Watch today's digest as a video summary (generated by NotebookLM)
By the Numbers
Statistically Speaking
97
supported languages during a single conversation and
Google's new AI can watch, listen, and talk back in real tim
2
launched at 64% cheaper than a frontier
Cheap, small, open models keep crashing the party
53%
to 69% without extra cost
"Passing the test" is quietly the wrong metric for AI agents
One Thing to Tell Your Friends
One Thing to Tell Your Friends
Chinese AI labs secretly funneled more than 151 million real user conversations through Claude to train their own models - by quietly rerouting their own customers' chats through it without telling them.
Summary
TL;DR
Top Stories
Trends
The referees are moving inside the labs they judge, Whether to "slow down" AI has become an open public fight, and Cheap, small, open models keep crashing the party.
Creative AI
Describe a building in plain English and get a real, editable 3D model.
Dev Tools
Inside a real "AI software factory" and An enterprise HR brain you can plug into any AI assistant.
Research
A 7-billion, An AI that aced a task once will not always do it again, and Can large language model (LLM) "judges" grade expert work? Partly..
Business
A dating and Voice AI is now a benchmark battleground.
Education
Google is running a free "badgeathon" for teachers, Trusted, specialized AI is coming to corporate learning, and A case for "tapestry over hustle" in the AI era.
Surprising
An AI trained on a railroad board game got better at finance, AI models show consistent "personalities" under pressure, and LLM agents beat specialized robots-of-the.
Worth Watching
AI that keeps a lab's knowledge after the people leave, One rulebook to define "recursive self, and Carbon.
GitHub
Leading repos: alibaba/open-code (+2,751), JustVugg/colibri (+2,035), and debpalash/VoiceStudio (+2,081).
HuggingFace
Leading models: deepseek-ai/DeepSeek-V4.1 (326,000), Qwen/Qwen3.8 (7.7M), and Lightricks/LTX (1.58M).
Product Hunt
Top launches: Resurf (295), Visiby (140), and Epilude Notetaker (109).
API Pricing
What this means: No material price changes among the flagship text models versus yesterday - Google's Gemini 3.1 Pro remains the value leader at $2 input, and Groq's open-model hosting stays an order of magnitude cheaper for lighter work.
arXiv
ZGCM-1 — Results competitive with Qwen3-235B and GLM-5.1 on math and agentic-search benchmarks, plus about a 4.2x improvement in early-training time-to-target.
FYI
Hot off the Presses
01
Google's new AI can watch, listen, and talk back in real time - in 97 languages
What this means for you: The voice assistant on your phone is about to feel less like a menu and more like a person who can see your screen, follow you across languages, and keep talking while it thinks.
Google released two real-time voice models on September 15: Gemini 3.8 Live, tuned to be cheap and fast, and Gemini 3.8 Live Extended Thinking, tuned for harder, multi-step problems. Both can take in what a camera sees in near real time and run background tools (like looking something up) without pausing the conversation. The "Extended Thinking" version can reason and speak at the same time, using natural filler like "Let me check that..." so it does not go silent mid-answer.
- #1 on the leaderboard - Extended Thinking topped Artificial Analysis' Speech-to-Speech Quality Index (a ranking of how good voice AIs sound and answer) with a score of 82.6.
- Mid-sentence language switching - it detects when you change among 97 supported languages during a single conversation and follows along automatically.
- Already inside Google's apps - live now via the Gemini application programming interface (API) and Google AI Studio, and wired into Search Live, Docs, Gmail, Keep, and the Gemini app.
- Everything it says is watermarked - all generated audio carries SynthID marking so it can later be flagged as AI-made.
02
The big AI labs just agreed on who gets to grade their homework
What this means for you: The companies racing to build powerful AI are, for the first time, agreeing to a shared rulebook for the outsiders who check their safety work - a small but real step toward independent oversight.
xAI, OpenAI, and Anthropic all endorsed AEF-1, a new baseline standard for third-party AI evaluators (the independent groups paid to probe models for danger before release). The standard spells out how evaluators handle conflicts of interest, funding ties, when they must step aside, and what they have to disclose. It was published by the AI Evaluator Forum, a body formed in December 2025.
- Why a standard is needed - outside auditors are often funded by the very labs they inspect, so the rules focus on keeping that relationship honest.
