BOTS FIRED

TUESDAY, OCTOBER 20, 2026

Mistral and MiniMax expand models as agents reach Macs, businesses and robots.

AI releases are moving across speech, local computers, business workflows and physical systems. This edition pairs newly available tools with research that tests their limits, from code migration and demographic fairness to protein binding. Deals, proposed power capacity and phased rollouts add scale, while access and production reliability remain separate questions.

🤖 FRONTIER MODELS

Mistral launches Voxtral 3 speech models and a voice-agent SDK

Mistral introduced Voxtral 3 on October 20, expanding its speech models and launching a voice-agent SDK. The family covers transcription and speech generation, while the SDK links listening, reasoning, tool use and speaking in a streaming agent workflow. Mistral offers the models through an EU-hosted API and self-deployment options. According to its Voxtral 3 announcement, Voxtral Small Transcribe 3 handles more than 50 languages; the text-to-speech offering covers more than 35 languages and more than 50 dialects. Those coverage figures apply to different models and tasks, so they should not be read as one universal capability. Mistral also reports low latency and favorable speech benchmarks, but those comparisons are company measurements. The practical change is a packaged route for developers to assemble voice agents that can hear a request, act through tools and respond aloud, with deployment choices that may matter to organizations handling regional data requirements.

Inception releases Mercury 3 diffusion language model through APIs

Inception released Mercury 3 on October 19, bringing its diffusion-based language model to commercial APIs. The model is available through Inception's own platform as well as OpenRouter, Baseten and Vercel, according to the launch announcement. Inception reports a one-million-token context window and generation at 1,068 tokens per second on NVIDIA GPUs. It says the new model is more than 40 times larger than Mercury 1 and is designed for applications where each response delay compounds, including browser agents, live tutoring and customer support. The company also reports quality and speed comparisons against other fast models, alongside examples from early-access partners. Those are vendor and partner results under particular tests, not independent evidence that every workload will be faster or better. Mercury 3 is an API launch, with no model-weight release claimed in the announcement. Its significance is a larger test of whether diffusion language models can serve demanding interactive tasks at production speeds.

OpenAI tests coding models on COBOL and other low-resource languages

OpenAI published two coding evaluations for low-resource programming languages on October 19. MultiLangBench asks models to build a SQLite-compatible database engine across 20 languages. OpenAI reports that GPT-6 Astra passed 87.4% of its tests on average at maximum reasoning effort, compared with 26.8% for GPT-5.6 Sol. That setup tests sustained implementation rather than a short syntax quiz. A second evaluation checks whether models can migrate COBOL programs while preserving their observable behavior. There, GPT-6.1 Sol reached 86.6% accuracy at low reasoning effort and 97.9% at higher effort, according to OpenAI. The company is releasing the MultiLangBench setup and test suite so developers can inspect the tasks. Its low-resource coding report describes progress on code that still supports essential services. These are company-designed evaluations, not measured success rates for production migrations. Teams modernizing legacy systems still need domain experts, representative tests and checks for business rules that may exist only in code.

Skyfall introduces SKY-001 for CAD and building-plan design

Skyfall introduced its SKY-001 engineering world model on October 19. The company says it breaks complex designs into hierarchies of assemblies, parts and construction steps. Its article demonstrates two bounded tasks: turning a product image into a CAD construction sequence and turning a building outline into BIM floor plans and rooms. Skyfall reports stronger speed and lower cost than several frontier language models on its own CAD and BIM benchmarks, alongside comparable or higher shape-overlap scores. Those comparisons reflect company-run tests, not independent evidence of broad engineering capability. The SKY-001 release describes the model's visual planning method and benchmark design. Access is currently through a waitlist, so readers should not assume a public API or production deployment. The work offers a different approach to geometry-heavy design, but manufacturability, safety and performance across unfamiliar engineering tasks remain unproven by the published demonstrations.

Viggle releases JST-2 for scenes, characters and 3D motion

Viggle released JST-2 on October 19 for generating 3D scenes, characters and motion. The company describes a shared representation of persistent entities in three-dimensional space, with an encoder, generator and decoder that can produce new views or changes in motion. JST-2 can make a scene from visual input, create an animation-ready character from an image, and generate full-body motion from text or video. Viggle says a scene takes about one second of model inference on an NVIDIA B200; that figure excludes network latency and does not describe every workflow. Its JST-2 launch article includes company-run comparisons against other character tools and links to a creation experience. Those results should be read as vendor benchmarks. The article lists text-guided scene expansion and finer-detail output as upcoming work. For creators, the immediate test is whether one model can keep a scene, character and motion coherent through the export steps their projects require.

