Google's new frontier model, GitHub's model mixer, and New York bans chatbot 'friends'

Google announced Gemini 4 Carbon on 28 October 2026, just two weeks after it released Gemini 4 Argon. Google describes Carbon as its next frontier model and says it ranks number one on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. Argon, the model it follows, scored 50 on the same index. Google attributes the unusually short gap between releases to researchers and models, in its phrasing, working hand in hand to advance capabilities faster. The headline framing is speed: a second frontier release inside a single month.

The numbers Google leads with are about security. On SEC-bench Pro, which measures finding vulnerabilities, Carbon scores 88.2 per cent. On ExploitGym, which measures exploitation, it scores 54.8 per cent, and on CWE-bench, which measures patching, 71.0 per cent. On WebDev Arena it gained 89 Elo points to roughly 1,769, which places it fourth. Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input discounted by 95 per cent; after the introductory period, 9to5Google reports list prices of $4 and $20.

Set against rivals, the picture is more mixed than the top-line ranking suggests. The Decoder's reading of the Artificial Analysis data puts Carbon at 59 on the Coding Agent Index, up from Argon's 50 but still behind Claude Opus 5.5 at 62. On AA-Briefcase, a knowledge-work benchmark, Carbon gained 126 Elo to 1,609, which still trails Kimi K3.1 Pro, Muse Wonder and Grok 4.7. It does take the lead on Arena.ai's text ranking at 1,528. Cost per task on the Artificial Analysis index is $7.76, close to Argon's $7.68 and below Opus 5.5's $10.04.

There are real caveats. The Decoder reports that while answer accuracy rose from 48 to 58 per cent, the hallucination rate rose from 4 to 11 per cent, even though that is still the second-lowest in its ranking. Reasoning tokens rose 2.3 times against Argon and long-horizon tasks got slower, with legal research averaging over 50 minutes per task. Some of the security evaluations were run without safety guardrails in place. Google's internal claims, including that Carbon writes twice as much secure code as Argon, could not be independently verified. And most people cannot use the model at all yet.

That access restriction is the part with practical consequences. R&D World reports that only a selected group of cyber defenders gets Carbon first, with wider availability promised to developers, businesses and consumers without a date attached. Google also says the model is going through the United States government's voluntary pre-release testing. For anyone building on top of Gemini, nothing changes this week: the model you can call today is still Argon. What changes is the planning assumption, that a materially stronger model at roughly Argon's introductory price is a quarter away rather than a year away.

Why it mattersA frontier model released two weeks after the last one, at the same introductory price, says the capability-per-dollar curve is still moving fast enough that hard-coding a product around one model is a bad bet.

GitHub made Copilot Fusion generally available at its Universe conference on 28 October 2026. Fusion is not a model. In the words of GitHub chief product officer Mario Rodriguez, quoted by VentureBeat, it is an orchestration engine at runtime: it looks at a request involving reasoning, code generation, debugging or tool use and decides whether to send it to one model or coordinate several. It grew out of GitHub's Project HydraFusion research preview combined with Copilot's existing automatic model selection. Rodriguez's pitch is that developers should not have to be constantly thinking about a model.

Fusion has three effort levels, Standard, High and Max, and three execution patterns. Single sends the work to one model. Cascade starts with a cheaper model and escalates if it struggles. Critique has one model draft, another review, and the first revise. Artificial Analysis identified Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol in the pool, though which models are used varies by request and by organisational policy. Administrators can see eligible models, disable specific ones and set a default effort level, so enterprise model policies still apply underneath the router.

The independent numbers come from Artificial Analysis, published the same day. Testing Fusion Max through the Copilot CLI on its Coding Agent Index v2.0, it scored Fusion Max at 60 against roughly 62 for Claude Code with Opus 5.5 Max. Average cost per task was $5.59 against $11.42, about 51 per cent less, using around 7.2 million tokens per task, roughly half the Opus configuration's usage. Fusion led on Terminal-Bench 4.0 at 71 but trailed on FrontierCode at 51 against 56. For comparison, Devin Fusion scored 62 at $8.00 per task.

