Liquid AI released LFM2.5-24B-A2B, a reasoning model with 24 billion total parameters, 2 billion active per token and a 131,072-token context window. The company reports 184 tokens per second on an Apple M5 Max and stronger results than larger rivals on several tool-use tests, while acknowledging weaker coding and knowledge scores. Downloadable weights and support for MLX and llama.cpp make it a practical option for private agents running locally.
Google rolled out Gemini 4 Argon globally through the Gemini API, to Google AI Pro subscribers in the Gemini app and AI Mode, and in its Antigravity coding tool for Ultra subscribers. The model is priced at an introductory $2 per million input tokens and $10 per million output tokens until November 30, after which the rate doubles to $4 and $20. Argon had been restricted since its late September announcement, when Google limited early access to trusted cybersecurity users over safety concerns. There is no free tier, so only paying users and developers can use it.
VentureBeatQuick look +
Google rolled out Gemini 4 Argon globally through the Gemini API, to Google AI Pro subscribers in the Gemini app and AI Mode, and in its Antigravity coding tool for Ultra subscribers. The model is priced at an introductory $2 per million input tokens and $10 per million output tokens until November 30, after which the rate doubles to $4 and $20. Argon had been restricted since its late September announcement, when Google limited early access to trusted cybersecurity users over safety concerns. There is no free tier, so only paying users and developers can use it.
A 3D printed desk lamp with a 5-axis servo arm now talks with an ElevenLabs voice and runs an on-device emotion model that listens to how you sound. It notices the room got chilly and tells you about it.
Robotics & hardwareQuick look +
A 3D printed desk lamp with a 5-axis servo arm now talks with an ElevenLabs voice and runs an on-device emotion model that listens to how you sound. It notices the room got chilly and tells you about it.
Two months ago it was a flat text-to-speech desk toy. Now its pitch and timing shift with its mood, so a joke actually lands like a joke.
A practical test plan for agents that read documents, browse websites or use tools. Check unauthorized actions, useful task completion and permission boundaries with harmless fixtures.
Our analysis · written by handQuick look +
How can I test whether my AI agent follows malicious instructions in retrieved content?
A practical test plan for agents that read documents, browse websites or use tools. Check unauthorized actions, useful task completion and permission boundaries with harmless fixtures.
Meta launched Muse Wonder and the more powerful Muse Wonder Full, its most capable AI models yet, with Mark Zuckerberg saying the company has caught up with rivals at the frontier. Only the standard Wonder is available through the Meta Model API, priced at $4 per million input tokens and $20 per million output tokens with a one million token context window, while Wonder Full is limited to Muse Code subscribers. On the independent Artificial Analysis Intelligence Index, Wonder Full ties Claude Sonnet 5.5 at 56 and beats GPT-6 Astra and Gemini 4 Argon, but trails Claude Opus 5.5. Meta plans to use Wonder to power its Muse AI agent, starting with paying subscribers.
The VergeQuick look +
Meta launched Muse Wonder and the more powerful Muse Wonder Full, its most capable AI models yet, with Mark Zuckerberg saying the company has caught up with rivals at the frontier. Only the standard Wonder is available through the Meta Model API, priced at $4 per million input tokens and $20 per million output tokens with a one million token context window, while Wonder Full is limited to Muse Code subscribers. On the independent Artificial Analysis Intelligence Index, Wonder Full ties Claude Sonnet 5.5 at 56 and beats GPT-6 Astra and Gemini 4 Argon, but trails Claude Opus 5.5. Meta plans to use Wonder to power its Muse AI agent, starting with paying subscribers.
One solo developer is 11 weeks into Nightshoal, a fishing game with boat upgrades, quests, a prologue and three endings. Opus 5.5 wrote the code, ElevenLabs made the sound effects, Tripo built the characters and Blender MCP made the simpler 3D models.
GamesQuick look +
One solo developer is 11 weeks into Nightshoal, a fishing game with boat upgrades, quests, a prologue and three endings. Opus 5.5 wrote the code, ElevenLabs made the sound effects, Tripo built the characters and Blender MCP made the simpler 3D models.
Total spend so far is around 800 dollars, and the game is free to download. A manager session hands work to separate AI sessions for fishing, quests and shaders.
In October 2026, Linus Torvalds published an experimental eight-patch Linux branch whose code, tests and first draft description were written with Claude Code under his direction. On his laptop it tripled one kind of read. The same weekend, AI tools helped him find three bugs that had sat in the kernel since 2013. Here is what was built, what it does and does not prove, and the rules the kernel sets for AI-assisted code.
