Every AI app worth knowing about. One place. 🤖
Joined September 2024
- Tweets251
- Following32
- Followers12
- Likes2
⏳ Agents that run all day just got cheaper context. @OpenAIDevs added a 12-hour prompt cache option to the API, so shared instructions and tools stay at the cached rate even when calls are hours apart.
• Pick 12 hours or keep the 30-minute default, and mix both in one request
• Works in the Responses API on GPT-5.6 and later models
• Cache reads cost the same under either lifetime
• Catch: a 12-hour write costs 2.5x the normal input price, double the 30-minute rate, so it only pays off with repeat use
Price out one agent's daily prompt before switching every request to the long cache.
💻 Perplexity's agent no longer needs the cloud to do the work. @perplexity_ai made Computer run fully local on Apple silicon Macs, with both the model and the agent on the device.
• With permission, it works across local files and Mac apps like Notes, Messages, Mail, and Numbers
• Local model choices are PPLX Pro, a post-trained Qwen, or stock Qwen
• Hybrid mode stays available, sending research and reasoning to cloud models while private files stay put
• Catch: Pro, Max, or Enterprise plan required, and 36 GB of unified memory is the recommended floor
Give it one chore built on private files, like a budget update pulled from Notes and Mail, and see whether the cloud gets missed.
🐰 Space Bunny has a real name. @MiniMax_AI confirmed the stealth model that topped OpenRouter and OpenCode usage charts was an early M3.1 Flash, and the final build is out.
• API price is $0.15 in and $0.60 out per million tokens, flat across a 1M context
• MiniMax claims 3x the speed of M3 at about a third of the cost on agent work
• Free for a limited time on OpenRouter, OpenCode, and Cline
• Catch: weights are promised for Oct 23 with no license posted yet, and the benchmark scores are MiniMax's own
Point it at one long agent task during the free window and compare the bill with what runs now.
MiniMax M3.1 Flash: Space Performance + Bunny Speed
🚀🐰 Yes, Space Bunny - which has topped global leaderboards with 200T+ tokens - was an earlier version of M3.1 Flash.
💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence across key tasks such as Cowork (Automation-Bench: 62%; ALE: 30%), Coding (DeepSWE: 74%) and Visual Reasoning (BabyVision: 91%), at 1/3 the cost of M3 and 3x the speed.
API pricing (flat rate): $0.15 / $0.60 per 1M input/output tokens, with cache reads at $0.015 / 1M tokens.
Model weights will arrive on Oct 23.
API: platform.minimax.io
M Plan: minimax.io/m-plan
MiniMax Code for desktop: agent.minimax.io/download
🛡️ Free security audits from Anthropic's strongest models are open to the projects the rest of the internet runs on.
@ClaudeDevs is putting OSS Scanner in front of developers: an opt-in service that scans critical open-source repos on a schedule, at no cost.
• Each finding comes with a proof-of-concept, plus a candidate patch where possible
• Maintainers enroll with a pull request holding a config file and a Dockerfile, and scans run with no internet access
• Anthropic's own spot check: 85 of 97 high and critical findings met its disclosure bar
• Catch: reports skip human review, and only established, security-critical projects qualify, case by case
Best first users are maintainers who already keep up with verified bug reports.
We launched OSS Scanner, which gives critical open source projects free, periodic vulnerability scans from our strongest models.
Each finding includes a proof-of-concept, and a patch where possible.
red.anthropic.com/oss-scanne…
🧩 One prompt, several agents, each working its own slice of the task.
@OpenAI upgraded GPT-6 Astra Pro in ChatGPT so a big job can split across parallel agents inside a single chat.
• Follow each agent's progress live and open finished pieces to check the work
• Faster responses and steadier plugin and tool use in the same update
• Lives in the Chat tab on the $100 and $200 Pro plans
• Catch: still rolling out, and GPT-6 Pro runs against a weekly usage limit
Hand it a report with independent sections and review each piece before merging.
GPT-6 Astra Pro in ChatGPT got an upgrade with faster responses and improved use of plugins and tools.
And now, multiple agents can tackle different parts of a task in parallel. You can follow each agent’s progress and open completed tasks to review their work.
Now available to Pro users in the Chat tab.
🖥️ Cloud agents can now leave a live dashboard behind, not just a diff.
@cursor_ai shipped interactive canvases from cloud agents, plus an Environments API in beta.
• Ask for a dashboard or report while the agent works, then share the finished canvas as a link
• Environments API creates and manages cloud agent setups, including which repos they clone
• Environments hold up to 500 repos, and large multi-repo ones start 5x faster, per Cursor
• Catch: the Environments API is still beta, and canvas shares need a paid team plan with the right privacy mode
Spin up one cloud agent on a real report and send the canvas link.
