Models & capabilities, read from an operations desk

Latest

Google puts Gemini 4 Hydrogen behind Search for all users

Google has moved Search to its Gemini 4 Hydrogen model for all users worldwide. It handles multi-part questions and typos better, runs wider simultaneous sub-queries, and shows more tables, images and small tools. Anyone researching vendors or customers in Google will see different answers and different linked sites.

What changes for operators — If your team uses Google to size up vendors, check a prospect's background or compare tools, expect answers laid out as comparison tables, and run a few of your usual multi-part research questions to see whether they now save time. If your own site is part of the buying journey, search the questions your buyers ask in full sentences and check which pages Google links. Google says the wider fan-out surfaces sites people would not have found before, so your pages could appear more often, or less, and the source gives no traffic data. Treat it as something to measure in your own analytics over the coming weeks. Support and sales-enablement owners should also make sure product pages and help articles answer each part of a compound question on their own, because that is the query shape the model is built for.

Google ships Gemini 4 Hydrogen at $1.50 per million input tokens

If you run support, sales ops or back-office automations on Gemini 3.8 Flash, Hydrogen is the upgrade to test first. Pick ten real tasks, such as ticket triage, contract lookups or spreadsheet reporting, and run them at the low, medium and high thinking levels. Record cost, latency and error rate for each. The flat price makes long-document jobs easy to budget, and the thinking levels let you pay for deep reasoning only on the steps that need it. Treat every benchmark in the announcement as the vendor's own claim until your data confirms it.

NVIDIA releases GR00T 2 open robot model and Isaac Lab 3.0

If you run sales, support or back-office operations at a 10-200 person B2B company, none of this needs action today. It matters if you run a warehouse, assembly line or fulfilment operation, or buy automation from someone who does. The practical step is to ask robot vendors which models and benchmarks they test against, and whether their results come from simulation or from physical trials. The source reports only vendor tests and pilots, not production rollouts, so treat any delivery date from a supplier as unproven until you see a live demonstration on your own parts.

Tether AI Research unveils 0.8B on-device agent model BotPsy

If you run sales, support or operations at a 10-200 person company, BotPsy is worth a small pilot rather than a roadmap change. Its strengths match routine actions: calendar entries, email and messaging, file handling and order or account lookups, where a model must call the right tool and refuse when no tool fits. Running it locally keeps data on your own machines and removes per-call cloud fees. Pick one narrow, low-risk workflow and measure it on your own data, because every number here comes from the vendor. The vendor excludes money, personal information and other sensitive uses, and English is the recommended language, so customer-facing work in Russian or Hindi needs separate testing. Check the licence terms when the weights are published.

DeepMind ships Gemini Robotics 2.5 to robotics partners

You cannot buy this model directly today: it is going to DeepMind's robotics partners. If your company handles physical work such as kitting, picking or packing, the practical step is to ask any robot or integrator vendor which model sits underneath, and to request success rates measured on your own parts and bins rather than lab demos. The Mercedes-Benz pilot gives you a reference: a 96% success rate on one repetitive task, reported by the vendor. For sales, support and back-office automation, nothing changes yet. The one portable item is ASIMOV-Jailbreak, an open-source benchmark for disguised unsafe requests, which shows the kind of adversarial testing worth asking any agent vendor about.

n8n adds Google's Gemini 4 Carbon to its AI Agent nodes

If your n8n workflows read emails, form submissions or webhook payloads from outside your company, Carbon is worth testing on the agent steps that touch that input, and a lighter model can stay on routine jobs like classifying or formatting. Run a handful of real cases through it before switching anything, and keep your existing injection safeguards. Check your token spend too: the $2 and $10 rates are introductory and double after November 30, 2026, so budget at $4 and $20 if the workflow will run past that date.

Meta AI releases Muse Ember, a small model for agent tasks

If your support or ops team already runs high-volume steps like ticket routing, field extraction from contracts or invoices, or tool calls inside an agent, Ember is a new price point to test. Take a sample of real tasks, run them on Ember and on your current small model, and compare cost per finished task as well as price per token. Check how long your prompts are: the price rises sharply past 200k tokens, so long-document jobs cost more than the headline rate. Cached input is the cheapest line on the price sheet, so workflows that resend the same instructions or reference documents benefit most.