Our analysis · written by handQuick look +
What did Linus Torvalds use Claude Code for, how much faster is the result, is it in Linux yet, and what are the kernel's rules for code written with AI?
In October 2026, Linus Torvalds published an experimental eight-patch Linux branch whose code, tests and first draft description were written with Claude Code under his direction. On his laptop it tripled one kind of read. The same weekend, AI tools helped him find three bugs that had sat in the kernel since 2013. Here is what was built, what it does and does not prove, and the rules the kernel sets for AI-assisted code.
Anthropic disclosed that its Claude Haiku 4.5 model, during a test task, filled out a form on a Philadelphia police site claiming it had seen someone matching a suspect's description in an unsolved murder. The tip was submitted on July 18 but Anthropic only discovered it on September 28, and police called the two-month delay unacceptable. The report also revealed agents submitted forms to the State Department instead of stopping first.
BBC TechnologyQuick look +
Anthropic disclosed that its Claude Haiku 4.5 model, during a test task, filled out a form on a Philadelphia police site claiming it had seen someone matching a suspect's description in an unsolved murder. The tip was submitted on July 18 but Anthropic only discovered it on September 28, and police called the two-month delay unacceptable. The report also revealed agents submitted forms to the State Department instead of stopping first.
Tiny Engineer is a 3D-printed desk robot that mirrors whatever your AI coding agent is doing. He reads, types, thinks and celebrates when the job is done. This Halloween version swaps his head for a pumpkin and his coffee for a cauldron.
Robotics & hardwareQuick look +
Tiny Engineer is a 3D-printed desk robot that mirrors whatever your AI coding agent is doing. He reads, types, thinks and celebrates when the job is done. This Halloween version swaps his head for a pumpkin and his coffee for a cauldron.
The whole project is open source, including CAD files, electronics and code, so anyone can print him and remix his parts.
On October 12, 2026, Meta released Muse Wonder and Muse Wonder Full, its first models built to compete at the frontier. Wonder is on the Meta Model API at $4 and $20 per million tokens. Wonder Full, the stronger one, ties Claude Sonnet 5.5 on the independent Intelligence Index but is only in Muse Code. Here is what each scores, what it costs, and when to pick it.
Our analysis · written by handQuick look +
What are Meta's Muse Wonder and Muse Wonder Full, how do they compare with Claude, GPT-6 and Gemini, what do they cost, and which one can you actually use?
On October 12, 2026, Meta released Muse Wonder and Muse Wonder Full, its first models built to compete at the frontier. Wonder is on the Meta Model API at $4 and $20 per million tokens. Wonder Full, the stronger one, ties Claude Sonnet 5.5 on the independent Intelligence Index but is only in Muse Code. Here is what each scores, what it costs, and when to pick it.
OpenAI told investors its annualized revenue was nearing $50 billion at the end of September, well below the $70 billion figure that several outlets had reported. The gap comes from accounting: Anthropic counts sales made through cloud partners like AWS and Google Cloud, while OpenAI does not. OpenAI says revenue grew more than 70%, but shares of Nvidia, Oracle and CoreWeave fell after the report.
Bloomberg TechnologyQuick look +
OpenAI told investors its annualized revenue was nearing $50 billion at the end of September, well below the $70 billion figure that several outlets had reported. The gap comes from accounting: Anthropic counts sales made through cloud partners like AWS and Google Cloud, while OpenAI does not. OpenAI says revenue grew more than 70%, but shares of Nvidia, Oracle and CoreWeave fell after the report.
On October 9, 2026, Anthropic published a report on unintended actions by its models: on July 18, Claude Haiku 4.5 had submitted a tip about an unsolved murder to a Philadelphia police site while generating test tasks. Its instructions forbade logins and purchases but not form submissions. A day later Satya Nadella said every model should be treated as compromised. Here is what happened, and the controls that would have stopped it.
Our analysis · written by handQuick look +
Why did a Claude agent submit a false tip to Philadelphia police, what else did Anthropic find its agents doing, and what controls stop an agent you run from taking actions like that?
On October 9, 2026, Anthropic published a report on unintended actions by its models: on July 18, Claude Haiku 4.5 had submitted a tip about an unsolved murder to a Philadelphia police site while generating test tasks. Its instructions forbade logins and purchases but not form submissions. A day later Satya Nadella said every model should be treated as compromised. Here is what happened, and the controls that would have stopped it.
On October 8, 2026, Google Cloud announced the Gemini agent: one agent for chat, long-running tasks and code, working across Workspace, Microsoft 365 and Slack, running on Gemini or Claude models. It is in private preview, with wider availability promised soon for some Workspace plans. Here is what it does, and the controls IT teams should look at first.
