Scoop
AI tool news · rumor vs. reality

The AI Wire ●

Rumors tracked. Announcements verified. Updated daily.

Showing Released · show everything

2026-11-07

RELEASED

Rumor: testers say a fast 'Gemini 4 Hydrogen' is hiding in Arena as gemini-3.8-flash — Google has announced nothing

Testers say Google's next Flash-tier Gemini 4 model, under the community codename 'Gemini 4 Hydrogen', has been turning up in blind Arena matchups under the label gemini-3.8-flash since early November. One account puts the latest checkpoint on November 5. The codename follows the element pattern of Google's confirmed Gemini 4 Argon, announced September 30, and Gemini 4 Carbon, which followed in October — but Google has published no Hydrogen page, model card or announcement.

The demos cluster on speed: community testers report detailed 3D scenes, voxel builds and motion graphics finished in about 1 to 5 minutes, and in one Reddit side-by-side Hydrogen finished a motion-graphics prompt in about 2 minutes against about 11 minutes for Gemini 3.8 Flash. A separate leak reports the string gemini-4-flash-preview in tokenizer code in Google's public SDK repository. Every speed, price and throughput figure in circulation — including a rumoured 300–500 tokens per second and a guessed $0.75/$3.75 per million tokens — is tester speculation, not a measurement Google has published.

As of November 8, Gemini 4 Flash is not out: it appears in none of Google's Gemini API docs, model list, changelog or blog, and there is no official model ID, context window or launch date. Treat Hydrogen as an unconfirmed leak until Google names it.

Update, November 12: confirmed and released. Google published two launch posts on November 12 — Hydrogen is available to developers through the Gemini API in Google AI Studio and Android Studio, in Gemini for business and in Google Antigravity for paid Pro and Ultra subscribers, and it now powers Google Search globally for all users. Google's posts give no price, context window or benchmarks, so the rumored speed and price figures above remain unverified. See the Wire's November 12 release entry.

RELEASED

Google ships ADK 2.14.0: A2A, execution environments and evaluation graduate to stable, with a Teams runner for agents

Google released Agent Development Kit 2.14.0 on November 7 — the release notes are dated November 4 and the tag was published November 7. Execution environments, the A2A agent-to-agent protocol, evaluation and the core runtime modules graduate from experimental to stable, alongside 26 auth, environment, app, CLI and cache modules and 16 toolset and integration features.

The release adds a TeamsRunner that connects ADK agents to Microsoft Teams chats and channels, a MongoDbMemoryService with vector-search retrieval and TTL expiration for agent memory, multi-run evaluation, and an adk generate_agent_card command that builds A2A agent cards from the CLI. A classifier for workflow and tool-gating questions uses OpenAI's Decisions API.

Google claims a 30-node workflow running 200 turns is 23% faster with 9% lower peak memory in this release. The release notes publish no baseline hardware, no comparison version and no second benchmark, so treat that figure as Google's own unverified claim. Stable status sets an API baseline for agent builders; it is not by itself a guarantee that an application built on it is production-ready.

2026-11-06

RELEASED

Salesforce introduces Koa Decisions, a Gemma-based decision model for Agentforce that reads text, images and audio

Salesforce introduced Koa Decisions on November 6, a decision model built from Google's open-weight Gemma that accepts text, images and audio. It is the next version of HyperClassifier, the router already embedded in Agentforce, and it joins the Koa family alongside the separate Koa reasoning model Salesforce built with NVIDIA, covered here in September.

Instead of generating text token by token, the model reads a request plus a user-defined list of categories and returns a selection, a yes-or-no decision or a score in one step — often a single token. The team's research preprint reports 85.5% accuracy on a production intent-routing benchmark, responses in about half a second, and median speed up to roughly 7.5 times faster than the frontier models tested under identical serving conditions. Those are the authors' own measurements; the predecessor's production figure — about 26 milliseconds against 1,446 for a general frontier model — belongs to HyperClassifier, not to Koa Decisions.

Anyone can try the demo after registering for a free Trailhead account, including non-customers, and Salesforce says adding a category such as returns needs no retraining. The wider integrations — Salesforce flows, Apex, prompt templates and a developer API — are planned, not generally available, so the demo does not establish availability beyond Agentforce routing.