Google opens Gemini 4 Carbon to paying API users

If your team already runs support triage, lead research or report generation on a Gemini model, do not switch everything to Carbon. Google itself says Argon remains the better fit for most software and knowledge work, and Carbon spends more tokens per task. Pick the five or ten slowest or most error-prone workflows, run them on both models, and record cost per finished task at the November 30 price of $4 input and $20 output per million tokens. Keep Carbon for multi-step jobs where Argon fails, and test prompt-injection handling yourself before giving any agent access to customer inboxes or CRM data.

Anthropic ships Mythos 5.2 and Fable 5.2, cuts Fable price

If your support or sales workflows run on Fable, the price cut and the speed gain reduce cost per ticket or per lead without any change to your prompts, though you should rerun your own test set before switching versions. If legal or a client contract blocked Anthropic models because of data retention, zero data retention support is the item to take to them now. Faster serving costs double, so use it only on flows where a customer is waiting live, such as chat or call assistance, and leave batch jobs on the standard tier.

Nemotron 3.5 Super VL lands on AWS SageMaker JumpStart

If your team spends hours keying invoice fields, reading chart-heavy reports or triaging screenshots from customers, this gives you a vision model you can run inside your own AWS account rather than calling an outside API. Start with the NVFP4 variant on the default G7 instance and test it on 20 of your real documents. Request the ml.g7 quota first, since it starts at zero in new accounts. Plan for the limits: one image per request, PDFs rendered to images page by page, and slow first responses on long prompts.

Salesforce launches Koa Decisions, a fast routing model

If your support or sales queue runs on routing rules or on an LLM call for every incoming message, this is the kind of component to test. A decision model returns a team name or a yes/no in milliseconds, so triage stops being the slow step, and the LLM is kept for drafting and summarising. Adding a category such as "Returns" needs no retraining, which cuts the cost of changing your routing. Two caveats: the 26 ms and 1.4 s figures are Salesforce's own measurement on Agentforce routing, and the source gives no pricing. The developer API and the Flow, Apex and prompt-template support are plans, not shipped features. Try the free Trailhead demo on your own real tickets before you plan around it.

NVIDIA releases four open Cosmos 3 simulation models

If you run sales, support or operations at a 10-200 person B2B company, nothing in your workflow changes today: these models generate driving sensor data and robot-arm video, and none of them handles documents, CRM records or customer conversations. The release matters only if you sell into or build for robotics, vehicles or warehouse automation, where an open, commercially licensed simulator can cut the cost of testing. Even then, the source stresses that a training run needs GPU infrastructure, so budget for that before promising anything to a client. For everyone else, file it as a signal of where open models are heading and move on.

Dharma-AI releases 2B Field Extractor for document-to-JSON

If your team keys in supplier invoices, requisitions, IDs or contracts by hand, or pays a frontier API to pull four fields from each page, this is worth a test on a sample batch. Send a schema with the key names your CRM or ERP already uses, check that missing fields come back as null, and compare GPU hours with your current API bill. The cost figures are self-reported and depend on your volume, hardware and latency needs, so measure on your own documents before planning around them.

Google ships Gemini 4 Argon, previews Carbon for cyber work

If your team runs sales, support or operations on a 10-200 person budget, the practical step is to rerun your own test set against Argon through the Gemini API, especially for contract review, finance checks and code-heavy internal tooling, and see whether it beats the model you use now. Carbon has no public date, so do not plan around it. EmbeddingGemma 2 matters if customer files or call recordings must stay on your own machines: a local index of documents, audio and video could be tested without sending data to a vendor. Google's post gives no pricing or limits, so check those before committing a workflow.

Mistral Large 4 ships as open weights on Hugging Face

If you run sales, support or operations at a 10-200 person company, this model is mostly a hosted option for you, not a box to rack. The reference deployment on Hugging Face's Inference Endpoints uses eight NVIDIA H200 GPUs, so renting by the hour or using serverless Inference Providers is the realistic route. Check two things before a pilot: the license restricts selling the model as a service, which matters if you resell automation to clients, and reasoning cannot be turned off, so expect every reply to carry reasoning overhead in latency and cost. Test it on your own tickets, CRM notes and documents before trusting the vendor-reported benchmarks.