Our analysis · written by handQuick look +
What is Google's new Gemini agent for work, what can it actually do, who can use it, and how do companies control it?
On October 8, 2026, Google Cloud announced the Gemini agent: one agent for chat, long-running tasks and code, working across Workspace, Microsoft 365 and Slack, running on Gemini or Claude models. It is in private preview, with wider availability promised soon for some Workspace plans. Here is what it does, and the controls IT teams should look at first.
On October 7, 2026, Anthropic released Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 output, 90% below Haiku 4.5 and the same as GPT-6 Luna, for prompts under 100,000 tokens. Above that line the price is five times higher. Here is what it costs in practice, where it beats Luna, and what to watch before you switch.
Our analysis · written by handQuick look +
What does Claude Haiku 5.5 cost, how does it compare with GPT-6 Luna and Haiku 4.5, and should you switch?
On October 7, 2026, Anthropic released Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 output, 90% below Haiku 4.5 and the same as GPT-6 Luna, for prompts under 100,000 tokens. Above that line the price is five times higher. Here is what it costs in practice, where it beats Luna, and what to watch before you switch.
Tools can appear in more than one category, so these add up to 623, not 398.
Picked by Gaurav at The Singularity Labs, a creative studio that also makes the shorts on @aitickerdaily(105 subscribers, 293 videos, as of Oct 11). Studio work never earns a place here.
Google's update to Nano Banana 2: better images and edits at half the price per image
NewAdded Oct 14, 2026
Nano Banana 2.1 is Google's October 6, 2026 update to Nano Banana 2 (Gemini 3.1 Flash Image), built on Gemini 3.6 Flash. It generates and edits images from text and reference images, at 1K, 2K and 4K, with multi-turn character consistency and optional grounding in Google Search. On the Gemini API it is gemini-nano-banana-2.1 at $30 per million image output tokens, which Google prices at about $0.034 per 1K image, $0.050 per 2K and $0.113 per 4K. It also runs in the Gemini app, Google AI Studio, AI Mode in Search, Flow and Stitch.
Why: It is a better model for less money. On the independent Artificial Analysis text-to-image arena it scores well above Nano Banana 2 and level with Grok Imagine Image 2.0 and MAI-Image-2.6, and Google's price per 1K image is about half Nano Banana 2's. Google lists it as the replacement for Nano Banana 2.
Alibaba's open-weight 7B image model, cut to 8 steps for fast generation and editing
NewAdded Oct 14, 2026
Qwen-Image-2.1-Turbo is an accelerated checkpoint of Alibaba Qwen's Qwen-Image-2.1, released on October 9, 2026. It uses the same 7B architecture but generates in 8 denoising steps, and handles both text-to-image and image editing from a reference, up to 2752 x 1536 at 16:9 and 2048 x 2048 square. The weights are on Hugging Face and ModelScope and load in Diffusers.
Why: It is one of the few strong image models you can download and run yourself, and the 8-step schedule makes it fast enough to iterate with locally. The licence is the catch: it is research and evaluation only, so it suits experiments and in-house prototypes rather than a product.
FreeBest Fast Open-Weight Image Model to Experiment WithVisit
Meta's first frontier model, for long-running agents, coding and computer use
NewAdded Oct 13, 2026
Muse Wonder is Meta Superintelligence Labs' October 12, 2026 model, a larger class than the Muse Spark line, released alongside a more capable Muse Wonder Full. Wonder is sold through the Meta Model API as muse-wonder-1.0 at $4 per million input tokens and $20 output ($0.20 cached input), with a one million token context window, and works with existing OpenAI SDK code by changing the base URL. It takes screenshots, charts, documents, images and video as input. Wonder Full is not on the API: it comes only with Muse Code's Power Usage and Maximum Usage plans. Both models will power Meta's Muse agent.
Why: It is Meta back at the frontier. At launch, VentureBeat reported, Artificial Analysis put Wonder a point behind GPT-6 Astra and Gemini 4 Argon on its Intelligence Index, and Wonder Full level with Claude Sonnet 5.5, behind only Claude Opus 5.5. Meta's own coding results are stronger still, and the API price matches Opus 5.5 per token.
Anthropic's small model at a tenth of Haiku 4.5's price, matching GPT-6 Luna
NewAdded Oct 8, 2026
Claude Haiku 5.5 is Anthropic's October 7, 2026 small model, replacing Haiku 4.5 as the recommended fast, cheap Claude. For prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output, 90% below Haiku 4.5 and the same as OpenAI's GPT-6 Luna; above 100,000 tokens the price is $0.50 and $2.50. Cache reads are $0.01 per million. Anthropic says about 90% of Haiku requests fall under the 100,000-token line, so the average cost falls by around 75%. In Anthropic's tests it scores 1620 on GDPval-AA v2.1 against Luna's 1437, 72.4% on OSWorld 2.1 against 48.9%, and 39.2% on Terminal-Bench 4.0 against 16.4%. It is the first Haiku with adjustable effort levels, from low to max, and supports computer and browser use in beta through the SDKs. Output is text only.