MiniMax rolls out M3.1 Flash across API, Code and M Plan

MiniMax officially rolled out M3.1 Flash on October 19. Its announcement says the model is available through the MiniMax API, M Plan and MiniMax Code, following an earlier teaser that called it forthcoming. MiniMax identifies Space Bunny, an anonymously operated earlier model, as an early version of M3.1 Flash. The company reports gains on agent, coding, computer-use and security evaluations, along with faster inference and lower task cost than its predecessor. Those numbers are company measurements under particular benchmarks and settings, not independent results for every workload. The M3.1 Flash release lists API prices of $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. It says model weights and a Pro variant will arrive later. Developers can now compare the available hosted model on representative tasks, while treating promised releases and benchmark generalization as open questions.

Cloudflare opens Clairic extraction model in Workers AI preview

Cloudflare made Clairic available as a hosted preview on Workers AI on October 19. The model is designed to pull typed fields from unstructured text and optional images, compute values, and return bounded estimates or short forecasts. Developers specify fields and types, then receive values with confidence measures rather than generated prose. The Cloudflare introduction presents Clairic as an early checkpoint at @cf/cloudflare/clairic and says open weights are planned, not released. Cloudflare reports favorable speed and cost comparisons with general language models, but those are its own tests. The company also warns that longer chains of arithmetic are difficult, forecasting is basic, and confidence intervals may be poorly calibrated on unfamiliar data. A correctly typed response can still be wrong. For builders processing invoices or other repetitive records, the preview offers a narrower model to evaluate. Production decisions still need task-specific accuracy checks and a route for low-confidence or high-impact cases to receive review.

🦾 AGENTIC AI

Perplexity brings fully local Computer agents to Apple-silicon Macs

Perplexity launched a fully local mode for Computer on supported Apple-silicon Macs on October 20. In this mode, the model, agent harness and task execution run on the device, with permissioned access to local files and apps such as Notes and Mail. The company announcement says local Computer is available to Pro, Max and Enterprise subscribers using the Mac app, and recommends at least 36 GB of memory. Qwen 3.6 and PPLX Pro are the initial local model choices. Perplexity previously offered Mac file access and a hybrid mode that combined local work with cloud reasoning; this release moves an entire task onto the Mac. Users can still choose to escalate to cloud models or services for current information, browser use and connected apps. The privacy benefit therefore depends on staying in local mode and granting appropriate app permissions, while actual speed and task success will vary by hardware and workload.

Wabi opens its shared personal AI assistant to all users

Wabi opened its personal AI assistant to all users on October 19. Founder Eugenia Kuyda told TechCrunch that Wabi works like a messenger with separate agentic threads for tasks and longer-term goals. People can share a thread as a group chat and use small custom apps within it, such as trackers or activity calendars. The company's current site offers phone and Mac downloads. In TechCrunch's founder interview, Kuyda said connecting an email account can prompt Wabi to create initial threads and apps, while users may request more. She said information access is scoped to each shared chat and visible when requested. Wabi existed before this launch as a mini-app platform, so the new event is broader public access and a revised assistant experience. The design could make shared planning more flexible, but reported features do not establish how reliably the agent completes real-world tasks or handles every connected account.

🦿 ROBOTICS & PHYSICAL AI

Dream Labs launches Dream-1 robot API and announces $90 million seed

Dream Labs launched Dream-1 through a public robot API on October 19. The company says its World Action Model can control previously unseen bimanual robots on tasks they were not specifically trained to perform. Users send observations and instructions to the API, which returns actions. Dream Labs also announced a $90 million seed round and lists usage from $3 per hour. Its Dream-1 technical release shows rollouts across robot setups and reports tests of transfer to unseen embodiments. The company describes its current focus as bimanual manipulation with parallel grippers, and says results vary with robot hardware and inverse-kinematics quality. It does not control every aspect of arm motion near obstacles or joint limits. These limits matter more than a single demonstration: a public API makes the model available for outside trials, while dependable performance on a new robot still requires task-level testing and suitable control hardware.