IBM updates Granite Switch: new ASR model, Decision adapter

If you run sales calls or a support line, the speech model matters most: transcripts arrive with punctuation and capitalization, so you can pass them straight into summarization or routing steps without a cleanup pass. Because 4.1.1 is backwards compatible and leaves the base model alone, a team already on 4.1 can swap in the new adapters and re-run its own call samples, with no migration project. The source gives no pricing, licence or accuracy figures for the new speech model, so test it on your own recordings and languages before relying on it. The Decision adapter gain is reported only against IBM's own test set, so treat it as unverified for your workflows until you have checked it.

Mistral publishes Large 4 weights under a modified MIT license

If you run sales, support or operations at a company of 10-200 people, the practical change is a new option for agents that write code, read documents or browse the web. You can test it through the API at $0.70 input and $2.10 output per million tokens, or through Microsoft Foundry, before deciding anything about hosting. Self-hosting a 1T-parameter model is a separate infrastructure decision, and the source gives no hardware requirements, so ask for them before you budget. The license matters only if you resell model access as a service. Run your own tasks, such as ticket triage, CRM data cleanup or report generation, through the low, medium and high reasoning modes and compare cost per finished task, because the vendor's benchmarks do not cover your workflows.

DeepMind opens Genie 4 world model to trusted testers

Nothing here changes a sales, support or operations workflow today: access is limited to trusted testers and a waitlist, and the Gemini API route is only for early access partners. The part worth tracking is the testing idea. DeepMind pairs Genie 4 with Gemini models so an AI system can be tried in a simulated environment before it touches the real one. If you plan to put agents into live processes, watch whether that kind of pre-deployment testing reaches the API. Until then, treat it as a signal and not as a purchase.

Liquid AI releases open-weight LFM3 family of 2B-class models

If you run support or back-office operations, the useful parts are the d1 decision models and the fine-tuning notebooks. d1 answers several questions about one ticket or invoice in a single pass, each with a probability, so routing, refund detection and urgency scoring need no separate prompts. The notebooks cover turning messy documents and receipt photos into JSON. Because the models are small and open-weight, a short fine-tune on your own tickets and terminology can run on private infrastructure. Treat every benchmark as Liquid AI's own and test on your data before relying on it.

NVIDIA adds StormScope EU nowcasting model to Earth-2

If you run operations at a 10-200 person company in energy services, logistics or insurance-adjacent work, you will probably not train this model yourself. The realistic route is through a weather vendor: Vaisala Xweather says it is adapting StormScope with lightning data and exposing the warnings through its SaaS products and AI-assistant integrations. So the useful question is for your current weather-data supplier: does it use regional nowcasts, and can the alerts feed your dispatch, scheduling or customer-notification workflows? If local conditions do not affect your delivery or service commitments, this item needs no action.

Canva launches Craft, its own image model, for all users

If your sales or support team already builds decks, social posts or one-pagers in Canva, Craft puts generation inside that workflow, and outputs can be turned into layered designs with Canva Magic, so staff edit rather than regenerate. The part worth watching for automation is the announced Canva API access and the planned MCP support for ChatGPT and Gemini: that would let a workflow request on-brand visuals without a person opening the editor. None of it is priced in the source, so test quality on your own templates before routing any process through it.

Amazon Polly upgrades Generative voices, adds Hindi and Finnish

If your support line, order-status IVR or voice agent already uses a Polly generative voice, the new engine arrives on its own and costs nothing extra, so the work is checking that your own scripts still sound right. Listen to the calls where the voice reads back order numbers, addresses, totals and confirmation codes, since that is where AWS says accuracy improved. If you serve Finnish or Hindi speakers, or Korean, German or Brazilian Portuguese ones, the catalog now has generative options to test against what you use today. If your flow depends on speech marks, those are still not available on generative voices, so check that before you switch. Before deploying across Regions, confirm the voice you picked exists in the Region you run in.

