Google: New Gemini 4 Carbon is Top-Tier and Profits are Soaring

Google Surges to the Top in Coding with Gemini 4 Carbon

On October 28, 2026, Google unveiled Gemini 4 Carbon, its most capable model to date. For now, access is restricted to a select group of cybersecurity companies via the curated “Fairwind Program.” Additional trusted partners are expected to be added gradually. Developers, enterprises, and consumers are set to receive Carbon “as soon as possible” once Google has gathered feedback and refined its guardrails. According to the company, it is also participating in the U.S. government’s voluntary process for pre-deployment model testing.

In the Intelligence Index v5.0 by independent evaluator Artificial Analysis, Carbon ranks first with 59 points under high reasoning effort. Following behind are Claude Opus 5.5 with 58 points, GPT-6 Astra with 57, and Google’s predecessor Argon with 50. In the benchmark overview published by Google itself, Carbon achieves the top score in 15 out of 17 rows. OpenAI retains its lead in a software engineering test, and Anthropic in a computer-use evaluation. On the WebDev Arena, Carbon climbs 89 Elo points to fourth place with 1,768 points.

Google places a strong focus on cybersecurity. On SEC-bench Pro, which measures finding real-world vulnerabilities, Carbon reaches 88.2 percent according to the company. On ExploitGym, which tests whether a model can construct a functional exploit from a vulnerability, it reaches 54.8 percent, and 71.0 percent on the patch benchmark CWE-bench. According to Google, these figures were measured without active guardrails. Internally, Carbon is said to be more than twice as successful at generating secure code compared to Argon. The Chrome team reports patches that are more than ten times better compared to another lab’s frontier model. Google subsidiary Wiz sees Carbon leading in benchmarks for web penetration testing.

Carbon is built for long, multi-step tasks. Google writes that internal teams have seen the model run for days, working through dead ends, identifying root causes in large codebases, and developing patches. According to the company, Carbon is more resilient to prompt injection, takes shortcuts less frequently, and halves the rate of unsubstantiated claims compared to Argon. Google has not yet published a technical report, architecture details, or a parameter count.

Independent tests show limitations. Compared to Argon, Carbon improves significantly in coding and knowledge work, but trails several competitors on professional tasks. In a test of factual knowledge, it records a higher hallucination rate than its predecessor.

At launch, Carbon costs $2 per million input tokens and $10 per million output tokens, with a 95 percent discount for cached inputs. → venturebeat, blog, blog, theverge

Synthszr Take: Thirteen days passed between Argon’s unveiling on September 30 and its broad rollout, with Carbon following 15 days later. Data centers aren’t the bottleneck here; the pace is set by a handful of security firms in the Fairwind Program whose evaluations take about two weeks. As long as guardrails are approved manually, a one-point benchmark lead shrinks to a brief window of time. The next race will be decided by whichever provider can get its models through third-party evaluators the fastest.

OpenAI Sells Astra-Level Performance at a Tenth of the Price with GPT-6.1 Terra

OpenAI has released GPT-6.1 Terra, a model that, according to AlphaSignal, reportedly matches the performance of GPT-6 Astra at one-tenth of the price. The API costs $1 per million input tokens and $5 per million output tokens—half the price of GPT-5.6 Terra and GPT-6.1 Sol, according to the report. Terra sits in the middle of OpenAI’s model lineup and is designed for production use in everyday tasks. According to AlphaSignal, the model can autonomously read screens, click through applications, and fill out forms. → AlphaSignal

Synthszr Take: OpenAI is cutting the price of its predecessor in half precisely in the model tier that drives volume. For agents reading screens and filling out forms, every click consumes tokens. These ongoing costs previously determined whether automating a workflow was economically viable. Now the math works for insurers, banks, and shared service centers handling millions of operations—and that is where the margin ends up. OpenAI earns less per token, betting that volume will grow faster than prices fall (a nod to Jevons).

GitHub Undercuts Claude Code on Price

GitHub has introduced Copilot Fusion, a system that autonomously decides which models to deploy and how they collaborate on coding tasks. The announcement was made Wednesday at the company’s GitHub Universe conference in San Francisco. The product is generally available immediately. Fusion combines the orchestration from the HydraFusion research project with Copilot’s existing automated model selection. Depending on the task, it employs a single model, a cascade with quality checks and potential escalation, or a workflow where one model critiques another’s work. Developers only need to specify the effort level: Standard, High, or Max. → thenewstack.io

Synthszr Take: With Fusion, GitHub becomes a procurer of model capacity. Developers merely select Standard, High, or Max; Microsoft decides which model crunches the numbers in the background. As a result, the halved API costs primarily benefit Microsoft’s margin, because the router directs each task to the cheapest model that passes the quality check. This relegates Anthropic from a branded product to an interchangeable supplier. While enterprises may maintain the model list, the routing logic—the true asset—remains with GitHub.