- Anthropic went further - it said it will embed external evaluators with office space, laptops, and desk access on par with its own internal risk teams.
- The open worry - one researcher quoted in coverage warned future models could "appear aligned under evaluation while hiding misalignment," which is exactly what independent testing is meant to catch.
03
The "AI misuse" report has a second story: labs allegedly copying each other at scale
What this means for you: Some of the "new" cheap AI models on the market may have been trained partly on a rival's answers - and the fight over who copied whom is becoming the industry's next big conflict.
Previously: September 10 - Anthropic published a threat report and said it disrupted attempts to use Claude for biological-weapons research and a network of roughly 70 fake news sites.
Today: Follow-up reporting and analysis this week put the spotlight on a different part of the same report - "distillation," where competitors harvest a top model's answers to train a cheaper copy. Anthropic says it detected and cut off large-scale distillation tied to seven China-based labs, and the numbers are the eye-opener.
“More than 151 million Claude exchanges, peaking at nearly 3 million per day.”
- 151 million and counting - Anthropic attributed more than 151 million Claude exchanges to Alibaba between May and July 2026, peaking near 3 million per day from over 3,500 accounts it calls fraudulent.
- Routing real customers through Claude - Anthropic alleges Moonshot and DeepSeek passed their own users' live chats through Claude and kept the responses as training data.
- A crowded list - the disrupted distillation campaigns were attributed to Alibaba, DeepSeek, Moonshot, Xiaomi, Zhipu, SenseTime, and MiniMax.
04
A startup is teaching AI to work by making it play board games
What this means for you: One promising path to more capable AI assistants is not a bigger model, but smarter training - and it can come from unexpected places like a 19th-century railroad game.
Good Start Labs, a $3.6 million-funded startup, trains AI inside game environments to see whether game skills carry over to real work. Its headline finding: how you train matters far more than which game you use. When it trained a mid-size model on "1830" (a cutthroat railroad-tycoon board game) as a hands-on agent that adapts turn by turn and uses tools, the model got better at outside financial-research tests. A simpler training setup only made it better at the game itself.
- Two real transfers so far - the railroad game improved finance-research scores, and training on the strategy game Diplomacy produced better customer-support behavior.
- Models have "personalities" - in Diplomacy tests, one lab's model planned betrayals while Claude Opus 4 refused to deceive.
- The honest caveat - broad generalization is still unproven; two cases is a signal, not a law.
Trends & Themes
Trends & Themes
The referees are moving inside the labs they judge
Why this matters to you: Whether AI is safe increasingly depends on outsiders who can see what the labs are doing - and this week they got both a rulebook and a desk in the building.
The pattern this week is oversight getting more formal and more embedded at the same time. That is progress, but it also ties the auditors closer to the audited - the exact conflict the new standard is trying to manage.
- A shared standard - xAI, OpenAI, and Anthropic all cosigned AEF-1 for third-party evaluators.
- Physical access - Anthropic said it will give external evaluators laptops, offices, and internal-team-level access.
- The catch, in the labs' own words - researchers warn a model can look safe during a test and still hide problems, so access has to be deep, not cosmetic.
Whether to "slow down" AI has become an open public fight
Why this matters to you: The people building AI are now arguing in public about whether they are going too fast, and the outcome shapes how quickly these tools land in your job and your apps.
This debate ran through several sources today, from newsletters to a widely shared departure post. The takeaway is that the disagreement is no longer polite or private - it is now a named, public split among the field's most powerful people.
- Slow-down camp - Anthropic's Dario Amodei argues labs should "slow down long enough for safety reasons," and Sam Altman broadly agrees.
- Full-speed camp - Jensen Huang and US AI policy voices frame slowing as optional, not required.
- Fuel on the fire - a former lab researcher's public resignation warning that labs are "gambling with our lives" went viral this week, and at least one current researcher agreed.
Cheap, small, open models keep crashing the party
Why this matters to you: The best AI is getting dramatically cheaper, which means more of it ends up free or nearly free in the tools you already use.
The throughline: capability is leaking downward to small, open, tool-using models. For everyday users, that means fewer paywalls and more "good enough" AI running cheaply, or even on your own machine.
- Tiny but mighty - a new fully open 7-billion-parameter model, ZGCM-1, reports math and search results competitive with models 30x its size by leaning on tools instead of memorized facts.