IBM Research ships a four-model Granite forecasting ensemble

If your team forecasts anything as a time series, such as ticket volume, order flow or staffing load, the practical change is one less model-selection decision. You load a single recipe, get a median plus a 0.1 to 0.9 range, and use the range for planning. Before trusting it, backtest on your own history and compare it with the best single member. Check that the 80% interval actually covers about 80% of outcomes, and measure runtime, because four checkpoints load instead of one. For a small team without a data scientist, this is a reasonable first forecaster to test, not a proven winner.

Anthropic's Claude 5.5 family: Opus, Sonnet and Haiku released

If your support or sales workflows already run on Claude Opus 5 or Sonnet 5, these releases are a prompt to re-test the same tasks on the 5.5 versions and compare cost per ticket or per lead summary. Anthropic describes Haiku 5.5 as built for summarization, subagents and browser use, so it is the candidate for bulk steps like call summaries or inbox triage. Use your own tasks as the benchmark, because the cost figures are Anthropic's own and depend on the workload.

IBM's ALTK-Evolve tops τ²-bench retail with GPT-4.1 memory

If you run support on a 10-200 person team, the useful point is where the gain comes from: a maintained rulebook in the system prompt, not a bigger model. Rules such as authenticating before any order action, confirming changes explicitly and stating refund timing are the kind your team already writes down. The price is input tokens, about 65% more per retail conversation with the full 292-rule set, and a smaller rule set can be the more reliable choice in some domains. Test on your own transcripts, since the evaluation here is small and run on simulated customers.

Cloudflare previews Clairic, an extraction model on Workers AI

If your support or finance team reads order numbers, dates, amounts and damage descriptions out of inbound emails, receipts or invoices, this is a cheaper way to fill those fields than prompting a general LLM for JSON. The confidence score and null probability give you a rule for routing: auto-process high-confidence values, send low-confidence ones to a person. The estimate and forecast fields could also feed lead scoring or volume planning. These are Cloudflare's own examples, so run it on a sample of your real documents before wiring it into anything that moves money.

Amazon Bedrock adds TwelveLabs video model Pegasus 1.5

For a 10-200 person B2B company, this matters only if you hold long video: recorded demos, support sessions, training material or webinars. The source's own examples come from media and sports, so any fit with sales or support recordings is our reading, not the vendor's claim. If you do have such a library, you can define a segment once, such as a pricing discussion or a product shown on screen, and get consistent JSON for every file. That feeds a CRM, a ticket queue or a search index without anyone scrubbing through footage. If you have no long video, there is nothing here for you to act on. Check the Region and the separate AWS Marketplace billing before you plan around it.

Meta releases Muse Spark Open, an 816B open-weight model

For a 10-200 person B2B company, the useful fact is that a frontier-scale model with tool calling and image, audio and video input is now open-weight, so support or sales workflows that handle screenshots, call recordings or screen captures could run without a closed API. The cost is hardware: the smallest validated hosted setup in the source is eight GPUs, so most teams will not self-host. The realistic route is a managed endpoint or a pilot. Test one narrow workflow, such as ticket triage with screenshots or call transcription, and compare quality and cost per task against what you use today. Because reasoning cannot be turned off, budget for the extra tokens.

Reflection releases Beam, a 501B open model under Apache 2.0

For a 10-200 person B2B company, Beam is a model to test through an API, not to host yourself. Self-hosting means roughly 280 GB of weights even in the 4-bit NVFP4 format, and the post suggests four B200 GPUs. Hosted access through Baseten, Fireworks or Together lets an ops lead run it on real tasks first, such as ticket triage or order checks. Treat the numbers as vendor-supplied. The tool-calling scores are modest: Tau3 Banking is 36.1% and AutomationBench (public split) is 39.3 for the BF16 checkpoint. Run your own workflow before replacing a current model.

NVIDIA opens 2.8T-token Indic dataset and translation pipeline

A 10-200 person B2B company will not pretrain a model on this, so the direct effect is indirect. If you handle support tickets, sales email or documents in Hindi, Tamil, Bengali or other Indian languages, expect model vendors to cite this kind of data, and test any claim on your own tickets, not on benchmarks. The more reusable part is the method: translate, score against a rubric, repair what fails, accept only above a threshold. The same loop fits translating a knowledge base or support macros, provided native speakers review a sample first. NVIDIA notes that the translating model also acts as judge in the twelve-language workflow, and that human review coverage varies by language, so automated scores should not replace your own spot checks.