Amazon Introduces Five New Alexa Devices, from Echo Dot to $500 Echo Show 16

According to its own blog, Amazon has unveiled five new devices, all of which will ship with the latest version of its Alexa voice assistant. The new Echo Dot costs $99.99 and includes a smart home hub and an eero Wi-Fi extender for the first time. For $109.99, there are kids' versions that include a one-year subscription to Amazon Kids+. According to the manufacturer, the Echo Studio with Home Camera for $249.99 can recognize objects in the room, such as a cereal box for reordering. These are joined by the Echo Time bedside alarm clock with sleep tracking for $199.99 and an Alexa Photo Frame with a 10-inch touchscreen. → About Amazon

Synthszr Take: The most revealing line is on the kids' Dot: a ten-dollar surcharge in exchange for a year of Amazon Kids+. Amazon prices these devices like Prime—as an admission ticket to recurring revenue. The Echo Dot brings a smart home hub and an eero extender along, turning it into household infrastructure that people are reluctant to remove. The camera that recognizes a cereal box closes the loop to the shopping cart. For device makers without their own subscription and store, the only thing left in this game is the margin on plastic.

Nvidia releases open transcription model that adapts to technical jargon via prompt

With Nemotron 3.5 Transcribe, Nvidia has released an open transcription model that allows users to provide instructions before processing a recording, reports The New Stack. For example, users can specify that it is a medical conversation in English and supply a list of technical terms. Speaker diarization, questions about the audio content, and the choice between a verbatim or cleaned-up transcript are also possible. According to Nvidia, the model ranks first on Hugging Face’s Open ASR leaderboard for English with an average word error rate of 5.33 percent. Technically, it is a Mixture of Experts model with 30 billion parameters, 4 billion of which are active, making it significantly larger than traditional transcription models with under one billion parameters. → The New Stack

Synthszr Take: Fourteen GPU hours on a single card nearly halved the error rate in the medical benchmark. For those considering a switch, that figure matters more than first place on the leaderboard. Proprietary services like AssemblyAI sell accuracy per audio minute, while customization for technical vocabulary remains with the vendor. With open weights, it belongs to the enterprise itself, as do patient data. Clinics, law firms, and insurers with recorded conversations are likely to be the first to switch. Call centers will stick with their current providers for now, because live streaming is missing.

Meta tests AI Tamagotchi clone in the Philippines

According to The Verge, Meta has quietly launched an app called Waddle World, a virtual pet game in the style of a Tamagotchi. The virtual pixel creatures chat with users and draw little sketches illustrating the conversation. In its listings on the Apple App Store and Google Play Store, Meta describes the app as a “16-bit pet care game” that is actually “a daily journal disguised as a pet.” Users are meant to talk to their pet rather than type, while playing casual mini-games such as stacking groceries or word puzzles on the side. According to the company, there are neither streaks nor scores; the app is intended to be a “warm little escape” from everyday life. It was discovered by social media consultant Matt Navarra. → The Verge

Synthszr Take: Building now costs Meta so little that pet journals and fashion sketches land unannounced in the Philippines and Indonesia. What is scarce is real estate on the home screen. Every new app has to fight for a daily visit, and its fiercest rival is Instagram. That is why Waddle World conspicuously dispenses with streaks and points—coming from the company that turned engagement mechanics into an industry. If it manages to secure daily returns without coercion, Meta gets something far more valuable than clicks: spoken journal entries.

Alphabet grows by 25 percent and Search booms

In the third quarter, Alphabet increased its revenue by 25 percent to $127.6 billion, while profit dropped by 25 percent to $26.1 billion, according to The New York Times. The company attributed the profit decline in part to valuation adjustments in its equity investments, including SpaceX. Cloud revenue nearly doubled, surging 103 percent to $30.7 billion and contributing significantly to growth. Search rose 12 percent to $63.3 billion, which was below analyst expectations according to Richard Kramer of Arete Research. Capital expenditures for data centers and other infrastructure totaled $51.4 billion for the quarter, leaving the full-year forecast of $195 to $205 billion unchanged. → www.nytimes.com

Synthszr Take: Pichai reports search queries, not clicks, and the choice of metric reveals the new division of labor. Every additional query lands first in a Gemini response, assembled from publisher content and delivered on Google’s own page. The release said nothing about increasing traffic for publishers. The increase in commercial queries hits them even harder: that is where their affiliate and advertising revenues used to live, and that is where Google is now posting 12 percent growth in its search segment.