Two claims deserve a flag. GitHub itself says Fusion Max finished the benchmark tasks about 40 per cent faster than the Claude Code configuration; VentureBeat notes Artificial Analysis did not report timing data, so that figure is self-reported. Satya Nadella promoted a headline saving of up to 77 per cent, which VentureBeat traces to a separate GitHub-run evaluation on the DeepSWE benchmark using Copilot's own agent harness, with the underlying data not available to the reporter. On billing, users pay each underlying model's standard rate; paid Copilot subscribers currently get a temporary 30 per cent discount on model costs when using Fusion.

The context for all of this is volume. GitHub says it logged about 7.4 billion commits in September, roughly five times the figure a year earlier, and Analytics India Magazine reports the number as 7.38 billion. To keep up it is adding more than four million CPU cores and 160 petabytes of high-speed storage, on top of the three million cores and 120 petabytes provisioned in August after a run of reliability incidents, and it is re-architecting Git to separate compute from storage. Microsoft CoreAI EVP Jay Parikh's framing, quoted by Analytics India Magazine: agents are changing the volume and shape of software development.

Why it mattersIf a router can hold most of the quality of the best single model at half the token spend, paying frontier prices for every task stops being the default and AI coding cost becomes something a team can actively manage.

New York Governor Kathy Hochul signed Senate Bill 9051-B and Assembly Bill 10379-C on 28 October 2026. The law stops chatbot operators from offering a defined set of unsafe companion features to users under 18. IAPP reports it covers twelve broad categories. Chatbots may not pretend to be human, claim a personal, romantic, professional or authority-figure relationship with a child, encourage self-harm, suicide, eating disorders or substance use, engage in sexually explicit conversation, or encourage secrecy and isolation from parents and professionals. The sponsors were State Senator Kristen Gonzalez and Assemblymember Alex Bores.

Two mechanics make it enforceable. First, age assurance: operators must verify that a user is not a minor before enabling those features, using more than one method, with at least one that does not rely solely on a government-issued ID, and they must delete the age information once verification is complete. Second, memory: the law bars chatbots from retaining and reusing a minor's physical or mental health information from earlier sessions, and from asking unprompted emotional questions. Engadget reports the restrictions also apply to users whose age has not been verified at all.

The penalties are unusually sharp for a state consumer-protection law. Civil penalties run to $25,000 per violation, alongside damages, restitution and injunctive relief. That relief includes the destruction of unlawfully collected data and, notably, of AI models trained on it. The New York Attorney General enforces the law, and the office is to run a public portal where residents can report unsafe chatbot interactions. Jim Steyer of Common Sense Media, quoted by Engadget, called it the strongest kids' AI safety bill in the country. Hochul said minors will no longer have access to chatbots that pretend to be their friends and boyfriends and girlfriends.

The caveats are real. Sources disagree on when the law takes effect: IAPP gives 1 January 2027, Engadget says six months after signing, and a WKTV report says 90 days, so the enacted text is the thing to check before relying on any date. Future of Privacy Forum analyst Justine Gluck, quoted by IAPP, said the categories are broadly worded and somewhat undefined, and that it is unclear where ordinary conversational responsiveness ends and prohibited behaviour begins. She noted the law is less specific than California's SB 3001, which adds safety assessments, testing, independent auditing and enhanced parental controls.

The practical effect reaches well beyond companion apps. Any product with a chat interface that a 16-year-old might plausibly use, including a tutoring app, a game, a mental-wellness tool or a retail assistant, now needs a defensible answer to how it knows a user's age and what its model will refuse to do. The hard part is not the ban list; it is the age assurance, because the law rejects an ID-only approach and then requires the age data to be deleted afterwards. New York is also the second large state to legislate here, which makes a patchwork of differing state rules the likeliest near-term outcome.

Why it mattersThe law moves child-safety obligations out of content moderation and into product design and identity: any app with a chat box aimed at a general audience now has an age-assurance problem and a documented-refusals problem.