⚙️ HARDWARE & CHIPS

Google opens TorchTPU to bring native PyTorch workloads to its chips

Google open-sourced TorchTPU on October 20, giving PyTorch a native backend for its tensor processing units. Announced at PyTorch Conference, the release lets developers use familiar PyTorch interfaces for training and inference on Google TPUs. The Google Cloud announcement also describes integrations with vLLM and SGLang serving engines, TorchTitan training and Ray orchestration. The public repository includes TPU operation kernels, compilation caching, distributed training examples and support for torch.compile. Google says developers can bring existing workloads over with relatively small code changes, although model compatibility and performance still depend on the workload and setup. Its published throughput and utilization figures come from specific Google or partner tests, rather than a general guarantee. PyTorch already had TPU access through other paths; this release matters because Google is opening a native stack and connecting it to tools developers already use across GPUs and CPUs.

AWS opens TorchNeuron beta for native PyTorch on Trainium 2 and 3

AWS opened TorchNeuron public beta on October 20 as a native PyTorch backend for Trainium 2 and Trainium 3. Its October 19 release notes describe registering a neuron device through PyTorch's extension system, letting supported models and tensors move to the accelerator with standard device calls. It supports eager execution for development, torch.compile for supported graph regions and distributed training APIs. The AWS release notes specify compatible PyTorch, Python, operating system and instance versions, so this is not blanket compatibility with every workload. Unsupported operations in eager mode can fall back to the CPU, and compiled mode has shape restrictions. Hugging Face says its Transformers and Diffusers libraries now support the native backend, reducing the special integration code needed for supported models. The useful shift is a more familiar PyTorch path onto Amazon's chips. Teams still need to test operator coverage, performance and migration effort on their own models before counting savings.

🏢 ENTERPRISE & APPS

Google adds a Gemini business hub, brand controls and finance tools

Google announced a new Gemini business hub and tools for small businesses on October 20. The company post describes a landing page for brand work, financial analysis and customer insights, along with connections to Google Ads, Workspace and outside services including Canva, Shopify, Wix and Xero. Owners can upload brand materials such as a website, fonts or color palette to guide image generation. Google also says Gemini can turn spreadsheets, receipts and purchase records into charts and interactive financial views. These are product features and workflows, not evidence that Gemini's advice is accurate enough for financial decisions without review. The business landing page, marketing and finance skills, and additional app connections are beginning a global rollout over the coming days, so users may not see them yet. Google introduced Business Profile integration and Business notebooks in June; this announcement adds a distinct hub and broader tool set rather than creating business support from scratch.

Google adds guided slide and video creation to Gemini Notebook

Google added a co-creation flow for Slide Decks and Video Overviews in Gemini Notebook on October 19. Instead of moving straight from a prompt to a finished artifact, users can ask for slides or a video in chat and review suggested topics drawn from their notebook sources. A new Review toggle exposes outline notes for slides or producer notes for video before rendering. Users can request changes, including an executive summary or a different scene order, and then choose Proceed. The Google announcement says people who already know what they want can prompt directly into that review flow, while the existing studio path remains available for quick generation. The feature is rolling out globally to Gemini Notebook users over the following week, so access is not universal yet. For source-heavy presentations, the change puts an explicit review step between source selection and a generated deck or video, without claiming that it improves factual accuracy automatically.

Exa releases Exa 3.0 with a lower-cost Lite API tier

Exa released Exa 3.0 through its API on October 19, adding a Lite tier priced at $4 per 1,000 requests. The existing Auto tier is listed at $7 per 1,000 requests and uses more retrieval compute. Exa also made Dynamic Highlights generally available after an earlier preview; it says the feature selects relevant text across results while using fewer tokens than full-page retrieval. The company expanded its index with commerce listings, local businesses and current data such as news, sports and weather. Its Exa 3.0 announcement reports leading scores on eight search benchmarks, but those comparisons are Exa's own tests and should not be read as independent rankings. The practical change for builders is a choice between lower per-request price and more retrieval work, plus shorter context from highlights. Teams evaluating search for agents should test their own query mix, citation quality and total task cost. The stated prices describe launch terms and can change.