- Price war on coding models - Cognition's SWE-2 launched at 64% cheaper than a frontier model while staying near the top on coding tasks.
- Open models dominate the download charts - Hugging Face's trending list is full of open releases like DeepSeek, Qwen, and MiniCPM pulling millions of downloads.
"Passing the test" is quietly the wrong metric for AI agents
Why this matters to you: An AI that works in a demo can still fail when you actually rely on it - and the industry is starting to measure that gap.
Three separate research threads today - IBM's consistency work, an "agent iteration" framework, and a paper on agents recovering from their own mistakes - all point the same way: reliability, not raw smarts, is the frontier that decides whether agents are trustworthy.
“A top AI agent passed 77% of the time on average, but nailed the same task on all five tries only 53% of the time.”
- The gap, measured - IBM found a top agent averaged a 77.4% pass rate but succeeded on all five tries for only 53% of tasks.
- Why it happens - when the model's next-word choice is a near-tie, tiny technical differences between runs can flip the outcome.
- A cheap fix helped - applying consistency guidelines lifted "succeeds every time" from 53% to 69% without extra cost.
Creative AI & Media
Creative AI & Media
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
Business & Industry
Business & Industry
Education
GenAI in Education
Surprising
Surprising & Under-the-Radar
An AI trained on a railroad board game got better at finance
Good Start Labs found that training a model as a hands-on agent inside the 1830s tycoon game "1830" improved its scores on unrelated financial-research tests - evidence that how you train can matter more than what you train on.
AI models show consistent "personalities" under pressure
In strategy-game tests, some models planned betrayals while Claude Opus 4 refused to deceive - a reminder that different models behave differently in the same high-stakes spot, not just at different skill levels.
LLM agents beat specialized robots-of-the-mind when conditions change
A physical-task study found general LLM agents matched purpose-built reinforcement-learning systems in steady conditions but clearly beat them when the environment shifted - suggesting reasoning may generalize better than narrow training when the world is unpredictable.
Debate: is publishing "who misused our AI" oversight or geopolitics?
Anthropic's naming of specific foreign labs and actors is praised as transparency by some and questioned by others as a move that could inflame US-China tensions during ongoing AI talks. Both readings can be true at once.
Worth Watching
Signals to Track
01
AI that keeps a lab's knowledge after the people leave
Why this is worth watching right now: it targets the quiet crisis of expertise walking out the door when staff move on.
A new system called LabAgent customizes an AI agent to a specific research group's methods and records how to fix errors, so work can continue after key people depart. It reportedly outranked general commercial agents across several life-science tasks and reproduced a published result. If it holds up, small teams could stop losing years of hard-won know-how to turnover.
02
One rulebook to define "recursive self-improvement"
Why this is worth watching right now: the field is trying to formally define AI that improves itself before such systems arrive.
A new framework, Generalized Agent Iteration, puts ordinary AI training and "AI that rewrites itself" on the same map, giving researchers shared language to spot where self-improving systems could go wrong. It is theory today, but it is the kind of groundwork that shapes future safety rules for autonomous AI.
03
Carbon-aware AI routing
Why this is worth watching right now: it quietly turns "which data center answers you" into a climate decision.
Research on carbon-aware routing picks where an AI request runs based on the cleanliness of the local power grid at that moment. As AI's energy use grows, this kind of behind-the-scenes routing could cut emissions without users noticing any difference in speed.
GitHub Trending
Top Repos Today
Rank yesterday: #2 - Rising ↑
⭐ Stars today: +2,751 · 📦 Total: 28,423
📜 License: Apache-2.0 · 👤 By: Alibaba (corporate)
🎯 Time to value: 20 minutes
📜 License: Apache-2.0 · 👤 By: Alibaba (corporate)
🎯 Time to value: 20 minutes
What it is: An open-source tool that uses AI to automatically review code changes, built and battle-tested inside Alibaba. It plugs into your code workflow and flags issues before a human looks. Why you'd want it: It brings big-company-grade automated review to any team for free, cutting the time reviewers spend on routine checks.
| ✓ Pros | ✗ Cons |
|---|---|
| Proven at large scale | Tuned to Alibaba's conventions |
| Fast and low-dependency | Needs an LLM provider to run |
| Active, fast-growing project | Newer, so docs still maturing |
GitHub - alibaba/open-code-review: Fast, efficient, battle-tested at Alibaba’s scale. Hybrid architecture code review tool: deterministic pipelines + LLM Agent, precise line-level comments, built-in multi-language ruleset (NPE, thread-safety, XSS, SQL injection), OpenAI & Anthropic compatible.