Meta ships Muse Wonder, with six reasoning levels

If any workflow of yours calls a Meta model, such as lead qualification, ticket triage or drafting replies, treat Muse Wonder as a candidate to test, not a switch to flip. Run your own sample of real tickets or CRM records through it at a low and a high reasoning level and compare the quality of the output. Then compare the cost and the wait for each answer, because the source gives no pricing. The release pace matters more than any single model. With a new Meta model arriving roughly monthly, keep the model name and reasoning level in one config setting, and keep a fixed test set so that re-checking takes an afternoon rather than a project.

IBM Granite Switch 4.1 adds decisions and speech

If your support or ops team currently runs separate services for transcription, ticket or call screening and drafting replies, this is a candidate for consolidating them onto one GPU deployment. A call could be transcribed, summarised and checked against your own flags in a single pass. Treat the speed and accuracy comparisons as IBM's claims: test the 8B model on your own call recordings and tickets before you plan around it. The source gives no pricing or licence terms here, so read the model cards first.

Meta ships Muse Spark Open, an open-weight model for Macs

For a team running sales, support or operations, the local Mac option is mostly out of reach: a requirement of at least 1 terabyte of RAM puts it in workstation territory. The practical route is the API platforms that already offer the model. If you pay per token for a closed model on coding or agent-style workflows, it is worth running a side-by-side test on your own tasks, since Meta's benchmark claims are its own and unverified. The open-weight format matters if cost control or the ability to inspect and adapt a model is a concern. Treat it as a candidate to evaluate, not a reason to change tooling this week.

Swiss AI's Apertus 1.5 adds image, audio input and 262k context

For a 10-200 person B2B team, the practical change is that an open-weight model can now take a screenshot, a scanned document or a chart as input, and read a long ticket history or contract in one prompt. Pick one workflow where staff paste images or long threads into a chat today, run the 8B model on it, and compare the output with your current tool. Treat audio as a pilot only, since the authors call it experimental. The model is still behind frontier open-weight models, so test it on your own documents before moving any process onto it.

NVIDIA releases open 74B agent model for RTX Spark laptops

For a 10-200 person B2B team, the useful part is not the laptop but the pattern: run routine agent work on a local model and send only hard queries to a cloud model. Switchyard, an open-source router, is built for exactly that split. Credential protection and monitoring in OpenShell matter once an agent is allowed near a CRM or helpdesk login. Treat all of it as vendor claims until tested. The source gives no laptop pricing, and the hybrid-routing and cost-saving claims are NVIDIA's own. A sensible first step is to pilot one narrow, repetitive task, such as classifying inbound requests, and compare results and cloud spend against your current setup.

Liquid AI ships open-weight 24B agent model for local use

If you run sales, support or operations at a 10-200 person company and want agents that call tools without sending every request to a hosted API, this is a model to test. Start with a narrow, rule-bound workflow such as ticket triage, CRM field updates or following a scripted process, where the reported strengths in tool use and instruction following apply. Do not give it knowledge-heavy questions without retrieval: Liquid AI itself recommends pairing the model with retrieval or web search for those. All scores come from the vendor's own evaluation, so run your own tasks before committing. The source does not state licence terms, so check the model card before commercial use.

Cloudflare's Clef-omni adds audio, video decisions

For a support or ops team using classification models to triage tickets, screen uploads for PII, or flag suspicious content, Clef-omni removes the need to stitch together separate speech-to-text and video-frame extraction steps before scoring — a customer complaint with a voice memo or video attachment can now go through one call. The Clef-flash price cut (now below Jev's pricing) and the 2x speed gain on Clef make it cheaper and faster to run these decision checks at the volume a 10-200 person company's ticket or document queue actually generates, though teams with inputs over 24k tokens per request will need to use the pricier Clef model instead of Clef-flash.

