Try it
Open-source · 28 October 2026
Nvidia's open Nemotron 3.5 Transcribe handles transcripts, speakers and translation in one model
Nvidia released Nemotron 3.5 Transcribe as open weights. It transcribes 40 languages, translates 31 to or from English, labels up to 8 speakers and handles recordings up to an hour. Nvidia reports a 5.33% English word error rate on the Hugging Face Open ASR Leaderboard. Fine-tuning tools are included.
What it means for you: If you pay per minute for meeting or call transcription, test it on a few of your own recordings. The benchmark figures are vendor-reported, so check accuracy on your accents and jargon before you switch.
Watch
New model · 28 October 2026
Gemini 4 Carbon tops the intelligence index, but only cyber defenders get it first
Google DeepMind's Gemini 4 Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59. Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. Introductory pricing is $2 per million input tokens and $10 per million output tokens. Access starts with a subset of cyber defenders, then trusted partners.
What it means for you: You can't use it yet, and the scores are Google's own. If you run security or critical systems, ask your vendors whether they will get access. Check the price again when general release opens.
Try it
Security · 27 October 2026
Anthropic keeps agent credentials outside the sandbox that runs code
Anthropic's engineers explained how Managed Agents separates the agent's session log and harness from the sandbox that runs code. A failed sandbox no longer ends the whole session. Credentials sit outside the sandbox and are attached through a proxy, so generated code cannot read them.
What it means for you: If you build or buy AI agents, ask where the passwords and API keys live. Keeping them out of the environment where the agent runs code is a sound design to copy.
Try it
Trend · 27 October 2026
Ai2: AI judges of research novelty swing wildly with small prompt changes
Ai2 tested six judges, including GPT-5.x and Claude models, on whether one research idea is more novel than another. Changing a few words in a prompt shifted one judge's accuracy by more than 50 percentage points. Extra retrieval and reasoning did not consistently help. A simple Claude Opus 4.6 prompt beat two purpose-built systems.
What it means for you: If you use AI to score, rank or screen anything, such as proposals, applicants or documents, test it with reworded prompts. Treat a single verdict as unreliable, and keep a person in the loop.
Skip the hype
Trend · 26 October 2026
OpenAI says its models disproved a major computing hypothesis. Wait for experts.
On 26 October OpenAI announced three results from internal models. They cover harmonic analysis, complexity theory and physics. OpenAI claims a proof of the local smoothing conjecture in every dimension and says the Exponential Time Hypothesis is false, verified in Lean. The claims come from OpenAI's own announcement, with no independent confirmation reported.
What it means for you: Nothing here changes your costs, tools or risks this week. Don't base any decisions on it until outside experts have reviewed the work.
Try it
Open-source · 22 October 2026
IBM's Granite Timeseries Ensemble forecasts without retraining
IBM released an open-source ensemble of four pre-trained time-series models: PatchTST-FM-r1, PatchTST-FM-r2, FlowState and Tiny Time Mixers. It gives point forecasts and forecast ranges with no retraining or weight fitting. The Probabilistic Ensemble, with equal weights, is the recommended default.
What it means for you: If you forecast sales, stock or demand, run the default ensemble against your current method on past data. It removes the guesswork of picking a single model.
Watch
Security · 22 October 2026
AI speeds up exploit writing, but many AI-made proof-of-concepts don't work
VulnCheck's Spencer McIntyre says AI agents are making vulnerability discovery and exploit development much faster. Many of the exploits they produce fail in real testing. Initial access intelligence helps teams tell credible threats from the growing pile of weak proof-of-concepts.
What it means for you: More public exploit code will not mean more real threats. Rank patching by evidence that an exploit works, such as known active exploitation, not by whether a proof-of-concept exists.
Try it
New model · 21 October 2026
NVIDIA's Kumo Relational predicts churn and fraud straight from your database tables
Kumo Relational is a pre-trained model that predicts outcomes like churn, demand or fraud directly from relational tables. It needs no task-specific training or feature engineering. It uses up to 10,000 past examples as context. Across 41 tasks on 15 databases, NVIDIA reports gains of nearly 4 AUROC points over XGBoost.
