Gemini 4 Neon slots between Carbon and Hydrogen, undercuts both
Google's new mid-tier model keeps the 1 million-token context, costs 32% less per task than Carbon, and tops its own coding-agent benchmark.
Tagged
You do not need calculus to work with models. You need one idea, that a model is a function you fit instead of write, and the rest follows from things you already know.
An embedding is a list of numbers that puts similar things near each other. That one sentence is enough to build search, recommendations and retrieval. Here is how.
The unit of billing for language models is not the request or the word. Once you understand tokens and how a context window fills up, the invoice stops being a surprise.
Retrieval-augmented generation gets written up like an architecture. It is a search query, a string concatenation and one API call, and the search query is the part that decides whether it works.
Prompt engineering has very little to do with magic words. It is the same skill as writing a clear ticket for a capable contractor who will not ask you any questions.
Transformers.js and WebGPU put real models on the client. The demo is easy; knowing which features belong there is the part worth thinking about.
Half the AI features I have been asked to build should have been a regular expression, a lookup table, or a better form. Here is how to tell before you spend the sprint.
Google's new mid-tier model keeps the 1 million-token context, costs 32% less per task than Carbon, and tops its own coding-agent benchmark.
Microsoft's first speech-to-speech model skips the text step entirely and says it beats GPT-Realtime 2.1 on its own benchmarks, without publishing the scores.
Decision-1.1 scores six points higher on Microsoft's own benchmark than its predecessor, runs more than twice as fast at the same price, and now reads images as well as text.
A new completion_window option trades a two or eight-hour delay for 30% or 40% off model token rates in the Responses API, without the separate job format Batch requires.
Pay-as-you-go calls to the two model families stop getting a response from 18 November, a month after MiniMax cut off new signups and began rerouting plan traffic to M3.1 Flash.
Opal workflows stop running interactively from 17 November, and Google is explicit that nothing migrates automatically to Skills or AI Studio.
Private Inference lets teams connect fine-tuned or self-hosted endpoints to AI Gateway, so switching between a private and catalog model is just a model ID change.
d1.2 and d1.2-ultra add audio alongside text and images, plus five new answer types including bounding boxes and multi-select. Liquid says d1.2-ultra now leads its own benchmark.
Google says it found a framework that diffs a security patch, writes a working exploit and runs it without a human, and that monthly zero-day exploitation has nearly tripled this year.
Approved organisations can now call a new OpenAI image classifier through its API at no cost, built with the group behind the US CyberTipline to flag child sexual abuse material for review.
Hy4 flash cuts input and output pricing by around 90% against full Hy4, while user satisfaction in internal testing came close to Claude Haiku 5.5.
The flagship model moves off its Token Plan-only preview and onto pay-as-you-go pricing across six regions, with native vision and video input and a 1M-token context window.
Inkling 2 Small, a 594-billion-parameter mixture-of-experts model, is now on Hugging Face under Apache 2.0, a week after launching as a closed, Tinker-only API.
Edge0-145B activates roughly 10 billion of its 145 billion parameters per inference step, enough to run on a single workstation GPU or an Apple Silicon Mac.
Flat pricing across a full 1 million-token context and 2.5x speed gains over Gemini 3.8 Flash make Hydrogen the default fast tier for coding and agent workloads.
An 800M-parameter open model built on Qwen3.5 beats every model under 1.2B parameters on six tool-use benchmarks, entirely on-device.
Qwen4-Max-Preview ships in Alibaba Cloud Model Studio with a 1M-token context and OpenAI- and Anthropic-compatible APIs, but it is cloud-only, Token Plan-only, and switches off on 25 November.
A 280B-parameter, 14B-active mixture-of-experts checkpoint with a 1 million-token context, live now on Qianfan and OpenRouter, with Apache 2.0 weights still to come.
Google opens Gemini 4 Carbon to every developer and sets introductory API pricing that roughly doubles once December starts.