TII releases Falcon-ASR, a 1.6B Arabic speech model

For a support or sales team handling calls in the UAE or wider Gulf, transcription accuracy on dialectal Arabic has been a persistent weak point in voice automation pipelines — standard ASR tuned for Modern Standard Arabic often mangles everyday Emirati speech, which breaks downstream QA scoring, CRM logging and agent-assist tools. Falcon-ASR's word-level timestamps and lower error rate on Emirati and Gulf dialects make it a candidate for teams building or rebuilding call-transcription, voice-ticketing or meeting-notes pipelines in that region; the model is available to test now via a Hugging Face demo, though API access and native applications are only planned, so production integration isn't ready yet.

AWS adds Claude Haiku 5.5, cuts subagent costs 75%

For a 10-200 person company running support ticket routing, document classification or knowledge-base Q&A through Bedrock, this is a direct cost lever: the same volume of quick, repeatable tasks now runs at roughly a quarter of the previous Haiku price, and the new effort controls let you dial down cost on simple lookups while keeping more reasoning budget for harder cases. Paired with Opus 5.5 as the planner, Haiku 5.5 becomes the cheap execution layer for many parallel subagents, which is the architecture pattern most relevant to teams automating high-volume, low-judgment work like inbound triage, data extraction or bulk document review.

Reflection AI's Beam undercuts OpenAI, Anthropic on price

For a 10-200 person company running AI agents for sales or support, the per-token cost of the underlying LLM is a direct line item, and Beam's pitch — Chinese-model-level performance at a fraction of Anthropic or OpenAI pricing — is worth testing against current vendor bills once Reflection publishes the weights and technical details. The same applies to Mistral's newly launched Large 4 ('le Chonk'), positioned for coding, finance and security tasks. Neither model has public benchmarks or pricing in this report, so any switch should wait for those details rather than being planned off the announcement alone.

Mistral Launches Large 4 Preview With Published Pricing

For a 10-200 person company running agents against email, spreadsheets, Slack or Salesforce, ML4 is a new candidate model worth benchmarking against whatever currently powers those workflows, since Mistral reports it ahead of Kimi K3, MiMo-V2.6-Pro and DeepSeek V4 Pro on exactly that kind of task. The published per-token price ($1.36 input / $4.18 output) lets ops teams run a direct cost comparison today through the Mistral Studio preview API, and the open-weight release planned for later this month adds a self-hosting option for teams that need to keep sales, support or operations data off third-party infrastructure for compliance or security reasons.

Google DeepMind Ships a Free On-Device Multimodal Search Model

For a 10-200 person B2B company, this is a building block, not a finished product: it means an engineer could index support transcripts, product screenshots, onboarding videos and call recordings into one searchable system that runs on a laptop or a modest server, with no data leaving the premises and no per-query API bill. That's directly useful for support teams who want to surface the right knowledge-base article, past ticket or recorded demo clip from a single query, and for ops teams building internal search over scattered multimodal documentation. The catch is that EmbeddingGemma 2 is only the retrieval half of a RAG system — someone still has to build the pipeline, pair it with a generative model, and maintain the vector database, so the near-term win is lower infrastructure cost and better data privacy, not a drop-in product.

New Arabic LLM Answers in Emirati Dialect, Not Standard Arabic

For a company running support or sales chat for customers in the UAE, this matters because generic Arabic language models tend to answer in formal Modern Standard Arabic even when a customer writes in everyday Emirati phrasing, which reads as stiff or foreign to a local audience. Falcon-Emirati-7B is built to recognize and reply in the dialect itself, including culturally specific references, which is directly relevant to anyone evaluating an Arabic-language chatbot, ticket triage system, or sales assistant for a GCC market. Teams already running Arabic support automation should treat dialect fidelity, not just translation accuracy, as a selection criterion, and test any candidate model against real customer phrasing before deployment.

New Benchmark Catches AI Agents Lying About Finished Work

If your support or ops team has an agent closing tickets, updating order status, or confirming refunds, this benchmark is a warning that the agent's own confirmation message is not evidence anything actually happened correctly. The same model can solve a workflow once and fail it the next four times, so a single successful pilot run tells you almost nothing about reliability at volume. For a 10-200 person company, the practical fix described here doesn't require buying the benchmark: check the terminal state of the record the agent touched (ticket status, order field, account balance) before trusting its summary, classify tool errors so retries target recoverable failures instead of re-running everything, keep the agent's tool access narrow to the workflow it's doing, and require human sign-off on any change that's expensive to undo, like refunds or cancellations.

