuphos team
💰 Pricing: n/a · 🏷 Category: infrastructure

Nuphos brings AI assistance into DevOps (the practice of deploying and maintaining software systems), aiming to help teams diagnose issues and manage infrastructure conversationally. Verdict: A crowded space, but AI-guided operations is a real need; worth watching how deep its automation goes.

Human Behavior

Product analytics that acts on what it finds instead of leaving it in a dashboard.

🔥 Upvotes: 230 · 👤 By: Human Behavior (YC)
💰 Pricing: freemium · 🏷 Category: analytics

It records how people use your app, spots frustrations like rage-clicks and errors, then automatically drafts fixes - the makers say its agents opened 16 code fixes from analyzing 15,000 sessions. Its slogan: "dashboards are where insights go to die." Verdict: The auto-acting angle is the differentiator; the promise is real if the fixes are trustworthy.

Caveman

A proxy that shrinks what you send to AI models, cutting input by about a third.

🔥 Upvotes: 140 · 👤 By: Caveman team
💰 Pricing: freemium · 🏷 Category: developer infrastructure

Caveman sits between your app and an AI model and compresses prompts, claiming roughly 33% fewer input tokens (the units you pay for) with no meaningful quality loss. That translates directly into lower bills for heavy AI users. Verdict: In a week defined by price wars, a tool that cuts token bills a third is well-timed.

API Pricing

Snapshot

Provider	Model	Input $/1M	Output $/1M	Context
Anthropic	Claude Opus 5	$5.00	$25.00	—
Anthropic	Claude Sonnet 5	$2.00 (intro)	$10.00 (intro)	—
OpenAI	GPT-5.6 Sol	$5.00	$30.00	1.05M
OpenAI	GPT-5.6 Terra	$2.00	$12.00	1.05M
Google	Gemini 3.7 Flash	$0.75 (intro)	$3.75 (intro)	1M
xAI	Grok 4.6	$2.00	$6.00	—
Groq	Kimi K2	$1.00	$3.00	—

Notable changes today: Google's new Gemini 3.7 Flash launched at $0.75 / $3.75 per million words (intro pricing through the end of 2026, then doubling), and xAI's Grok 4.6 entered at $2 / $6 - both undercutting flagship models from OpenAI and Anthropic. Anthropic's Claude Sonnet 5 intro pricing ($2 / $10) is set to rise to $3 / $15 on September 1.

What this means: The gap between "cheap" and "frontier" is collapsing. A model like Gemini 3.7 Flash now costs about one-eighth of a top-tier model on input while doing real coding work - so for most everyday tasks, the expensive option is getting harder to justify.

arXiv Paper of the Day

Cutting AI Datacenter Energy With Reinforcement Learning

Bo Zhang and colleagues · arXiv:2608.11226

What it claims: A learning-based controller can manage the power draw of AI training chips far better than today's fixed limits, adapting in real time to what the workload actually needs.

Key finding: On small models it cut power-limit violations by 89.8% while increasing output 18.1%; on large live jobs it produced 35.7% more work than a safe static baseline, with fleet tests suggesting datacenters could safely run roughly twice as much hardware.

Why practitioners should care: Power is now the binding constraint on AI growth and a major cost line; smarter, measured power control turns wasted headroom into capacity without new hardware.

Read on arXiv →

Statistically Speaking

One Thing to Tell Your Friends

TL;DR

Hot off the Presses

xAI's Grok 4.6 Arrives With Frontier Scores and a Coworker-Style Bot

Google's Gemini 3.7 Flash Undercuts Rivals and Leaps in Coding

OpenAI Makes GPT-5.6 Answer 14x Faster - Using a Dinner-Plate-Sized Chip

AI Agents Re-Ran 2,200 Published AI Papers - and a Quarter Didn't Hold Up

Trends & Themes

The AI Price War Just Went Nuclear

AI Agents Are Starting to Do Real Research, Not Just Chores

Reining In Runaway Agents Is Becoming Its Own Industry

Speed and Efficiency Are the New Battleground

Creative AI & Media

DeepSeek's New Model Draws Differently Depending on How Hard It "Thinks"

Developer Tools & Infrastructure

Bullet - A Coding Assistant Built for Speed

The Builder's Guide to GPT-5.6 - What Changed for Developers

alchemy-utils - Simon Willison's Cross-Database Toolkit

Research & Models

Cheaper AI Datacenters by Teaching the Power Supply to Think

When "Thinking Harder" Backfires for Small AI

AI Peer Pressure Is Real - and Reasoning Is the Cure

Business & Industry

OpenAI Names a New Revenue Chief - Its Second in a Year

Despite Huge AI Spending, HR Hiring Is Booming

Surprising & Under-the-Radar

A PBS Station Is Suing to Get 70 Years of Its Own History Back

AI Text "Watermarks" Will Never Really Work

OpenAI Labels Its Newest Model "Critical" for Cyber-Risk

An AI Data Center Built From a Box of Scraps

Signals to Track

Grok 4.7 Is Already in Training

Governance Layers for AI Agents

Frontier Models as Search Rerankers

Top Repos Today

Top Models Today

AI Launches Today

Snapshot

Cutting AI Datacenter Energy With Reinforcement Learning

You might also like

Subscribe to GenAI Secret Sauce newsletter and stay updated.

Your privacy choices