Fast, efficient, battle-tested at Alibaba's scale. Hybrid architecture code review tool: deterministic pipelines + LLM Agent, precise line-level comments, built-in multi-language ruleset (NPE,…
GitHub
Rank yesterday: #1 - Falling ↓
⭐ Stars today: +2,035 · 📦 Total: 33,765
📜 License: MIT · 👤 By: Independent developer
🎯 Time to value: 30 minutes
📜 License: MIT · 👤 By: Independent developer
🎯 Time to value: 30 minutes
What it is: A lightweight tool for running large "mixture-of-experts" AI models (a design where only part of the model runs per query) on hardware you already own. It keeps dependencies minimal. Why you'd want it: It lets you run frontier-style models locally without a specialized server or heavy setup.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs big models on modest hardware | Command-line, not beginner-friendly |
| Very lightweight | Performance varies by machine |
| Permissive MIT license | Setup takes some patience |
GitHub - JustVugg/colibri: Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦 - JustVugg/colibri
GitHub
Rank yesterday: #3 - Holding steady ➡
⭐ Stars today: +2,081 · 📦 Total: 30,877
📜 License: MIT · 👤 By: Independent developer
🎯 Time to value: 15 minutes
📜 License: MIT · 👤 By: Independent developer
🎯 Time to value: 15 minutes
What it is: A local, free alternative to paid voice-cloning services, offering voice cloning and audio creation across 646 languages, all on your own computer. Why you'd want it: You get studio-style voice tools without a subscription or sending your voice to a cloud service.
| ✓ Pros | ✗ Cons |
|---|---|
| No subscription, runs locally | Voice cloning raises consent issues |
| Huge language coverage | Needs a capable machine |
| Simple to get started | Quality varies by voice |
GitHub - debpalash/VoiceStudio: VoiceStudio is the open-source, fully-local ElevenLabs alternative — voice cloning, voice design, video dubbing, dictation, transcription & audiobook creation in 646 languages.
VoiceStudio is the open-source, fully-local ElevenLabs alternative — voice cloning, voice design, video dubbing, dictation, transcription & audiobook creation in 646 languages. - debpalash/Voic…
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +593 · 📦 Total: 3,299
📜 License: Apache-2.0 · 👤 By: alphaXiv (startup)
🎯 Time to value: 25 minutes
📜 License: Apache-2.0 · 👤 By: alphaXiv (startup)
🎯 Time to value: 25 minutes
What it is: A toolkit that turns coding agents into research agents, helping AI read papers and run literature-style investigations instead of just writing code. Why you'd want it: It repurposes tools you may already use for coding into an AI research assistant.
| ✓ Pros | ✗ Cons |
|---|---|
| Builds on existing agent tools | Early-stage project |
| Focused, useful niche | Research quality still varies |
| Open Apache license | Small community so far |
GitHub - alphaXiv/OpenResearch: Turn your coding agents into research agents
Turn your coding agents into research agents. Contribute to alphaXiv/OpenResearch development by creating an account on GitHub.
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +437 · 📦 Total: 105,666
📜 License: MIT · 👤 By: earendil-works (startup)
🎯 Time to value: 20 minutes
📜 License: MIT · 👤 By: earendil-works (startup)
🎯 Time to value: 20 minutes
What it is: A broad AI-agent toolkit with one unified way to call many AI providers, plus a command-line tool for building and running agents. Why you'd want it: It saves you from wiring up each AI provider separately when building your own assistant.
| ✓ Pros | ✗ Cons |
|---|---|
| One API for many providers | Large surface area to learn |
| Big, active community | Broad scope can feel heavy |
| CLI included | Fast changes between versions |
GitHub - earendil-works/pi: AI agent toolkit: unified LLM API, agent loop, TUI, coding agent CLI
AI agent toolkit: unified LLM API, agent loop, TUI, coding agent CLI - earendil-works/pi
GitHub
Rank yesterday: Holding steady ➡
⭐ Stars today: +205 · 📦 Total: 2,798
📜 License: MIT · 👤 By: Independent developer
🎯 Time to value: 30 minutes
📜 License: MIT · 👤 By: Independent developer
🎯 Time to value: 30 minutes
What it is: An open-source sales system with built-in AI agents and WhatsApp integration, aimed at small teams that want automation without enterprise pricing. Why you'd want it: It bundles customer management and AI outreach in one free, self-hosted package.
