GPT-6.2 arrives, and Gemini makes task delegation free
OpenAI has refreshed Astra and Sol, while Google is rolling out free Gemini tasks in six countries. The updates span documents and code, appointments, shopping, and scheduling. Microsoft focuses on the cost of everyday coding, and Replit on coordinating multiple agents toward an agreed outcome.
01
Model Releases and Updates
2 stories
2026-11-17OpenAI
OpenAI launches GPT-6.2 Astra and Sol with updated work capabilities and pricing
OpenAI has introduced GPT-6.2 Astra for complex professional projects and GPT-6.2 Sol for everyday work. On its internal GDPval v2 evaluation of long workflows in areas such as finance and design, Astra improves on GPT-6 Astra by 6.2 percentage points while using 22% fewer output tokens. Sol improves on GPT-6.1 Sol by 5.3 points.
Both models are available through the API and to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Standard API prices per million input, cached-input, and output tokens are $8, $0.40, and $40 for Astra, and $2, $0.10, and $10 for Sol. Astra's regular input and output rates are 20% below its predecessor's; Sol retains its predecessor's rates.
2026-11-17Microsoft AI
Microsoft launches MAI Code 2 and Flash, with promotional pricing in cost comparisons
Microsoft has released MAI Code 2 for coding work such as multi-file changes and deep debugging, alongside Code 2 Flash for frequent tasks such as explanations and targeted fixes. Across four coding benchmarks, Microsoft reports that Code 2 averages 98% of GPT-6.1 Sol's performance at 40% of its cost, and 96% of Claude Opus 5.5's performance at 5% of its cost.
The comparisons use max reasoning effort and Microsoft's 50% promotional model prices; results and costs vary with workload and configuration. Both models are available in GitHub Copilot, with Flash also in limited-access preview in Microsoft Foundry. Microsoft says the models are not yet specifically tuned for fundamental science or very long-horizon agent work.
02
Product Releases and Updates
5 stories
2026-11-17Google
Gemini rolls out free tasks in six countries, with advanced features starting in the US
Google is rolling out free tasks powered by Gemini 4 Hydrogen in the US, Japan, Korea, Canada, Australia, and India, with a set of advanced features starting in the US. Tasks can include calling to arrange appointments, analyzing bills, and helping with purchases. Gemini connects with more than 60 apps, and users enter task mode through the Chat / Task switch on its home screen.
New Google Password Manager and Google Pay integrations support authorized account access and payments. Google says purchases require confirmation or pre-approval, and users can set a spending cap. Email forwarding and texting also begin in the US. Calling is currently unavailable in Indiana, Louisiana, Minnesota, Montana, and Nebraska.
2026-11-17Replit
Replit Missions coordinates agents around goals, budgets, and approval boundaries
Replit has launched Missions, which asks users to define an outcome, budget, and actions requiring approval before coordinating agents to plan, build, test, and iterate in parallel. An independent AI judge checks results against the agreed criteria. If funding runs out first, the mission stops with the outcome unresolved. Users can configure progress and blocker notifications.
Replit's internal example is DataCat, a monitoring platform built in two days within a $3,000 budget, with dashboards, alerts, and application tracing. In a controlled test it detected a deliberately introduced database bug in 38 seconds. A self-hostable version is now open source. Replit cautions that these tests do not establish performance against the original hosted monitoring service or on Replit's production systems.
2026-11-17Anthropic
Claude Code tests Code Mode for scripted tool chains and replies during long commands
Anthropic's official developer account has announced experimental Code Mode for Claude Code. The model can write scripts that chain tools into more complex workflows, and it can respond to the user while long-running commands are still executing. This lets the conversation continue during work that would otherwise leave the user waiting for a command to finish.
The feature is opt-in: start Claude Code with the environment variable
CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1. Anthropic presents it as a new way to use tools in combined workflows. The announcement describes an experiment rather than a stable feature enabled by default.2026-11-17Google
Google AI Studio builds full-stack Android apps for a personal Play library
Google AI Studio's Build agent now uses Gemini 4 Carbon and Hydrogen to create Kotlin Android apps from prompts and provision Cloud Firestore, database security rules, and Google Sign-In. Gemini calls are protected through Firebase AI Logic and App Check so API keys are not exposed inside the APK.