🧬 SCIENCE & HEALTH

Anthropic reports lab-tested Claude designs that bind scorpion toxins

Anthropic reported that 17 Claude-designed proteins bound scorpion neurotoxins in laboratory tests. Its October 20 research report describes a 64-hour design campaign guided by an initial prompt and four human follow-up instructions. Claude generated more than 318,000 candidates; researchers selected 180 and added 31 more with Claude's help. Adaptyv Bio and Twist Bioscience tested all 211 designs, and 17 met Anthropic's binding criteria against individual toxins. The company has released the designs and experimental data for further study. Binding is an early molecular result, not proof of a therapy. Anthropic says it has not established whether the proteins neutralize venom in cells, and it does not claim an antivenom. None of the designs intended to cover multiple toxins qualified against more than one. The work shows AI-assisted design producing lab-confirmed binders quickly, while the central questions of neutralization, safety, affordability and clinical benefit remain open.

⚡ ENERGY & INFRA

AEP and X-energy study 960 MW Indiana nuclear project with Amazon funding

AEP and X-energy announced an agreement on October 19 to study a 960-megawatt nuclear project in Indiana, with Amazon funding development. The companies are evaluating the former Breed coal plant site in Sullivan County. The plan calls for twelve Xe-100 small modular reactors rated at 80 megawatts each. Amazon intends to purchase the electricity through a long-term contract if the project reaches a final investment decision, is built and begins operating. The AEP announcement describes a feasibility effort, not a construction approval or power already on the grid. That distinction matters for data center energy forecasts: the 960 megawatts is proposed capacity, while site evaluation, financing, regulatory approvals and construction still lie ahead. AEP says its agreements include financial safeguards intended to protect existing customers. The partnership gives Amazon a route to support new nuclear supply, but the actual power and delivery schedule depend on decisions the companies have yet to make.

🧭 AI SAFETY

Ai2 releases FairStress to test how firmly models hold fair decisions

Ai2 introduced FairStress on October 20 to examine demographic fairness beyond a model’s final answer. The benchmark tests whether equivalent decisions remain equally resistant to persuasive arguments when demographic identities change. That matters because a model can initially select the stronger candidate in both versions of a scenario yet abandon that choice more readily in one. The FairStress release reports asymmetries across open-weight and frontier models, including differences in explanations and a narrower analysis of internal activations. Researchers also found that changing a benchmark’s question format and scoring method can change the apparent direction of bias. FairStress uses an equality-centered definition of fairness and does not claim to represent every competing definition. The team released data and code and is incorporating the benchmark into Olmo-eval. For builders, the practical implication is to test decision stability under matched challenges, rather than treating one balanced set of final answers as sufficient evidence.

🌍 GEOPOLITICS & POLICY

US publishes order removing DeepSeek tools from civilian agency IT

The United States publicly filed a notice on October 19 requiring federal civilian agencies to remove covered DeepSeek technology from their IT systems. The Department of Homeland Security notice, formally published in the Federal Register on October 20, also excludes DeepSeek-developed products and services from those agencies' future procurements. Its removal language covers DeepSeek and affiliate products plus third-party products and services developed using DeepSeek implementations. The scope is federal civilian agency procurement and information technology, not a nationwide ban on private or consumer use. The underlying exclusion and removal order was signed in July; October's new event is its public filing and publication. DHS issued it under the Federal Acquisition Supply Chain Security Act after a council recommendation. The notice does not by itself show which systems agencies currently use or how long removal will take. For federal buyers and suppliers, the published scope makes model and implementation provenance a concrete procurement and compliance question.

💸 DEALS & FUNDING

Asymptote raises $5.5 million for agent monitoring and runtime controls

Asymptote Labs announced a $5.5 million seed round on October 20, led by Tower Research Ventures. Lytical Ventures, 1984 Ventures and South Park Commons also participated. The company plans to expand engineering, product development, hiring and sales around infrastructure that records and controls AI agent activity. Its MIT-licensed Beacon project collects activity from Claude Code, Codex, Cursor and more than 20 other agent harnesses, according to the company. The funding announcement distinguishes that open-source telemetry layer from Asymptote’s enterprise platform, which adds prompt-level and tool-level policy enforcement. Beacon launched in May, so the fresh development is the financing rather than a new product debut. Asymptote says security teams use its tools in production, without naming customers in the announcement. Tower’s security organization is evaluating Beacon, a separate claim from a completed enterprise rollout. The investment backs a shared record of agent actions; claimed prevention capabilities still require testing in each deployment.