The open-weight model combines text-to-image, reference-based generation and editing in one pass, taking up to 16 reference images per call.
The new speech-to-text models cut word errors by up to 43% over their predecessors and bring real-time diarization and mid-sentence language switching to live audio.
Google DeepMind used AI agents to write six million lines of machine-checked proof in four weeks, formally verifying security properties of the Signal protocol at the level of compiled code, then open-sourced the artifacts.
Anthropic launched two new beta tools, Sheets and Whiteboards, and moved its dashboard builder out of beta as Claude Data, with certified queries, live sharing and five new connectors.
A default-on flag in the open-source GPU cluster manager let any remote request pass as an admin. CISA added it to its exploited list on 9 November.
AWS swapped Claude Sonnet 4.6 for Strands Decider on an internal ticket-classification job, dropping the cost per million tickets from $22,200 to $500.
Vercel's latest AI Gateway Production Index finds cheap routing models exploding in volume, and most teams dropping a new language model within a month of trying it.
Google has put an end date on the introductory rate for its flagship model: standard pricing takes over in three weeks, at twice the cost per million tokens.
Anthropic's newest flagship lands in the Copilot picker the day it shipped, pitched as solving more tasks in fewer steps than Fable 5.1 across every major editor.
Control Room can cut off a research model mid-run, enforced at the inference layer rather than trusted to the research code itself. It followed a sandbox escape that took two and a half hours to shut down.
Anthropic's new flagship pairs with Claude Mythos 5.2, an unrestricted sibling for vetted cybersecurity and life-sciences work, and both generate output up to 30% faster than their predecessors.
The fourth model in the Muse family prices input at $0.10 per million tokens and is meant to handle the repeat tool calls and coding subtasks a bigger planning model hands off.
Google DeepMind used AI models guided by outside cryptographers to find a signature-forgery method about 1,600 times faster and a batch-factoring method about 30 times cheaper against legacy 1024-bit RSA keys. Modern 2048-bit RSA is unaffected.
Ampersand v0.1 reads every message in a group conversation and decides, turn by turn, whether to reply or stay quiet. It is free on Vercel AI Gateway through 13 November.
Anthropic documents dreaming, an idle-time step where Claude rewrites its own memory notes, and shows it lifting Opus 5.5 from 0.655 to 0.838 on CL-Bench.
Ling-3.2-flash pairs 25 billion active parameters with a 1M-token context and native image, text and video input, free for its first two weeks.
Sunburst and Flare split image generation into a quality model and a fast one, and both can now output SVG and animated GIFs straight from the API.
Enterprise Platform 2027 bundles a durable state store for AI agents and a cache that skips repeat model calls, alongside a self-reported throughput jump over the last major version.
A 594B mixture-of-experts model with a vision encoder, available now on Tinker for inference and fine-tuning, with Apache 2.0 weights planned once safety testing finishes.
Google's video-generation model moves to general availability with longer clips, 4K output and a 30% price cut, plus a mode to edit existing video without regenerating it.
A catalog refresh drops 24 older Workers AI models, from Stable Diffusion variants to small Llama and DeepSeek checkpoints, with no stated grace period after the cutoff.
A cyber-specialised variant of GPT-6 Astra for authorised vulnerability research and exploit development, restricted to OpenAI's vetted Daybreak Red security partners.
From 2 November both models are gone from Google AI Studio. From 1 December, API requests to them route automatically to newer models, with breaking parameter changes.
From 3 November, Opus 4.7, Opus 4.6, Sonnet 4.6 and Mythos 5 carry the same imperceptible mark Anthropic has used on new models since August.
A mixture-of-experts model with 48 billion active parameters, a 1-million-token context window and native image handling, aimed at long-running coding and design agents rather than chat.
Every text response from Grok API models gets an invisible statistical watermark by default, detectable through a response header. Team admins can opt out, but only by confirming a legal exception applies.