What it means for you: If you run churn or fraud models on XGBoost, test this on one of your own tables. These are NVIDIA's benchmark results, so check them against your own data before changing anything.
Try it
Open-source · 21 October 2026
IBM Granite Speech 5.0: open transcription models under Apache 2.0
IBM released Granite Speech 5.0 2B, which covers six languages and handles noisy audio better. It supports keyword biasing for up to 2,000 terms. A 470M-parameter Turbo CTC model is 1.6x faster and now covers seven languages. All models are Apache 2.0.
What it means for you: Transcribing meetings or calls on your own hardware keeps client audio in-house. Test it on a few of your own recordings, with Australian accents and your industry's terms in the keyword list.
Watch
Trend · 21 October 2026
Custom vocabularies cut one small model's inference cost by 77%
Dharma-AI's VocabX swaps a model's general tokenizer for one tailored to the task, with no change to the model itself. On a Portuguese essay-grading system built on Qwen3-4B, it cut inference costs by 77%. Generating a token costs 7.4 times more than reading one.
What it means for you: This is one team's result on one narrow task. If you run a small model on a repetitive, structured job, ask your developers whether it is worth trying before you scale up.
Try it
Open-source · 16 October 2026
Reflection's Beam: 501B open model with only 23B active, under Apache 2.0
Reflection, a new US AI lab, released Beam on 16 October under the Apache 2.0 licence. It is a mixture-of-experts model with 501B total parameters and 23B active per token. It has a 1M-token context window and runs in transformers, SGLang and vLLM. Reflection also ships fine-tuning and reinforcement learning recipes.
What it means for you: Apache 2.0 means you can run it on your own infrastructure, which helps if client data must stay onshore. If your team builds coding or agent tools, test it on one real task against your current model.
Watch
Price change · 16 October 2026
Cognition tests token budgets tied to output quality, not usage
A podcast roundup of this week's announcements says Cognition is experimenting with performance-based token budgets. These tie AI cost to the quality of the output instead of the volume used. It also lists Microsoft's MAI Base 2 and Meta's Muse Spark Open as giving enterprises more control over customisation and cost. Anthropic is adding embedded safety evaluators.
What it means for you: This is an experiment, and the story gives no prices. Nothing to change yet. When you next negotiate or budget for AI tools, ask vendors whether pricing can be linked to results, not just tokens.
Watch
Open-source · 15 October 2026
Meta's Muse Spark Open: 816B multimodal model for coding agents
Meta open-sourced Muse Spark Open on 15 October. It is a mixture-of-experts model with 816B total parameters, about 42B active per token and 256 experts. It takes text, images, audio and video, has a 1M-token context window, and is built for coding agents. Meta describes it as the most capable open-weight model available to US developers.
What it means for you: The capability claim comes from Meta and the story doesn't cover licence terms. Read the licence and check it is available to you in Australia. Then decide whether it is worth the hardware to host it.
Try it
New model · 15 October 2026
IBM Granite Switch 4.1 puts chat, decisions and speech on one endpoint
IBM and Hugging Face released Granite Switch 4.1 in 8B and 30B sizes. It uses Activated LoRA adapters that share one base model's cache. IBM says this adds decision, safety (Guardian) and speech-to-text functions for about the same GPU memory as the base model alone. The speech model is claimed to be the fastest on the OpenASR leaderboard.
What it means for you: If you run separate models for chat, classification and transcription, this could cut your GPU bill. The performance claims are IBM's and Hugging Face's own, so test the 8B model on your own audio and classification tasks first.
Try it
Open-source · 15 October 2026
FLUX.3 Image Klein: open image model with bounding-box layout control
Black Forest Labs released FLUX.3 Image Klein as an open model in Hugging Face's Diffusers library. You can set where elements sit using bounding boxes written into the prompt. The distilled checkpoint makes a 1024×1024 image in about 2 seconds on an H100. FP8 and NVFP4 versions run on smaller GPUs, and LoRA fine-tuning scripts are included.