Tenstorrent closes $1.5 billion Series E for AI server expansion

Tenstorrent announced on October 19 that it closed a $1.5 billion Series E at a $6.2 billion post-money valuation. Trace Capital and Innovation X led the round, according to the company's financing announcement. Tenstorrent says the funding will help scale deployment of its Galaxy AI servers and expand commercialization of its AI accelerator and CPU intellectual property. That points to a company trying to move from chip design and early systems into broader server delivery, a capital-intensive phase where supply, software and customer adoption all matter. The announcement is a financing event, rather than the introduction of a new chip or a new Galaxy product; Tenstorrent had announced Galaxy Blackhole availability earlier. Its valuation, intended use of funds and commercial plans are company statements, and the release does not independently establish future shipment volumes or market share. The meaningful development is fresh capital earmarked for turning its existing hardware and IP into wider deployments.

Salesforce signs $5.66 billion agreement to acquire Braze

Salesforce announced on October 20 an agreement to acquire Braze for about $5.66 billion in enterprise value. Braze Class A shareholders would receive $47.50 in cash per share. The deal would bring Braze's real-time customer messaging and AI decisioning capabilities alongside Salesforce Marketing Cloud, Data 360 and Agentforce, but that combination depends on closing. Braze currently coordinates customer interactions across apps, websites, email, SMS and WhatsApp. In its transaction announcement, Salesforce says it expects the deal to close in its fiscal year 2028, subject to Braze shareholder approval and required regulatory approvals. That schedule makes the present milestone an agreement, not completed ownership or an integrated product. The enterprise value is also distinct from the per-share cash consideration. For enterprise buyers, the strategic direction is clearer than the delivery timetable: Salesforce wants live customer signals tied more closely to its broader CRM and agent systems, while customers must wait for the deal and any subsequent integration.

KEY TAKEAWAYS

Mistral paired new speech models with a voice-agent SDK - Teams can test end-to-end spoken workflows with hosted or self-deployed options.

Inception put Mercury 3 on multiple API platforms - Developers can compare a larger diffusion model on latency-sensitive applications.

OpenAI published two new coding evaluations - Legacy-software teams can inspect the tests, while validating migrations against their own systems.

Skyfall introduced SKY-001 for design tasks - Its CAD and BIM results merit scrutiny while access remains on a waitlist.

Viggle released JST-2 across three 3D tasks - Creators can test scene, character and motion workflows without assuming future features are live.

MiniMax released hosted M3.1 Flash access - Developers can measure its price and performance on their tasks; weights are still planned.

Cloudflare opened a hosted Clairic preview - Typed extraction may fit narrow tasks, but early-checkpoint accuracy needs local validation.

Perplexity added fully local Computer execution on supported Macs - Sensitive multi-step work can stay on device when users keep tasks in local mode.

Wabi opened public access to its shared assistant - Users can test its group-thread design while checking permissions and task reliability.

Dream Labs opened a robot-control API - Builders can test Dream-1 on supported hardware while checking transfer and safety limits.

Google released a native open-source PyTorch backend for TPUs - Developers gain another route to run and serve models on Google accelerators.

AWS opened a native PyTorch backend for Trainium in beta - Model teams gain a familiar interface, subject to operator and version limits.

Google is rolling out a Gemini business hub and new workflow tools - Small firms gain more connected brand and finance options as availability expands.

Google added an editable planning step in Gemini Notebook - Source-based slides and videos can be reviewed before rendering as rollout proceeds.

Exa released Exa 3.0 and a Lite API tier - Builders gain a lower listed request price, with quality claims still needing independent checks.

Seventeen designs met laboratory binding criteria - The findings offer research leads, while venom neutralization and treatment remain unproven.

AEP and X-energy agreed to study an Indiana reactor site - Amazon's funding backs feasibility work, while 960 MW remains proposed capacity.

Ai2 released a decision-stability benchmark - Fairness checks can examine matched persuasion tests alongside final answers.

DHS published a federal civilian DeepSeek exclusion and removal order - Agency buyers and suppliers now have a public scope to assess against their systems.

Asymptote raised a $5.5 million seed round - Its existing open telemetry project and paid enforcement platform serve different roles.

Tenstorrent closed a $1.5 billion financing round - The company says it will use the capital to scale AI server deployment and commercialization.

Salesforce agreed to acquire Braze - The $5.66 billion enterprise-value deal and proposed product integration remain subject to closing.

BotsFired · Bots Keep Firing 🔥