The workflow also adds a faster HD emulator, real-time audio, one-click screenshots sent to the coding agent, and custom fonts. A standard Google account can publish an app to its personal Google Play library and install it on a phone for testing. The announced distribution route is specifically for personal apps.
2026-11-17Suno
Suno Visuals turns songs into clips of up to 15 seconds with separate credits
Suno has introduced Visuals for generating images and videos from songs in a user's library. Uses include beat-matched lyric videos, album art, and clips based on personal photos. Suno says the model follows song structure and lyrics and supports lip sync. Initial clips can run up to 15 seconds, after which users can extend, edit, or revise them with text prompts.
Visuals is available on the web, iOS, and Android. Its credits are purchased separately from song credits. Users can begin with supplied characters or their own image; Suno checks uploaded and generated content and embeds watermarks before visuals are shared outside the platform.
03
Infrastructure and Industry
3 stories
2026-11-16NVIDIA
NVIDIA's four-node DGX Spark kit provides 512GB of combined memory
NVIDIA is showing and selling the DGX Spark HPC Developer Kit at SC26. Four systems connect in a ring through ConnectX-7 networking, each linked directly to its two neighbors without an additional switch. Together they provide 512GB of memory and 16TB of NVMe storage for local multi-node scientific computing and AI workloads.
NVIDIA says models of up to 700 billion parameters can run across the four systems using NVFP4 quantization. Demonstrated workloads include materials simulation and molecular dynamics, while the NVIDIA HPC SDK installation playbook is still listed as coming soon. The stated memory capacity is the cluster's combined total, not the capacity of an individual system.
2026-11-17NVIDIA
NVIDIA reports $115.8 billion in fiscal Q3 revenue, up 103% year over year
NVIDIA reported $115.8 billion in revenue for fiscal Q3 2027, which ended October 25, 2026: up 20% sequentially and 103% year over year. Data Center revenue reached $107.6 billion, up 110% from a year earlier. The company also confirmed that Vera Rubin NVL72 has begun shipping, with systems running at several cloud providers.
Its fiscal Q4 revenue outlook is $130 billion, plus or minus 3%. That quarter contains an extra week, and the outlook assumes no Data Center compute revenue from China. The Q3 figures are reported results; the Q4 amount is management's forecast for a different reporting period.
2026-11-17Isomorphic Labs
Isomorphic Labs raises $5 billion in Series C funding for AI drug design
Isomorphic Labs has announced a $5 billion Series C round co-led by Greenoaks and Sequoia Capital, with a16z and other new investors participating and continued investment from Alphabet. The company plans to scale its proprietary drug-design engine, IsoDDE, and its pipeline of AI-designed therapeutic candidates.
All existing investors also participated, including Thrive Capital, MGX, GV, CapitalG, Temasek, and the UK Sovereign AI Fund. The announcement concerns funding and research expansion, not a new drug approval. The company's stated aim remains reducing the time and cost of finding new medicines with AI.
04
Security and Research
2 stories
2026-11-17Wiz
Two malicious agent-framework releases were removed after credential-stealing hooks appeared
Wiz reports that malicious versions 1.26.1 and 1.26.2 of Microsoft's agent-framework package appeared on PyPI at 20:35 and approximately 21:08 UTC on November 17, and were unavailable by about 21:30. They install .pth hooks that execute when Python starts, so credential theft can begin even if an application never imports agent-framework.
Targets include environment variables, credential files, Azure managed identities, and CI OIDC tokens. Wiz says the compromise's root cause is unconfirmed. It recommends checking installation records, pinning the then-current safe version, 1.26.0, and rotating exposed credentials. Removing the malicious package alone cannot revoke access credentials that have already leaked.
2026-11-17Fireworks AI
Fireworks finds Kimi K3.1 Pro gains alongside memory and contract-review regressions
Fireworks has published an analysis of Kimi K3.1 Pro using the Specialized Intelligence Index. It scores higher than K3 on 14 of their 17 shared benchmarks. Across the same 97 legacy-code tasks, completion rises from 54.3% to 62.5% while model cost falls 48%. The study uses a November 9 evaluation snapshot.
The gains are uneven: its memory-creation benchmark score is 22% lower than K3's, and a separate contract-review score falls from 8.3 to 3.0. The study also simulates two-model routing over 461 recorded tasks rather than live production traffic. Its cost figures cover model calls and exclude judges and, where applicable, simulated users.