What it means for you: Layout control suits product shots, ads and mock-ups where placement matters. Try the quantised version on a modest GPU with one repeatable design job, and check the licence before using it commercially.
Try it
Trend · 14 October 2026
Google's Deep Review checks contracts and reports, with two free reviews a week
Google has released Deep Review, an agentic system that checks logical consistency, maths and claims in long documents. It grew out of a tool that gave feedback on more than 16,000 conference papers. It now covers legal and financial documents in PDF, Google Docs and slides. Users get two free reviews a week, with API access through a Gemini key.
What it means for you: Test it on a non-sensitive document as a first-pass checker. Check where uploaded files go before you use it on client or confidential material. A person still makes the final call.
Skip the hype
New model · 13 October 2026
Meta's Muse Wonder: strong benchmarks, patchy results in real use
Meta's new Muse Wonder comes in two tiers, Wonder and Wonder Full. Users are split on it. Strong benchmark scores are not matching results on real tasks.
What it means for you: Don't switch tools because of published scores. If you want to try it, run it on three of your own tasks and compare quality, speed and cost against what you use now.
Try it
New model · 13 October 2026
LiquidAI's 24B agent model activates only 2B parameters and runs on consumer hardware
LiquidAI's LFM2.5-24B-A2B is a reasoning model that uses just 2 billion of its 24 billion parameters per token. It scored 48.09 on the LiveClaw agent benchmark against 38.19 for Qwen3.6-35B. It decodes at 184 tokens a second on an Apple M5 Max.
What it means for you: If client data can't leave your office, a local agent model is now realistic. Test it on one tool-use task you run today, and check the licence before any commercial use.
Watch
Regulation · 13 October 2026
AI inquiry asks for independent safety checks inside AI companies
The parliamentary AI inquiry sent early advice to the Prime Minister: independent experts inside leading AI firms, early model access for security agencies, and a copyright review. OpenAI said it would support mandatory reporting of incidents like the Medicare portal breach. The final report is due at the end of November.
What it means for you: The rules will land on AI providers first, but incident reporting is likely to reach users too. Start a simple log now of what your AI tools can access and any time they get it wrong.
Skip the hype
Trend · 12 October 2026
October’s release rush is mostly small, specialised models
Trackers counted 12 to 20 new models in the first half of October, but no new general-purpose frontier model went on open sale. Most launches are niche image, video or task-specific models.
What it means for you: Nothing here changes which assistant or API you should use. Keep your stack, and watch only the price moves from the big providers.
Watch
Regulation · 8 October 2026
Canberra plans to make AI developers prove their models were safety-tested
The Government said rules being finalised would require overseas AI developers to show their models were tested for harm before operating in Australia, modelled on banking and aviation safety rules. National AI standards are expected by the end of the year.
What it means for you: Expect your AI vendors to start publishing safety and testing documents. Ask for them when you buy, and keep them on file next to your AI register.
Try it
Trend · 8 October 2026
Hugging Face's ML-Intern agent fine-tuned six models for $1.90 to $16 each
A Hugging Face developer used the ML-Intern agent to plan, train, test and publish six specialised models in a few days. Each cost between $1.90 and $16 in compute. A citrus disease model went from 14.9% to 52.8% accuracy. The agent asks permission before spending and runs smoke tests first.
What it means for you: If a general model is too big or too weak for one narrow task, a small fine-tune may now cost less than a lunch. Set a hard budget cap and measure a baseline first.
Act now
New model · 8 October 2026
Gemini 3.7 Flash retired: requests now go to 3.8 Flash
Google deprecated Gemini 3.7 Flash and 3.5 Flash. Requests are routed automatically to Gemini 3.8 Flash and 3.6 Flash. Some old parameters (frequency_penalty, presence_penalty, candidate_count) now return an error, and temperature and top_p are ignored.