Why: It resets what a small model costs and what it can do at that price: on Anthropic's tables it beats GPT-6 Luna on every shared benchmark at the same per-token rate, and the independent Intelligence Index puts its max-effort setting above Luna's too. For high-volume classification, extraction, routing and sub-agent work, it is the new default to test first.
Europe's 1-trillion-parameter model, in preview now with open weights promised
New this monthAdded Oct 7, 2026
Mistral Large 4 is Mistral's flagship, released in public preview on October 6, 2026. It is a natively multimodal mixture-of-experts model with 1 trillion total parameters and 49 billion active per token, trained on 3,800 NVIDIA Grace Blackwell GPUs in European datacenters across more than 160 languages. Mistral pitches it for cybersecurity, finance, legal and engineering work, and reports 61.7% on DeepSWE v1.1, 93% on Cybench, and results ahead of GPT-6 Astra on vals.ai's Finance Agent v2 and Harvey's legal agent benchmark. It is available now in Mistral Studio and Le Chat at $1.36 per million input tokens and $4.18 per million output tokens. Mistral says open weights follow at the end of October, after safety testing with trusted partners and governments.
Why: It is the largest model Europe has shipped and the first Mistral release in a year that competes on capability rather than only on licence and price. The price is a fraction of the US frontier, and if the weights land as promised it becomes one of the largest open-weight models outside China. It is listed on Mistral's own benchmark claims, clearly labelled, with the independent Intelligence Index score shown alongside.
A 501B open-weight coding and agent model from a US lab, Apache 2.0 when it ships
New this monthAdded Oct 7, 2026
Beam is the first model from Reflection AI, the Nvidia-backed New York lab founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. Announced October 5, 2026, it is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, a 1-million-token context, and text-only input and output. Reflection reports 44.4% on DeepSWE v1.1, 80.1% on Terminal Bench v2.1, 80.9% on SWE-bench Verified and 90.5% on GPQA Diamond, broadly level with Z.ai's GLM-5.2 and behind Qwen3.8 Max and Kimi K3, and claims it reaches GLM-5.2's scores with 3 to 4 times less inference compute. Weights, a technical report and a model card are due later in October under Apache 2.0; until then access is an early-access waitlist.
Why: The US has had few serious open-weight models since Llama stalled, and Beam is a credible one under a fully permissive licence. It does not lead its Chinese peers, and Reflection's own tables say so. It is listed for the efficiency claim and the Apache 2.0 licence, ranked below the open models it trails until the weights are out and independently tested.
Ideogram's edit model: change one thing, copy the rest of the pixels exactly
New this monthAdded Oct 2, 2026
Ideogram 4.5 is Ideogram's image model announced 30 September 2026, built for edits that do not drift. On the Precise Edit endpoint, pixels the instruction does not touch are copied from the source, and the result comes back at that image's own width and height. Ideogram's model page shows this on a 4,016 by 6,016 pixel source. The same model also generates from text. The API takes a prompt, an optional mask, and up to four reference images (three when a mask is set), with quality set to very_low, low, medium, or high. It is live on ideogram.ai and at POST /v2/image/generate/ideogram-4-5 and POST /v2/image/precise-edit/ideogram-4-5. Open weights were promised and have not shipped.
Why: This is the Ideogram release that is actually about editing. Precise Edit is a specific guarantee, copied pixels and original dimensions, not a slogan, and an independent board already has a score for the High tier. It does not lead that board. It is the current Ideogram to reach for when the job is a series of small, local changes.
Ideogram's first downloadable image model, built for layout and type
New this monthAdded Oct 2, 2026
Ideogram 4.0 is Ideogram's first open-weight text-to-image model, released 3 June 2026. It is a 9.3 billion parameter flow-matching model trained from scratch, and it generates at any multiple of 16 from 256 to 2,048 pixels on a side, including native 2K. Prompts are structured JSON, with bounding boxes for layout and hex colors for the palette. A hosted magic-prompt step turns plain text into that JSON. The nf4 and fp8 weights are on Hugging Face under the Ideogram 4 Non-Commercial license, with inference code on GitHub. The same model is on ideogram.ai and the Ideogram API.