| ✓ Pros | ✗ Cons |
|---|---|
| AI agents built in | Self-hosting takes effort |
| WhatsApp integration | Smaller ecosystem |
| Free and open | Still maturing |
GitHub - melgarafael/DeskcommCRM: Open-source AI sales OS — self-hosted CRM with native AI agents + WhatsApp (WAHA). Open alternative to Kommo, Octadesk & Intercom for any business that sells by chat. MCP-ready, multi-tenant, LGPD.
Open-source AI sales OS — self-hosted CRM with native AI agents + WhatsApp (WAHA). Open alternative to Kommo, Octadesk & Intercom for any business that sells by chat. MCP-ready, multi-tenant, L…
GitHub
HuggingFace Trending
Top Models Today
A fast, low-cost version of DeepSeek's model that handles both text and images, topping the trending charts on strong price-performance.
📥 Downloads (30d): 326,000 · 📜 License: MIT
👤 By: DeepSeek · 🎯 Task: Image-Text-to-Text
📐 Size: 763B (mixture-of-experts)
👤 By: DeepSeek · 🎯 Task: Image-Text-to-Text
📐 Size: 763B (mixture-of-experts)
What it is: A "Flash" (speed-optimized) model from Chinese lab DeepSeek that reads both pictures and text and answers cheaply. It uses a mixture-of-experts design so only part of it runs per query. Why you'd want it: Frontier-level answers at a fraction of typical cost, with an open MIT license.
| ✓ Pros | ✗ Cons |
|---|---|
| Very cheap per task | Huge total size to host |
| Handles text and images | Best run via a provider |
| Open license | From a lab named in misuse reports |
deepseek-ai/DeepSeek-V4.1-Flash · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Alibaba's mid-size open model that reads text and images, pulling millions of downloads as a practical workhorse.
📥 Downloads (30d): 7.7M · 📜 License: Apache-2.0
👤 By: Alibaba (Qwen) · 🎯 Task: Image-Text-to-Text
📐 Size: 27B
👤 By: Alibaba (Qwen) · 🎯 Task: Image-Text-to-Text
📐 Size: 27B
What it is: A 27-billion-parameter open model that handles text and images at a size many people can actually run. It is a general-purpose assistant model. Why you'd want it: A strong, permissively licensed all-rounder that fits on serious consumer or single-server hardware.
| ✓ Pros | ✗ Cons |
|---|---|
| Massive real-world adoption | 27B still needs a strong graphics card (GPU) |
| Text and image input | General, not specialized |
| Apache license | Competition is fierce |
Qwen/Qwen3.8-27B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
A popular open image-to-video model that animates still pictures into short clips.
📥 Downloads (30d): 1.58M · 📜 License: Open (LTX community)
👤 By: Lightricks · 🎯 Task: Image-to-Video
📐 Size: not published
👤 By: Lightricks · 🎯 Task: Image-to-Video
📐 Size: not published
What it is: A model that turns a still image into a short moving clip, widely used by creators. It is one of the most downloaded open video models. Why you'd want it: Free, local video generation for social clips and prototypes without a subscription.
| ✓ Pros | ✗ Cons |
|---|---|
| Very widely used | Video gen is compute-heavy |
| Turns photos into clips | Short clips only |
| Open community license | Quality varies by input |
Lightricks/LTX-2.5 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
A tiny 3B-class model built to run on phones and laptops while staying capable.
📥 Downloads (30d): 272,000 · 📜 License: Apache-2.0
👤 By: OpenBMB · 🎯 Task: Text Generation
📐 Size: 3B
👤 By: OpenBMB · 🎯 Task: Text Generation
📐 Size: 3B
What it is: A small text model designed to run on everyday devices, part of the push to put useful AI on-device. It trades some raw power for portability. Why you'd want it: Private, offline AI that runs on hardware you already own.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs on phones/laptops | Less capable than big models |
| Fast and private | Short on world knowledge |
| Open license | Best for lighter tasks |
openbmb/MiniCPM5-2B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
A new mid-size open model climbing the charts as a fresh general-purpose option.