Microsoft’s profits surge as it discloses Azure figures for the first time

In the first quarter of its fiscal year 2027, Microsoft generated $92.3 billion in revenue, reporting across just two segments for the first time, according to The Verge. Revenue rose 19 percent, while net income grew 32 percent to $33.6 billion. The previous segments—Productivity and Business Processes, Intelligent Cloud, and More Personal Computing—have been phased out. Replacing them are “Agents and Infra” and “Devices and Consumer.” Agents and Infra combines Azure, Microsoft 365 for commercial and consumer customers, and server products, accounting for $76.8 billion, or 83 percent of total revenue—an increase of 25 percent. → www.theverge.com

Synthszr Take: Microsoft delivers the Azure revenue figure that analysts have demanded for years, while simultaneously depriving them of an answer on margins. Agents and Infra blends data centers requiring massive capital expenditures with Office licenses that have generated near-pure profit for decades. The costs of agent infrastructure thus disappear into the blended average. Amazon showed with AWS how effectively such a reshuffle obscures segment-by-segment comparisons. The only thing that remains visibly isolated is what is shrinking: Windows, with an announced decline in the mid-30 percent range.

SK Hynix reports record profit and expects memory shortage until 2030

According to Quartz, SK Hynix reported a record operating profit of 83.33 trillion won in the third quarter of 2026 and expects demand for memory chips to outstrip supply until 2030. Revenue rose 32 percent quarter-on-quarter to 104.62 trillion won, surpassing 100 trillion won for the first time in company history. Operating profit was 632 percent above the previous year’s figure, beating analyst expectations of 76.72 trillion won, with operating margins reaching 80 percent. The company cited the technical difficulty of cutting-edge manufacturing processes and long lead times for new fabs as reasons for the shortage. According to SK Hynix, it has signed multi-year supply agreements with 14 key customers, which together cover more than half of its production capacity for 2027 and 2028. → Quartz

Synthszr Take: More than half of the capacity for 2027 and 2028 is already sold to 14 customers. This means the AI roadmaps of the coming years will be decided by SK Hynix’s order book. New fabs take years to build, and additional capital will do little to relieve this bottleneck. Everyone else will buy the remainder at prices that explain an 80 percent operating margin. Compute discipline is turning into a procurement issue, and the 14 contract customers are pulling ahead.

Anthropic Buys a 1-Person Biotech Startup Ahead of Its IPO

Anthropic has acquired Mithril Medicine, a biotech startup operated solely by founder Cody Tranbarger. The acquisition was announced on October 27, one day after filing its IPO prospectus. According to Mithril’s own technical report, the acquisition was already closed in July 2026. Backed by Atlas Venture, Mithril develops therapies for two rare forms of muscular dystrophy, centered on LAMA2-related congenital muscular dystrophy. Tranbarger is joining Anthropic, continuing existing programs and planning to launch new ones.

Tranbarger calls his operating model 'pipeline-in-a-person.' Claude coordinated contracts, logistics, and experiments across eleven external laboratories, with the report also listing eight paid consultants. According to Anthropic, one program progressed from concept to encouraging results in mice in less than 18 months. Treated animals showed roughly twice the grip strength of the control group. The report is not peer-reviewed. Treatment was randomized by litter rather than individually, which is why the authors recommend repeating the study with a blinded evaluation. Anthropic considers initial human trials possible by late 2028 at the earliest.

Mithril documents one workweek in detail, which the company itself selected as its best. During it, Claude evaluated a 341-page production report covering 113 antibody constructs in under eight minutes. Following an objection from a researcher, the model reduced the testing plan to 89 constructs and adjusted the associated documents in four minutes. Of 17 drafted emails, twelve were sent after human review. The report records six errors, at least three of which reached external laboratories, and two remained undetected for over four days.

In parallel, Anthropic presented interim results from six research partnerships: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines, and The Column Group. At Generate, researchers used Claude Code to design a component of a CAR-T-Zelltherapie that is roughly 70 percent shorter and produced about ten times as many receptors in the laboratory. For Divergence, Claude generated around 35,000 antibody designs in roughly a month. According to the company, three of them bound the allergy molecule IgE better than the approved drug Xolair. At Enveda, Claude reviewed 1,061 pages of manufacturing documentation, finding all eleven deficiencies identified by human reviewers plus seven additional ones, reducing review time by 89 percent. The Column Group is launching a separate company, Pleiades, to deploy Claude in drug development programs. Anthropic noted that the work remains predominantly preclinical.

The acquisition comes amid preparations for an IPO on Nasdaq under the ticker ANT. → WinBuzzer, Pharmaceutical Technology, The Information, The Neuron, Quartz, Nate’s Substack, Anthropic, RuntimeWire, blockchain, unite

Synthszr Take: Completed in July, announced one day after the prospectus. The timing reveals the intended audience. Mithril gives investors a second revenue narrative in which Claude develops proprietary drug candidates. The evidence is thin: a self-selected best week, an unblinded mouse study. Yet it is enough for a valuation premium, because no one will test whether the pipeline holds up in humans before late 2028. Meanwhile, Bristol Myers Squibb and the other partners are looking at a future competitor in drug discovery programs.