Cloudflare Ships Open-Source Clef Models for Support Triage

The example Cloudflare publishes is the exact workflow most support teams already run by hand or with brittle rules: a message comes in, someone decides if it's urgent, which team owns it, and how severe it is. Clef does all three in one API call with typed, probability-scored outputs your existing code can act on directly, no LLM prompt engineering or free-text parsing required. For a 10-200 person company this means ticket routing, escalation flags, and severity scoring can run as a deterministic step inside an existing support stack, with humans only pulled in when the model defers. The new RL fine-tuning option (via Cloudflare's hands-on team first, self-serve later) also means a company with a backlog of labeled support tickets or sales qualification outcomes can tune the model on its own data rather than relying on a generic classifier.

Claude Sonnet 5.5 lands on AWS Bedrock

For a company already running Bedrock-based coding assistants, alert triage, or document drafting, this is a straight cost-and-latency lever, not a new capability to evaluate from scratch: routing well-defined, high-volume tasks (SQL generation, first response to alerts, spreadsheet edits, routine document work) to Sonnet 5.5 while reserving Opus 5.5 for release debugging, security review, or contract redlining should lower per-task spend without touching the existing IAM, CloudTrail, or Guardrails setup. Teams with fixed monthly AI budgets for IDE coding agents or support-ticket triage get the clearest immediate win, since Sonnet 5.5 is pitched specifically at continuous or at-scale workloads with a fixed spend cap.

H Company Ships Holo4, Open-Weight Agents That Work Any Software Interface

For a 10-200 person B2B company, the practical draw is not the benchmark score but the interface flexibility: most internal tools, legacy CRMs, and vendor portals have no usable API, forcing manual clicking today. A model that can fall back to GUI control when an API is missing, then switch to API calls where one exists, is directly applicable to automating order entry, ticket triage, or data reconciliation across tools that were never built to be automated. The open weights (BF16, FP8, NVFP4, GGUF) mean a technical team could self-host rather than pay per-call frontier pricing, though the reported OSWorld 2.0 gap versus Opus 5.5 (61.7% vs 81.8%) signals this is a cost-performance tradeoff, not a like-for-like replacement, and should be piloted on a specific workflow before being trusted with production tasks.

Microsoft Foundry Adds Two Cheaper GPT-6 Tiers for Production Agents

For a company running support or ops agents on Azure, this is a direct cost lever: instead of sending every ticket-classification or data-extraction step through an expensive reasoning model, teams can now put GPT-6 Luna behind routine, high-volume steps and reserve Sol or Astra for the parts of a workflow that actually need judgment. At Global Standard short-context pricing, Luna runs $0.10 per million input tokens and $0.50 output versus Sol's $2/$10 and Astra's $10/$50, which changes the math on cost-per-task for anyone running thousands of routine agent calls a day. The practical move is to map an existing agent workflow, identify which steps are classification/routing/summarization versus multi-step reasoning, and split the model assignment accordingly rather than defaulting to one model for everything.

TypeSafe AI ships Jev, a decision layer to replace LLM calls

For a support or ops team that currently pays an LLM to tag ticket urgency, screen resumes or decide which model handles a request, Jev is a much cheaper intermediate step: it returns a structured score or choice with a confidence value instead of free text, so you can auto-route high-confidence cases and send low-confidence ones to a human. The catch is the 67.8% accuracy figure and the lack of any explanation for how it reached a decision, so it's a triage layer, not a replacement for judgment on anything with real downside if it's wrong — and adversarial inputs (a crafted support message, a doctored CV) can still hijack its reasoning without strong guardrails around it.

Together AI ships a $17 recipe for custom support-ticket classifiers

For a support team fielding a few hundred tickets a day, this is a concrete blueprint for a first-line triage bot: feed in the customer message plus your existing intent list, get back a single label in under a token's worth of output, and route accordingly. The training cost (~$17) and time (~25 minutes) make it cheap enough to iterate on your own historical tickets rather than rely on a general-purpose LLM call for every classification, and the fixed temperature=0, max_tokens=8 setup makes outputs predictable enough to feed directly into a routing rule or CRM webhook without a human review step for high-confidence cases.