📥 Downloads (30d): 17,900 · 📜 License: Apache-2.0
👤 By: Edge0 · 🎯 Task: Text Generation
📐 Size: 35B (mixture-of-experts, ~3B active)
👤 By: Edge0 · 🎯 Task: Text Generation
📐 Size: 35B (mixture-of-experts, ~3B active)
What it is: A 35-billion-parameter model that only activates about 3B at a time, keeping it efficient to run. It is a preview release gaining early traction. Why you'd want it: Bigger-model quality at closer to small-model running cost, thanks to the mixture-of-experts design.
| ✓ Pros | ✗ Cons |
|---|---|
| Efficient to run for its size | Preview, not final |
| Open license | Small track record |
| Rising quickly | Limited documentation |
Edge0/Edge0-35B-A3B-preview · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Product Hunt
AI Launches Today
Personal context library for Mac
🔥 Upvotes: 295 · 👤 By: Resurf
💰 Pricing: freemium · 🏷 Category: productivity / on-device AI
💰 Pricing: freemium · 🏷 Category: productivity / on-device AI
Resurf builds a private, on-device library of your context on macOS so AI tools can draw on what you have read and worked on without shipping it all to the cloud. It targets people who want personalized AI help while keeping their data local. Verdict: The clear standout of the day, and a bet that private, on-device context is the next battleground. Product Hunt
SEO tracking for AI-powered search engines
🔥 Upvotes: 140 · 👤 By: Visiby
💰 Pricing: freemium · 🏷 Category: marketing / analytics
💰 Pricing: freemium · 🏷 Category: marketing / analytics
Visiby tracks how your brand shows up inside AI answer engines, not just traditional search results, as more people get information from chatbots instead of blue links. It helps marketers see whether AI assistants mention or recommend them. Verdict: Useful and well-timed as "being cited by the AI" becomes the new SEO. Product Hunt
Privacy-focused meeting notes
🔥 Upvotes: 109 · 👤 By: Epilude
💰 Pricing: freemium · 🏷 Category: productivity
💰 Pricing: freemium · 🏷 Category: productivity
Epilude records and summarizes meetings with a privacy-first design, aimed at people who like AI note-takers but not the idea of every call living on someone else's server. Verdict: A crowded space, but privacy is a real differentiator worth watching. Product Hunt
API Pricing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Anthropic | Claude Opus 5 | $5.00 | $25.00 | 200k |
| OpenAI | GPT-5.6 (flagship) | $5.00 (promo $4.00) | $30.00 (promo $20.00) | not published |
| Gemini 3.1 Pro (preview) | $2.00 | $12.00 | 1M | |
| Groq | GPT-OSS 20B | $0.075 | $0.30 | not published |
What this means: No material price changes among the flagship text models versus yesterday - Google's Gemini 3.1 Pro remains the value leader at $2 input, and Groq's open-model hosting stays an order of magnitude cheaper for lighter work. Notable today: Google launched its Gemini 3.8 Live voice models, which are priced separately from these text APIs; real-time voice pricing is a new line item to watch as that market heats up. Batch processing and prompt caching still cut effective rates by roughly half at most providers. (OpenAI and Groq figures come from third-party pricing trackers, as their pages were not directly readable.)
arXiv Paper of the Day
ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search
Open-source team · arXiv:2609.13356
What it claims: A fully open 7-billion-parameter model can match models roughly 30x larger on math reasoning and agentic search by pairing deliberate internal reasoning with active tool use, rather than trying to memorize the web.
Key finding: Results competitive with Qwen3-235B and GLM-5.1 on math and agentic-search benchmarks, plus about a 4.2x improvement in early-training time-to-target.
Why practitioners should care: It ships weights, code, datasets, and training logs, giving teams a genuinely reproducible recipe for a small, tool-using model they can run and adapt cheaply.
Key finding: Results competitive with Qwen3-235B and GLM-5.1 on math and agentic-search benchmarks, plus about a 4.2x improvement in early-training time-to-target.
Why practitioners should care: It ships weights, code, datasets, and training logs, giving teams a genuinely reproducible recipe for a small, tool-using model they can run and adapt cheaply.
×
Click anywhere or press ESC to close
Member discussion