RELEASED

OpenRouter rebuilds Auto Router and launches Auto Code, claiming Opus 5.5-level coding at 74% lower cost per task

OpenRouter launched a rebuilt Auto Router (openrouter/auto) alongside a new coding-specialized router, Auto Code (openrouter/auto-code), on November 6. OpenRouter says the new stack outperforms the previous router by up to 4x on its benchmarks, with the routing decision itself about 3x faster at p90 — 0.65 seconds against 1.96. Each request is weighed on eleven factors plus live price, provider health and cache status, inside the caller's guardrails, zero-data-retention settings, allowlists and bring-your-own-key policies. Auto Router already handles more than 2.5 trillion tokens a month, according to OpenRouter.

Auto Code is the headline claim: across five benchmarks it posts an aggregate cost of $1.64 per task against $6.43 for Claude Opus 5.5 at high effort — roughly 74% cheaper at, OpenRouter says, comparable scores, with beta traffic producing a quarter as many invalid tool calls as the old router. Those are OpenRouter's own tests, not independent results.

Existing openrouter/auto callers need no slug change, but they do get a different, pricier-leaning model mix by default overnight; setting cost_tier=low restores the older cheap bias. Teams pinned to Auto for mixed or coding traffic should re-check their cost dashboards before assuming yesterday's routing still holds.

RELEASED

OpenAI ships Codex CLI 0.167.0: an agent command center that flags unread results, and backend switching that keeps your drafts

OpenAI released Codex CLI 0.167.0 as a stable build on November 6, tagged at 01:40 UTC and listed the same day in the ChatGPT and Codex changelog. It is a release about managing and recovering agent work, not about a new model — the notes carry no benchmarks or productivity figures.

Running `codex agents --no-daemon` now opens the agent command center without a background daemon, and both the command center and the session picker mark sessions with unread results, so finished background work stands out. The session picker can switch between connected backends while keeping drafts, saved chats stay readable after a provider change, supported terminals preview local images while composing, and patch details can be expanded during approval prompts.

Two pieces stay gated. A Google Cloud Vertex AI setup wizard sits behind the `gcp_setup_wizard` flag, under 'Sign in another way,' and system-proxy support on macOS and Windows remains opt-in and experimental.

RELEASED

OpenAI rolls out Remix for ChatGPT Sites: fork a teammate's site and revise it in chat

OpenAI began rolling out Remix for ChatGPT Sites on November 6, giving Business and Enterprise users a way to copy a teammate's Site and revise it by describing the changes in chat, then keep editing their own version. OpenAI announced the feature in a post on X, pitching a familiar office problem: a teammate built a good site, and everyone else has notes.

The move pushes ChatGPT deeper into shared office work, where a first version of an internal tool becomes a template others branch from — with editing, hosting and sharing staying inside one paid workspace.

RELEASED

Rumor: a 'Claim your business in ChatGPT' banner with mention tracking surfaces — OpenAI has announced nothing

A banner reportedly spotted inside ChatGPT and posted on November 6 invites businesses to 'Claim your business in ChatGPT,' alongside an analytics view counting how often a business appears in responses. Analyst Juozas Kaziukenas posted the screenshot; he reports the tool measures impressions and clicks daily, and the banner carries a red 'unreleased/beta' label and fine print saying you will need to verify you are authorised to represent the business.

The banner's three cards cover tracking how often your business is mentioned in ChatGPT conversations, adding products and keeping business details current, and letting a business agent answer customer questions with context you control. Kaziukenas calls it part of a larger Business Hub he says he first saw steps of in August, and says it does not show prompts, user data or competitor benchmarks.

OpenAI has confirmed none of it. Search Engine Roundtable, Search Engine Land and RuntimeWire all covered the same single screenshot the same morning, and checks of OpenAI's small-business announcement, business page and release notes found no mention of claiming, analytics or mention tracking. This stays rumor, unresolved.

Update, November 9: Digiday added two unnamed sources with knowledge of the tool and reported that OpenAI is testing the dashboard inside a wider Business Hub that has not launched yet, plotting daily impressions and clicks, still with no prompts, user data or competitor rankings shown. OpenAI declined to comment. Digiday also reported that OpenAI claimed a $1 billion annualised ChatGPT ads run rate by September. The dashboard remains a test, not a launch, so this stays rumor, unresolved.

Update, November 10: Confirmed. Reuters reported that OpenAI launched "Businesses" in beta within ChatGPT, with Business Hub for claiming and managing a profile, aggregated mentions/views/clicks insights, and Business Agents rolling out first to eligible US e-commerce and online service businesses. See the November 10 Businesses launch entry. This rumor is now confirmed.