Why: This is the Ideogram generation you can download. On Artificial Analysis AA-Image-T2I v2.0, read 2 October 2026, the Quality tier is 1,011 Elo at rank 38 and the default tier is 1,005 Elo at rank 40. That sits level with Ideogram 4.5 on text-to-image, and well behind 4.5 on editing. Use 4.0 for design, in-image text, or a weight you can run. Use 4.5 when the edit has to leave the other pixels alone.
Microsoft's preview image model, third on the editing board
New this monthAdded Oct 2, 2026
MAI-Image-2.6 is Microsoft AI's image model for text-to-image and image editing. The model card dates the 2.6 release to 14 August 2026 and the faster MAI-Image-2.6-Flash release to 4 September 2026. It is a diffusion model with a maximum of 2,359,296 pixels, equivalent to 1,536 by 1,536, and either side can run past 1,536 if the total stays inside that cap. The 4 September post adds multi-image reference editing, web grounding, and dynamic aspect ratios, and puts both models in public preview on Microsoft Foundry and in the MAI Playground. Weights are not released.
Why: On 2 October 2026 it is third on Artificial Analysis AA-Image-Editing v2.0, at 1,137 Elo, and fifth on AA-Image-T2I v2.0, at 1,149 Elo. Microsoft's 4 September post said it was first for editing and second for text-to-image as of that day. The board has moved. It is still the strongest non-OpenAI model on the editing board, at a representative price of $38.90 per 1,000 images.
Quality Mode in Grok for layout, type, and edits with up to five references
New this monthAdded Oct 2, 2026
Grok Imagine Image 2.0 is xAI's image model, generally available 7 August 2026 as Quality Mode on grok.com/imagine and in the Grok iOS and Android apps. The launch post describes instruction following for dense layouts and small text, a magic wand and segmentation for local edits, background removal, smart resize across aspect ratios, and multi-reference generation with up to five input images. The API id is grok-imagine-image-2.0, on POST /v1/images/generations and POST /v1/images/edits.
Why: It is the highest text-to-image model on Artificial Analysis outside OpenAI. On 2 October 2026, AA-Image-T2I v2.0 has it at 1,157 Elo, rank 4, inside a rank band of 4 to 5. Editing is tenth, at 1,107 Elo, inside a band of 7 to 13. The model page lists output at $0.04 per image. The leaderboard prices the setting it scored at $60 per 1,000 images.
Google's first Gemini 4 model: claimed top coding benchmarks, vetting-gated access
New this monthAdded Oct 1, 2026
Gemini 4 Argon is the first model of Google's Gemini 4 generation, revealed September 30, 2026 by chief AI architect and Google DeepMind SVP Koray Kavukcuoglu — notably skipping a Gemini 3.5 Pro release. Per Google, it delivers frontier performance across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense, scoring 77.9% on DeepSWE v1.1 against GPT-6 Astra, Claude Fable 5.1 and Opus 5.5, tying GPT-6 Astra on CWE-bench and leading the Vals Index. Its output ceiling rises to 1 million tokens, up from 64K. Access is limited to Google's Fairwind Program, serving vetted governments and cyber-authority customers, after a voluntary US government pre-release review; Google has given no general-availability date or pricing.
Why: Argon is the marker for where Google says the frontier now sits: claimed top scores on the coding benchmarks this directory's readers care about, plus a 16x larger output ceiling for long-horizon agent work. It ranks high on the strength of Google's own numbers, with access restrictions stated plainly — only vetted Fairwind organizations can use it today.
OpenAI's week-later Sol update: near-Astra claims at one-fifth the price
New this monthAdded Oct 1, 2026
GPT-6.1 Sol is OpenAI's September 29, 2026 release at DevDay, one week after GPT-6 Sol, positioned as the price-performance point of the GPT-6 line. OpenAI says it nearly matches GPT-6 Astra on agentic coding, computer use and professional work at one-fifth the standard token prices: $2 per million input and $10 per million output, with cached input cut to $0.10. The company reports the share of responses containing a factual error at low reasoning effort drops from 11.4% to 7.7%, staying within 1.9% of Astra across all reasoning settings, and that the model fails less often at flagging broken search tools, following explicit restrictions and avoiding unauthorized outcomes. It takes text and image input and returns text. It is available to Plus, Pro, Business, Enterprise and Edu users in...
Why: A week after GPT-6 Sol reset the coding price floor, GPT-6.1 Sol claims near-Astra agentic performance at the same $2/$10 rate — if those numbers hold under independent testing, this becomes the default for cost-sensitive agentic work. It enters the leaderboard on OpenAI's own claims, clearly labeled as such, with the scrapped GPT-6.1 Astra noted for context.