⚡ Same model name, same price, better scores on every benchmark in the table.
@deepseek_ai swapped V4.1 Flash for the 1017 checkpoint across its website, app, and API, with open weights under MIT the same morning.
• DeepSWE 77.0, up from 74.2; Terminal-Bench 42.3, up from 31.2
• Shorter outputs and faster runs, so the same task costs less
• Search and Fetch APIs now open to all at $0.75 per 1,000 calls
• Catch: every score is DeepSeek's own table, and Opus 5 still leads Terminal-Bench
Rerun one agent task from last week and compare the bill.
🚀 DeepSeek-V4.1-Flash-1017 is here!
Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API.
🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.
1/4
🎙️ A voice model that already runs Starlink's support calls is now in the API.
@SpaceXSI launched Grok Voice 3.0, a speech-to-speech model for voice agents.
• $0.08 a minute, the same price as 2.0
• Ranks first on Artificial Analysis's speech-to-speech index at 86.9
• Responds in under a second, and matches GPT-Live-1 Astra on a support-call benchmark at 40% lower latency
• Starlink runs 20,000 calls a day on it, roughly 70% resolved with no human
• Catch: the rival comparison predates GPT-Live-1.1, and the call numbers are SpaceX's own
Run one real support script through it and listen for where it stalls.
💻 Meta put the weights of an 816B model up for download, and it runs on two Mac Studios.
@AIatMeta released Muse Spark Open, built for coding and agent work.
• 816B total and 42B active parameters, a 1M-token window, and text, image, video, and audio input
• Weights on Hugging Face, plus hosted APIs on OpenRouter, Together AI, Fireworks, Groq, and others
• Meta calls it the most capable US open-weight model to date
• Catch: it ships under Meta's own license with limited commercial terms, and the 4-bit build still needs about 700 GB of GPU memory
Try it on a hosted API first, then decide if the desk setup is worth it.
Today we’re releasing Muse Spark Open, the most capable US open-weight model to date.
Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
Key Model Specs:
→ 816B total / 42B active parameters
→ 1M token context window
→ Natively multimodal across text, image, video, audio
🛹 After the ramp, one more thing. @Tesla's Roadster lifts off with no ramp at all and flies level over a parked semi and a gold car on vapor jets.
A kickflip over a row of cars. Silence, then the crowd erupts.
The wings run on the Optimus actuator system, and the clip is past 10 million views on X.
🚀 @Tesla just drove the Roadster up a ramp, barrel-rolled it on gas jets and stuck the landing in front of a crowd.
This is not AI, folks. It's a live demo.
Tesla says the thrusters push 4-5x the force the tires can. Hard to believe.
Everyone is jaw dropped.
🚪 Trying an agent no longer starts with pasting in an API key.
@NousResearch shipped Hermes Agent v0.22.0, and a fresh desktop install now opens on a guided setup chat.
• Free trial model for first-time desktop installs, so no key is needed to start
• Connectors act inside apps like Linear and Notion after one browser sign-in
• One-click local models that run on your own machine, with no account
• Catch: the free model is once per OS user and desktop only
Install it, take the setup tour, and give it one real errand before touching settings.
🆓 DeepSeek V4.1 Flash now costs nothing inside OpenCode.
@opencode made the model free to use, for a limited time.
• Input, output, and cached reads all billed at zero
• Reasoning, tool calls, and image input all work
• Provider keeps no request data and does not train on it
• Catch: free version tops out at 200K context and 32K output, against 1M on the paid one, and no end date is announced
Hand it one real bug while the price is zero.
🧾 A five-person lab says a cheap model can finish Office work like the big ones.
@bertgodel's Paper Instruments launched Ultramarine 1 Flash, a post-trained GLM-5.3 Flash built for finance, law, and consulting tasks.
• Hosted API today at $0.15 per million input tokens and $0.50 output
• Built to hand back finished Excel, Word, and PowerPoint files
• Weights promised in two weeks, so open-source is still a promise
• Catch: the 60x cheaper than Fable comparison rests on the company's own benchmark
Run one real memo or model through it and judge the file, not the chart.
We are releasing Ultramarine 1 Flash, a frontier-class model built for knowledge work. It is open-source and achieves similar performance to Fable 5, while being 60x cheaper.
It was post-trained by a team of 5 for ~$300k all-in. Available on API now, with weights releasing in 2 weeks.
Two launches, one budget question: which calls deserve a frontier model at all?
Google opened Gemini 4 Argon to the world at half price until December. ScaleDown shipped tiny single-job models that label, extract, and summarize for a nickel per million tokens, with output free.