RELEASED

OpenAI releases BioBasicsBench: 100 wet-lab tasks for AI agents, with GPT-6 Astra at 66.7%

OpenAI published BioBasicsBench on November 6, a 100-task benchmark built with enzyme-design company Ligo to test whether AI agents can handle the routine analysis work of a wet lab. The tasks use raw experimental evidence — gel and blot images, colony plates, microscopy, sequencing files and instrument exports — rather than the cleaned-up figures in published papers, and they ask agents to spot errors and to recognise when the evidence does not support a conclusion.

The tasks span five areas: protein methods, imaging and cytometry, assays and quantification, nucleic acid methods and molecular cloning. Ligo scientists wrote them from their own experiments, each with an expert grading rubric, and every task passed three reviews. OpenAI ran seven of its own models across 37 model-and-effort settings, five attempts per task, for 18,500 scored results.

At the xhigh reasoning setting, GPT-6 Astra earned 66.7% of the available rubric credit, ahead of GPT-6.1 Sol at 62.0% and GPT-5.6 Sol at 43.8%, while using about half as many output tokens as Sol. The score is a mean share of partial credit, not a pass rate.

OpenAI is candid about the failures: Astra picked the wrong colonies from a cloning gel after assuming a larger PCR product meant success, and overcounted a colony plate by about 12%. In the company's own words, higher benchmark scores do not yet mean models can make reliable judgments in the lab. No non-OpenAI models have been scored on the benchmark yet, and OpenAI has not said whether the tasks will be released for others to run.

RELEASED

Meta launches Spaces in Muse: describe an idea and Muse builds a persistent mini-app for it

Meta announced on November 6 that Spaces are now available everywhere Muse is available. Describe an idea to Muse and it builds a persistent, customised experience inside the app — a holiday gift tracker, a daily word game, a DIY planning board — that you can keep refining by asking Muse to add pieces, change the look or tweak how it works.

Spaces live in the app itself: you can pin them to the sidebar or main tabs, hide them, or schedule when they appear, so a weekly planner can show during the workweek and give way to weekend plans on Saturday. Meta's examples run from shopping hubs and music trackers to school dashboards, wedding planners, order trackers and freelance client hubs.

Connected apps let a Space update in the background with price changes, new concert dates or savings progress, and you can check off items and edit directly inside it. Mark Zuckerberg said he had built a BBQ game, an event tracker and a home control dashboard this way. Sharing and collaborative Spaces are still under exploration, so the launch is centred on individual use.

RELEASED

Meta rolls out specialized Muse and Meta AI models with Active Memory — 2.4x less likely to forget critical information

Alongside Spaces, Meta AI Research introduced specialized versions of its existing models for Muse and Meta AI on November 6, trained using learnings from Muse Wonder. The models bring in Active Memory, meant to retain relevant earlier information and distinguish temporary, permanent, current and stale information, and Active Focus, meant to keep work moving toward a goal when obstacles arise by trying harder, recovering or finding alternatives.

Meta's own figures: the current Muse model scores 47% on its Personal Superintelligence Index, up from 38% for the previous model, correctly used critical context in 90% of Active Memory tasks versus 76% before, and is 2.4 times less likely to forget critical information. Meta also reports a 67% improvement in completing everyday tasks when unexpected issues arise.

RuntimeWire notes the rollout reaches all users after Muse passed 10 million setups in North America. In Meta AI, the same changes are meant to break complex problems into steps, produce sharper follow-up questions and accommodate changes of direction mid-conversation. All performance figures are Meta's own claims.

RELEASED

Cursor ships Voice Chat with Agents: realtime voice calls with your coding agent, powered by Grok Voice

Cursor's November 6 changelog adds Voice Chat with Agents — a two-way realtime voice call with the agent in your chat. You can describe a change out loud, ask what the agent is doing while it works, and hear short spoken progress notes and a brief answer when it finishes, instead of reading the full written reply. It works in the Agents Window, the IDE, on the web, and in Cursor for iOS, where a call can be started with 'Hey Siri, call Cursor' and muted or ended from a Lock Screen Live Activity.

Voice chat is separate from dictation, the older speech-to-text input: the voice agent listens and replies, and hands the actual coding work to the agent running your selected model. Speech runs on Grok Voice, with selectable voices including Ara, Eve, Leo, Rex and Sal, and Cursor says that with Privacy Mode on, voice chat runs with zero data retention.

The practical limits are in Cursor's docs: a signed-in account is required — API-key sessions cannot start a voice call — a daily limit on calls applies but the number is not published, and a call hangs up after about ten minutes of silence when no work is running. Updating Cursor is all it takes to try it.

