AI news, 29 October 2026

Google announces Gemini 4 Carbon, but you can't use it yet

Two weeks after Argon, Google announced Gemini 4 Carbon, its new top model, aimed at long-running coding and agent work and at cybersecurity. Google says it now ranks first on the Artificial Analysis Intelligence Index (59 points, a one-point lead) and leads several cyber-security benchmarks; those are Google's own claims. This week it's only open to vetted cyber defenders through Google's Fairwind Program, with developers, businesses and consumers promised access 'as soon as possible' and no date. The API will launch at an introductory $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20. 'It destroys Opus 5.5' reactions on X are enthusiasm, not tests.

Why it mattersIf you build on the Gemini API, a stronger model is coming, but budget for the post-launch price rather than the introductory one, and keep using Argon until Carbon actually opens. Same evening, Alphabet reported Google Cloud revenue up 103% and about 32 billion tokens a minute through its own model APIs (company figures).

Sourcesx.com post ↗ · blog.google ↗ · x.com post ↗ · x.com post ↗

Free API credits if you pay for ChatGPT Pro or Business

Paying ChatGPT Pro and Business customers can now link an API organisation they own and get API credits every month: Pro subscribers get $100, $200 or $500 depending on tier, and Business gets $20 per Standard seat and $100 per Premium seat, pooled and capped at $500 a month per workspace. The subscription has to be at least seven days old; Plus, trials, Enterprise and Edu don't qualify. Credits expire at the end of each billing cycle, don't cover Codex used through API keys, and don't raise your ChatGPT or Codex limits. Separately, OpenAI's help pages confirm the grandfathered Pro 200 plan's usage allowance drops on 30 October while the price stays at $200; OpenAI hasn't published the new limits, and 'about half' is a third-party estimate.

Why it mattersIf your team already pays for ChatGPT Business, you may be sitting on free budget to prototype an agent or automation on the API. Claim it this month, because unused credits don't roll over.

Sourcesx.com post ↗ · developers.openai.com ↗ · x.com post ↗

• ChatGPT (OpenAI)

Claude for Legal gets a gallery of ready-made tools

Claude for Legal (launched in May) now has a gallery of interactive artifacts you open and adapt: cite check, jurisdiction survey matrix, matter board, playbook deviation board, regulatory horizon scan and tabular contract review, plus partner-built ones from Goodwin and Harvey. Each opens with an example; you add your own documents and connected tools, every output links back to its source, and you can share it across the business. Anthropic frames it as Claude's first pass with a lawyer reviewing, and 'lawyers, we are cooked' takes on X are hype.

Why it mattersYou don't need to be a law firm: the tabular review and playbook board patterns work just as well on supplier contracts or HR policies, and the 'artifact as internal tool' idea fits any department.

Sourcesx.com post ↗ · claude.com ↗Steps to try, cost and caveat ↗

• Claude & Claude Code (Anthropic)

GitHub Copilot Fusion: pick an effort level, not a model

At GitHub Universe, Copilot Fusion became generally available on all Copilot plans. Instead of choosing a model you choose an effort level (Standard, High or Max), and Copilot routes or combines models: one model, a cheap draft escalated when needed, or one model critiquing and revising another. GitHub claims Fusion Max scored within two points of Claude Code with Opus 5.5 Max on Artificial Analysis' coding-agent index at about half the cost and in 40% less time. It's billed per model used, with 30% off for a limited time. GitHub also announced a Copilot app for agentic delivery. Spotted by @kimmonismus.

Why it mattersModel routing is becoming the default, which takes the 'which model?' decision off your plate. If your team codes with AI, compare a week's cost per task before and after switching it on.

Sourcesx.com post ↗ · github.blog ↗ · x.com post ↗

• GitHub
• GitHub Copilot

NVIDIA's open transcription model labels who said what

Nemotron 3.5 Transcribe is an open-weight speech model that transcribes 40 languages, translates to and from English, labels speakers, adds word timestamps, handles recordings up to an hour and answers questions about the audio; you can also give it names and jargon to listen for. It's free to use commercially (OpenMDW licence), but it's for recorded audio, not live calls, and NVIDIA's reference setup is a single high-end H100 GPU. The leaderboard rankings are NVIDIA's own. Spotted by @kimmonismus.

Why it mattersPrivate meeting and call transcription with speaker labels, without sending recordings to a cloud tool, matters for law, health and finance clients. Most small firms would use it through a hosted provider rather than their own GPU, and should still check names and speaker labels by hand.

Sourcesx.com post ↗ · huggingface.co ↗ · x.com post ↗

Mistral launches a Responses API that mirrors OpenAI's

Mistral released a Responses API in public preview for building agents. Each response comes back with the answer, the reasoning and any tool calls as separate items, and it follows OpenAI's Responses format, so agent code written for OpenAI is easier to move across. It works with all Mistral-hosted chat models, including Large 4, and it's stateless, so you send the conversation history each time.

Why it mattersIt makes switching an agent between OpenAI and a European provider much simpler, which helps when a UK or EU client asks where their data is processed.

Sourcesx.com post ↗ · docs.mistral.ai ↗

Kylon: shared rooms where your team and your agents work together

Kylon gives a company shared rooms where people, Kylon's own agents and outside agents such as Claude Code, Codex and Devin work as members, hand tasks to each other in threads and wait for a human to approve anything consequential. It connects to the usual business tools (Gmail, Slack, HubSpot, Xero and thousands more) and pricing starts at $50 a month per workspace for 500 credits. Its security audits are listed as in progress, and the customer quotes are its marketing. Spotted by @kimmonismus.

Why it mattersIt tackles a real pain: you being the copy-paste middleman between your coding agent and your team. Agencies can keep one workspace per client. Start with one workflow, like invoice chasing, before inviting everyone.

Sourcesx.com post ↗ · kylon.io ↗

• Kylon

Unify: the agent prepares the call, a person makes it

Unify, an AI outbound sales platform, added a dialer where agents do the pre-call research and talking points, a salesperson makes the call, and the call is logged to the CRM. There's a free plan for up to three seats, then $20 or $60 per seat a month. Spotted by @kimmonismus.

Why it matters'Agent prepares, human closes' is a realistic way to use agents in sales today. In the UK, remember that cold calls and emails to contacts you find this way fall under UK GDPR and PECR rules.

Sourcesx.com post ↗ · unifygtm.com ↗

• Unify

A near miss worth sharing: never auto-approve an agent in someone else's repo

Developer @Damir_Akaza says he nearly leaked production keys after running Claude Code with auto-approve inside an unfamiliar GitHub repository, the kind of setup where hidden instructions in the code can steer the agent. It's one person's account with a small audience, but the risk is well known.

Why it mattersIf you or a freelancer test code from GitHub with an AI agent, do it in a sandbox with no passwords or API keys on the machine, and keep approvals switched on.

Sourcesx.com post ↗

Claude for Legal: a regulatory horizon scan in a few clicks

Verdict: a polished promo rather than a tutorial; beginner level, no coding.

Sourcesx.com post ↗Steps to try, cost and caveat ↗

• Claude & Claude Code (Anthropic)