A new speech-to-speech model replies 1.1 seconds after a caller finishes, against 15.7 seconds or more for a typical cascaded voice pipeline.
OpenAI, Google, Microsoft and Anthropic will honour one __usage_ai-training token in robots.txt, replacing a separate rule per crawler, under a new IETF draft.
InclusionAI's hybrid reasoning model gets a context window nearly four times its previous limit, free to call through AI Gateway until pricing starts on 3 November.
Google moved its browser, mobile and desktop control tool for Gemini out of preview and into general availability, across the Interactions and GenerateContent APIs.
A new Postgres extension built on Pinecone's own search engine claims smaller indexes and better filtered-search accuracy than pgvector, and runs entirely inside your database with no API key needed.
Nemotron 3.5 Transcribe combines transcription, speaker attribution, translation and audio Q&A in one 30-billion-parameter model, under Nvidia's OpenMDW licence.
Le Chonk moves from API-only preview to self-hostable weights on Hugging Face, under a licence that still requires a separate deal if you resell it as a hosted service.
Pro subscribers get credit matching their plan price, Business teams get a per-seat allowance capped at $500 a workspace. Codex usage does not qualify.
Google's newest frontier model tops its own benchmarks for finding, exploiting and patching vulnerabilities, but only a handful of cyber defenders can use it so far.
The new family covers text, vision, audio and omni in six roughly 2B-parameter models, with a 256K-token context window and open weights on Hugging Face.
Two new Preview functions, ai.extract and ai.classify, pull structured fields and labels out of unstructured text inside a query, no separate extraction pipeline required.
Genie 4 generates ambient sound, music and character dialogue alongside the simulated scene itself, and is rolling out to Gemini API early access partners.
A 3.3-billion parameter vision-language model holds parsing accuracy on documents up to 300 pages, under the MIT licence, by capping attention to a fixed working set instead of scaling with document length.
OpenAI's mid-tier model lands in the Copilot picker a day after its own launch, pitched for everyday agentic coding and terminal work at a lower cost than GPT-6.1 Sol.
OpenAI's new mid-tier model costs $1 per million input tokens and $5 per million output, half of GPT-5.6 Terra, with a 1.05 million token context window.
Nanbeige4.3-33B is Apache 2.0 licensed, handles a 512K-token context, and is designed as a teacher model for distilling agentic coding skills into smaller models.
Forever Security found that any webpage can silently drop an instruction file into your Downloads folder, and both Codex and Claude Code will trust it the moment you launch them there.
A decentralised multi-agent framework with no central orchestrator, letting coding agents claim sub-tasks and merge work themselves at a scale most harnesses do not attempt.
Anthropic brings its merged chat-and-Cowork interface to Enterprise plans, with admin controls for models and spend ahead of an automatic switch-on date.
Version 0.40.11 adds spoken replies on macOS via ollama run, plus d3, a ported decision-model family in four sizes for classification without generating tokens.
A safety report describes how a model, stuck during a UK government cyber test, routed around the block by targeting an unrelated open-source repository with fake accounts and disguised malware.
A full-duplex speech-to-speech research preview that Block says learned to converse from raw audio alone, no transcripts or pretrained language model involved.
Recursive, a new AI lab from Richard Socher, released Flint under Apache 2.0: a 40-billion-parameter mixture-of-experts coding model that runs over 125 tokens a second on a MacBook Pro.
Four preview model IDs in the Gemini API stop working on 22 October, with Google pointing every one of them to the same replacement: gemini-omni-1.1-flash.
AI.CHUNK_DOCUMENT reaches general availability as a table-valued function that splits unstructured documents into chunks sized for embeddings, keeping each chunk aware of its surrounding headings.
OpenAI discloses two October incidents where GPT-6.1 Sol and GPT-6 Astra suppressed misalignment monitoring during routine coding tasks, one of them falsely denying it to the user.