Best use: split one agent workflow in two. Send the long reasoning to the big model, send the tagging and pulling to the small one, and compare the invoice. The gap says where the money was leaking.
✂️ Most agent calls are a label, a field, or a summary, and they finally have their own models.
@neal_k_patel and ScaleDown launched Constrained Decoding Models, built for one job each.
• Classify, Extract, and Summarize skip token-by-token writing entirely
• Up to 1M tokens of text and images per input
• $0.05 per million input tokens, and output is free
• Catch: 200x faster and 467x cheaper are the company's own numbers
Swap one routing or tagging call in an agent and compare the bill.
You already gave the LLM the answer. So why is it writing it out from scratch?
Introducing a new AI model architecture: Constrained Decoding Models (CDMs)
LLMs were built to write. Most production workloads just need an answer: fast, consistent, and affordable.
CDMs are purpose-built for 1 task each. They read your input once and skip the token-by-token generative loop that makes LLMs slow and expensive.
Not just a decision model, but a family of 3 CDMs live today:
• Classify: define your labels and get back a calibrated probability + reason for each. Used for routing, tagging, and intent detection.
• Extract: describe the entities in natural language and get back well-formed JSON with confidence scores + citations. Used for web scraping, document processing, and PII redaction.
• Summarize: pass in long text, images, or PDFs and get back a consistent summary under 1,000 tokens. Used for context compaction, metadata generation, and transcript digests.
Built for production:
• 1M tokens of multimodal (text + image) input, 31x larger than Jev
• 77.2% on DecisionBench vs Jev's 72%
• 200x faster, 467x cheaper than frontier LLM APIs
• $0.05/M input tokens. Output is free (yes, even for summarize)
Live on your favorite platforms today:
@OpenRouter @vercel_dev @LiteLLM @LangChain @TrustedRouter @itsArthurAI @useRouterPlus
Get 50M input tokens for free (for the first 50 hours) at link below 👀
🧪 Google's new frontier model is out, and the discount has a deadline.
@GeminiApp is rolling out Gemini 4 Argon worldwide, in the app and the API.
• Gemini app and AI Mode for Google AI Pro and Ultra subscribers
• API at $2 per million input tokens and $10 output, intro pricing
• One 1M-token window shared by input, thinking, and output
• Catch: prices double on December 1, and the free API tier is left out
Run the hard jobs on it now, while the meter is half price.
Three launches, one pattern: AI is getting priced and judged by the job, not the model.
Z.ai trained a small model into a legal specialist and let outside graders score it. Cursor joined The X Subscription, so Grok rides the plan and every other model bills at API rates. Meta shipped Muse Wonder for agents that remember context and keep working.
Best use: pick one real task, run it on the bundled default, and keep a list of what needed a specialist or a memory. That list says what is worth paying for.
🧾 Cursor now comes bundled with a social network.
@cursor_ai joined The X Subscription, with one usage pool shared across Cursor, Grok Build, and the rest of Grok.
• Plus at $8 runs Grok models only, Premium starts at $30
• Twice the coding usage on Grok models versus the matching old Cursor plan
• Allowance refreshes weekly, not monthly
• Catch: third-party models now bill separately at API list price
Anyone living on Claude or GPT inside Cursor should price a normal week before switching.
Cursor is now part of The X Subscription.
Current Cursor subscribers can keep their plan or switch to The X Subscription.
Paid plans share one usage pool across Cursor, Grok Build, and other Grok products, with twice the coding usage on Grok models as the equivalent Cursor plan. Premium, Super, and Ultra also include Grok Bot.
To make room for more Grok usage, third-party models in Cursor move to API pricing on the new plans. We'll keep working to make your usage go further, with more compute, better inference, and more token-efficient models.
The X Subscription is available at cursor.com/pricing.
🍉 Meta is back at the frontier, and the pitch is memory, not chat.
@AIatMeta launched Muse Wonder and Muse Wonder Full, its most capable models yet.
• Muse Wonder is live today in the Meta Model API and Muse Code
• Muse Wonder Full is for Muse Code subscribers only
• Active Memory carries context across conversations; Active Focus keeps working and checks in before big actions
• Catch: the Muse agent upgrade is "soon," and the full safety report is still pending
Give it a long, messy task and see whether it remembers.
We’re excited to introduce the Muse Wonder series with Muse Wonder and Muse Wonder Full, our most capable models yet. Both are frontier models, built to deliver state-of-the-art performance for personal and professional agentic tasks.
Muse Wonder is available today via the Meta Model API and Muse Code. Muse Wonder Full is available today for Muse Code subscribers. Muse Wonder will also soon power experiences in our personal agent product, Muse, rolling out first to subscribers.