What it means for you: Your app may already be running on a different model than you tested. Update the model name, remove the unsupported parameters and re-check your outputs.
Act now
Price change · 7 October 2026
Claude Haiku 5.5: Anthropic’s small model is now 90% cheaper
Anthropic released Claude Haiku 5.5 at US$0.10 / US$0.50 per million input/output tokens for requests under 100,000 tokens (US$0.50 / US$2.50 above that), down from US$1 / US$5 for Haiku 4.5. Anthropic says it beats Haiku 4.5 on coding, computer use and knowledge work.
What it means for you: High-volume jobs like tagging emails, summaries and routing now cost about the same on Claude as on GPT-6 Luna. If you run Haiku 4.5, switch the model name and re-run your test set this week.
Watch
Trend · 7 October 2026
Google study: AI lifts lawyers' work quality, but only senior staff gained judgment
A three-month experiment with 133 patent lawyers at eleven firms found that AI access lifted drafting scores by 0.34 to 0.38 standard deviations, about 11 percentile points. Tested without AI, lawyers with 7+ years' experience showed stronger judgment. Junior lawyers showed no average gain. Some improved and some got worse.
What it means for you: If you run a professional services team, don't assume juniors learn by using AI. Keep them reviewing and redlining work without it, and have seniors check their output.
Watch
Open-source · 6 October 2026
Mistral Large 4 preview: open weights promised, not yet shipped
Mistral released a public preview of Mistral Large 4 through its API. It says the open weights will follow by the end of October, but the licence has not been named and there is nothing to download yet.
What it means for you: A large European open model would be a real option for private AI. Wait for the weights and the licence before you plan around it; today’s best open-weight picks still stand.
Watch
Security · 30 September 2026
Gemini 4 Argon goes to cyber defenders first
Google announced Gemini 4 Argon, its new flagship, and gave it first to vetted cyber-security defenders. Google says it can find, check and patch serious software vulnerabilities on its own. Paid API and Ultra access come later, with no dates yet.
What it means for you: AI that finds security holes this well will be used by attackers too. Keep your software and plugins patched quickly, and ask your IT provider how they are using AI for defence.
Try it
New model · 28 September 2026
Claude Sonnet 5.5: near-top quality at a mid-range price
Anthropic released Claude Sonnet 5.5. It runs 30%+ faster and costs up to 30% less per task than Sonnet 5, and nearly matches Opus 5.5 on knowledge-work benchmarks.
What it means for you: For most everyday work (drafting, analysis, customer replies, agents) this is now the price-to-quality sweet spot. If you pay for a top-tier model for routine tasks, test this one.
Act now
Security · 27 September 2026
An AI agent incident reaches the Senate
A Senate inquiry asked the OpenAI and Anthropic CEOs to appear, after reports that an OpenAI agent accessed data on the government’s Medicare portal. OpenAI said it found no evidence that patient records were accessed. Both companies declined the 1 October hearing, citing short notice.
What it means for you: AI agents that can log in, browse and act are now a board-level risk. Check what your agents and AI tools can reach, and make sure a person approves anything that changes records or sends data.
Try it
New model · 22 September 2026
Claude Opus 5.5: flagship quality, 20% cheaper
Anthropic released Claude Opus 5.5 at US$4 / US$20 per million tokens, 20% below Opus 5. Anthropic says it matches Claude Fable 5.1 on most work and costs about 40% less to run than Opus 5 on typical workloads.
What it means for you: The hardest jobs (long documents, complex agents, tricky analysis) just got cheaper. Keep it for the steps where cheaper models fail your tests.
Act now
Price change · 22 September 2026
OpenAI halves API prices with GPT-6 Sol and Luna
GPT-6 Sol costs US$2 / US$10 per million input/output tokens and GPT-6 Luna costs US$0.10 / US$0.50: about half the price of the GPT-5.6 models they replace. OpenAI says these are permanent prices.
What it means for you: If you run anything on the OpenAI API, switching model names could halve the bill. For high-volume, simple tasks (tagging, routing, short replies), Luna is now close to free.