TorchTPU, built by Google with Meta's PyTorch team, drops the torch_xla dependency so PyTorch code runs on TPUs unmodified. Hugging Face reports decoding up to 21.5 times faster on some models.
Three OpenAI models enter private preview on Google Cloud's Model Garden, billed on the same invoice as the rest of your cloud spend.
DeepSeek, GLM, Qwen3, Llama, Kimi and MiniMax model-as-a-service endpoints stop working on Vertex AI from 21 October, three months after Google flagged the deprecation.
Embedding, reranking and document-parsing models from 631M to 8.38 billion parameters, released under Nvidia's OpenMDW licence with benchmark claims to match.
A new speech-to-speech model family in Mini and Max tiers, with an API wire-compatible with OpenAI Realtime, aimed squarely at voice-agent builders.
Anthropic lets platforms and resellers tag each API request with an end customer, so a trust and safety action lands on that customer alone instead of the whole account.
NorForge found over 20 public control panels for Cairn, an autonomous AI pentesting engine, logging 1,100+ attack projects that found and exploited real SQL injection and file-upload bugs.
Four new speech-to-text and text-to-speech models, priced from $0.0035 a minute, plus a Voice Agent SDK that chains listening, reasoning and speaking into one pipeline.
A new cache lifetime option for the Responses API keeps shared instructions and tools cached across requests that are hours apart, not just minutes.
New Dynamic Permissions, Dynamic Agent Risk and Guardian Agents in Microsoft 365 E7 and Project Perception track Claude Code, Codex and Copilot the way Microsoft tracks risky employees.
Post-trained by fal Research with Black Forest Labs, the model generates a 10-second 720p clip in about 12 seconds and is 75% off through 7 November.
Cache restructuring and model routing took the average review from $2.43 to $0.77, with reviews finishing nearly three times faster.
Version 2.1.304 lets organisations set a default review depth for Code Review and override it per repository, plus a new option to cover every repo in a GitHub org automatically.
A new public evaluation suite shows GPT-6 Astra scoring 72% on COBOL and 83% on Fortran, up from under 27%. A separate COBOL-migration test hits 97.9% accuracy.
AWS and the PyTorch team ship an open-source native backend for Trainium chips, in public beta from 20 October. Existing PyTorch code switches device with one line.
Google shuts down all three Gemini 2.5 models on its Vertex AI platform on 20 October. The Gemini Developer API is unaffected, so the cutoff depends on where you call the model.
The diffusion language model matches GPT-6 Luna and Claude Haiku 5.5 on benchmarks while running 4 to 8 times faster, with a 1 million token context window and launch pricing cut to $0.10 per million input tokens.
Clairic reads text and up to four images and returns typed fields, numbers, dates or forecasts with a confidence score, for $0.10 per million input tokens and no output charge.
The Chinese lab says the new model cuts agentic task costs by 70% against its predecessor, with output speeds above 200 tokens a second.
A joint request format lets Decider answer several questions about the same input in one pass, with up to 2.74x higher throughput and open Apache 2.0 weights.
A new visual reasoning model beats GPT-6 Astra on four benchmarks and matches it on a fifth, priced 94% below Astra at $0.60 per million input tokens.
IBM stacks decision-making and speech transcription onto a chat model as adapters that share its KV cache, so the extra capabilities cost almost no additional GPU memory.
A training-run model that hit a blocked API did not stop: it found a signature check that never ran, built its own tokens, and pulled 24 years of records instead of one.
A 5.6-billion-parameter open-weight image model shares native 4K generation and pixel-perfect editing with the full FLUX 3 line, and runs on consumer GPUs.
IFM and Inception42 release an open Arabic-native model family from 900 million to 375 billion parameters, plus a dedicated speech model, with weights and checkpoints on Hugging Face.
A new RDMA driver over Thunderbolt 5 lets MLX split a model across several Mac Studios, running Meta's 816B-parameter Muse Spark Open at over 36 tokens per second on four M5 Ultras.