RELEASED

Apple slips an official Xcode plugin into Claude Code: the agent can build, preview and run iPhone apps

An official Apple plugin has turned up in Claude Code's official plugin directory. One command installs it — /plugin install xcode@claude-plugins-official — and it needs a current Xcode on a Mac.

Once installed, Claude Code can build, run and test a scheme, render SwiftUI previews, drive simulators and real iPhones, read build logs and search Apple's developer documentation. That lets the agent see what an app looks like and fix it without the developer tabbing back to Xcode.

Apple has put out no press note. The marketplace entry, which lists Apple Inc. as its author, was merged into Anthropic's claude-plugins-official repository on November 6, and the entry does not ship a separate download: it runs xcrun on the Mac and uses whatever plugin the installed Xcode ships.

RELEASED

Anthropic ships Claude Code 2.1.319: MCP tasks run in the background, and unattended runs can deny instead of prompting

Anthropic released Claude Code 2.1.319 on November 6. The headline change is MCP tasks in the terminal: when an MCP server runs a tool call as a task, Claude Code now moves it to the background right away and tells Claude the result when it finishes, instead of leaving the session sitting inside the call. A new environment variable, CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND, restores the old wait-inside-the-call behaviour.

The release also targets unattended runs. A new setting, permissions.denyInsteadOfPrompting, denies a tool call that would normally stop and wait for a permission prompt and lets the turn carry on, in terminal sessions, subagents and agent teams. Auto mode tightens at the same time: when telemetry or feature flags are off, including behind LLM gateways, calls the server cannot check are now asked about or denied with no fallback to the built-in classifier.

The changelog rounds out with a /claude-api latency-optimize command and fixes for several permission-check bypasses, including a crafted environment prefix in Bash and a sandbox escape during worktree creation. Reporting on the release notes that MCP tasks are a terminal feature — the VS Code extension and the desktop app do not have them.

RELEASED

Amazon lets invited US customers shop by texting Alexa — weeks after blocking Meta's Muse from its store

Amazon has put its Alexa for Shopping agent into text messages. Starting November 6, a small invitation-only group of US customers can text a contact named 'Alexa Shopping' to search for products, compare options, build a cart, complete purchases and get delivery updates, the company announced. Amazon says millions more customers will get access in the coming weeks, and anyone not invited can join a waitlist.

The timing is pointed. Amazon blocked Meta's Muse agent from querying its store in September, and is now rolling out its own texting agent — the same message-an-assistant style Muse uses — just ahead of the holiday shopping season. Alexa for Shopping, formed in May by merging the Rufus chatbot with Alexa+, does not require a Prime membership for its core features.

Amazon says more than 400 million customers have used Alexa for Shopping in the past 12 months, and that use of its agentic features has grown more than eightfold since January. Digiday reports the text threads will carry Sponsored Products ads, though purchases the agent completes on its own stay ad-free.

2026-11-05

RELEASED

Thinking Machines Lab releases Inkling 2 Small — 594B MoE on Tinker, weights now on Hugging Face under Apache 2.0

Thinking Machines Lab, the lab founded by former OpenAI CTO Mira Murati, introduced Inkling 2 Small on November 5, the first model of its second generation. The sparse Mixture-of-Experts system has 594 billion total parameters with 17 billion active per token, takes text and image input, supports variable reasoning effort and a 1M-token context, and is live now on the Tinker platform for fine-tuning and sampling, with inference-only access in the Tinker Playground.

It scores 31 on the Artificial Analysis Intelligence Index, up from 13 for the first-generation Inkling Small, and Artificial Analysis says that once the weights ship it would be the highest-scoring open-weights model developed outside China. Tinker pricing is $0.30 per million input tokens and $1.00 per million output tokens, with cached input at $0.03.

The weights are closed at launch. The company says it intends to publish the full weights under Apache 2.0 in the week after the announcement, conditional on completed external safety testing and red-teaming under its July 31 staged-release framework — a gate it stresses is evidence-dependent, not automatic, since publishing weights is irreversible.

Update, November 14: the promised weights are out. Thinking Machines posted on November 12 that Inkling 2 Small's weights are available on Hugging Face under Apache 2.0 — the main checkpoint in NVFP4 across 25 safetensors files, plus an MXFP8 variant — a week after launch, on the schedule set at release, after the external safety testing the lab cited as the gate. API access has also widened beyond Tinker to Together, NVIDIA, Fireworks, Baseten, Nebius, Modal and Cloudflare, with day-one vLLM support. The lab has not published the results of that safety testing.