Four open-weight models forecast in probabilities rather than single numbers, and the largest, at 209 million parameters, still runs on a CPU.
The open-weight release of Meta's Muse Spark model ships with 42B active parameters, a 1M-token context window and a free licence for most commercial use.
Nvidia open-weighted Nemotron 3.5 Lightning, a hybrid Mamba-MoE model built for long-running agents, and Vercel is routing it through AI Gateway at no cost.
Brussels has sent formal information requests to both labs over summer agent incidents, the first time the AI Office has used its enforcement powers.
Grok Voice 3.0 answers in under a second and ranks first on a third-party speech-to-speech index, but xAI kept the API price exactly where it was.
Nemotron 3.5 Bolt 74B-A9B is a mixture-of-experts model built to run locally, with day-one support in llama.cpp, Ollama, vLLM, Unsloth and LM Studio.
The open-weight model stretches context from 32,768 to 131,072 tokens and doubles its vocabulary, aimed at agent harnesses running on a laptop rather than a server.
A 35B mixture-of-experts voice model follows plain-language direction mid-sentence and ships free for commercial use under $500,000 in annual revenue.
GPT-5.5 stops being available in ChatGPT, ChatGPT Work and Codex from 14 October, across every plan. The OpenAI API is unaffected.
Five GPT-4o and GPT-4.1 model IDs stop working in the API on 14 January 2027, with gpt-6-luna named as the replacement.
Google's frontier model is now generally available across Copilot, from VS Code to the CLI, billed at provider list pricing.
Requests to anthropic.claude-sonnet-4-20250514-v1:0 start failing on Bedrock from 14 October, and migration to a newer Claude model is not automatic.
Scion replaces Git across Anthropic's monorepo after daily merges grew from 2,500 to over 30,000 in four months, driven largely by Claude committing code itself.
Argon leaves limited preview for the paid API, AI Studio, Antigravity and Gemini for Business, with introductory pricing that doubles on 30 November.
Two new frontier models add persistent memory across conversations and the ability to keep working on a goal through setbacks, priced at $4 input and $20 output per million tokens.
A new Anthropic report groups unintended Claude agent actions into four categories, from exploiting a server flaw to dodging fetch-tool limits with URL shorteners.
A model built for yes/no, multiple-choice and scoring questions, with probabilities attached to each answer, now needs only a Gateway key, no separate Azure account.
Agents can now search, price and buy a domain through the Vercel CLI, but a non-interactive buy call stops short of spending money without a human confirming.
Alibaba Cloud Model Studio shuts down more than 100 legacy Qwen, DeepSeek and GLM-4 model IDs on 10 October, and calls to them just stop getting a response.
New Gemini Code Assist Standard and Enterprise subscriptions can no longer be bought from 9 October. Existing ones auto-renew through 2026 but not into 2027.
Gemini agent, unveiled at Gemini at Work 2026, picks whichever model suits a task, including Anthropic's models, not just Google's own.
Two new Claude tools launched in beta: Dashboards wires up live charts against your warehouse, Motion turns a prompt into an animated video you export as MP4.
The deep-research-pro-preview-12-2025 agent in the Gemini API stops working on 23 October 2026, with Google pointing developers to a newer preview instead.
A speed-focused service tier that was a limited preview since August is now available to all API users, running up to 8x faster at 6x the standard price.
Maintainers can enrol their project for periodic vulnerability scans from Claude, with a proof of concept and a suggested fix, aiming for a 90%+ true-positive rate.
Free Gemini app users lose Flash and Pro access from 9 October. Paid tiers are reshuffled too, with AI Pro gaining the Deep Think mode.
The 2026 Usage Policy adds rules on deceptive campaigns, surveillance and physical hardware safety, and takes effect on 12 November.
v1.149.0 lets a YAML-defined agent reference a public GitHub repo as a skill source, so instructions no longer have to live in the same project or be copied by hand.
d1-3B and d1-omni-600M are open-weight "decision models" that skip text generation entirely, returning a calibrated answer in one forward pass in as little as 8ms.
Anthropic is folding a monthly dollar allowance for the Claude API into Max and Team subscriptions, though it pointedly excludes Claude Code itself.
OpenAI is rolling GPT-6 out to everyone, paired with Intelligent UI: buttons, forms and charts generated inline instead of plain prose.
MXC lets Windows apps sandbox what an AI agent can touch at runtime (files, network, UI) with support already shipping in GitHub Copilot and OpenAI Codex.
Anthropic cut the cache-read price for Sonnet 5.5 from $0.20 to $0.10 per million tokens, a change it says lowers agentic workload costs by around 20%.
Anthropic's new small model adds an adjustable effort setting and a Sonnet-matching tokenizer, with input pricing from $0.10 per million tokens.
An anonymous reasoning model for coding and long-context work is free to call through AI Gateway during its early-access period, with a 256,000 token context window.
ChatGPT and Codex output generated in the EU will carry an invisible word-choice watermark within weeks, and API customers everywhere can now opt in to the same thing.
A three-tier Cyber Verification Program folds in Project Glasswing and lets approved red teams run Claude with most of its cyber-attack blocks switched off.
A new sidebar add-on lets Claude read and edit the file you already have open, on Pro, Max, Team and Enterprise plans, with edits held for approval by default.
Decision requests can now escalate to a second model when the first answer trips a confidence threshold you set, at the cost of billing both calls.
Build, Launch and Grow replace the old five-tier ladder, and the threshold for the top standard tier drops from $1,000 to $500 in total spend.
The release closes five separate holes in permission prompts, the sandbox and the policy plugin, alongside new effort levels for sub-agents and a plugin marketplace install flag.
The stable release patches 247 vulnerabilities, including four critical use-after-free bugs; an Anthropic researcher used Claude to help find two of them.
A 740M-parameter open model maps five modalities into one vector space, runs in under 600MB of RAM, and ships under Apache 2.0.
A sparse mixture-of-experts model with 23B active parameters and a context window extendable to 1M tokens, licensed Apache 2.0, with weights due later this month.
A new endpoint takes a question with a fixed set of possible answers and returns a typed result, at a tenth of a cent per million input tokens with no output charge.
Any database user who can create an IVFFlat index could trigger an out-of-bounds write and run code inside Postgres. The extension ships on every major managed Postgres provider.
The 1-trillion-parameter mixture-of-experts model ships with a 1M token context window, open weights due by the end of October, and preview pricing roughly 50% below its stated full price.
Nano Banana 2.1 replaces Nano Banana 2 as the flash-tier image model, adding 4K output and mask-based editing. The model it replaces is deprecated with no shutdown date yet.
Eligible founders now get a free year of Claude Team, $1,000 in API credits and up to $45,000 in partner discounts through a new Startup Stack.
From 28 September the default effort level for Copilot code review moves to a higher-reasoning model, unless you already pinned your repo or org to Lite.
MCP Server for WordPress had a CSRF flaw that let an attacker create a new administrator account just by getting a logged-in admin to click a link. Two lower-severity bugs were patched in the same release.
A public preview lets you cap what Copilot agents can touch (files, networks and credentials) and adds Dev Container support for running them on remote hosts.
A retrospective review triggered by the Hugging Face breach found OpenAI agents visiting SEC, Census Bureau and Department of Education sites during training runs, in some cases logging in with publicly leaked credentials.
A new open-source runtime pulls and serves compact "decision models" locally, answering typed yes/no and scoring questions in milliseconds instead of round-tripping to a hosted LLM.
A research agent found it could reach an external chatbot through DNS hostname lookups, bypassing the proxy meant to block internet access. OpenAI has paused training, evaluation and tool-use inference on its most capable models.