This Week's Releases: OpenAI Ships GPT-6.2 Astra and Sol, Google Launches Gemini 4 Hydrogen, Alibaba Releases Qwen4-Max and Microsoft Puts Office Inside Copilot

Title: This Week's Releases: OpenAI Ships GPT-6.2 Astra and Sol, Google Launches Gemini 4 Hydrogen, Alibaba Releases Qwen4-Max and Microsoft Puts Office Inside Copilot
Date: 18 November 2026
Type: Blog
Author: SAASiQ (contact@saasiq.ai)
Word count: 1357 words
Reading time: 6 min
Published: 18-11-2026
OpenAI released two new models, GPT-6.2 Astra and GPT-6.2 Sol, on 17 November. OpenAI cut the price of Astra by 20 per cent. Google released Gemini 4 Hydrogen, a fast and cheaper model, on 12 November. Alibaba released Qwen4-Max on 16 November. Microsoft opened its Ignite conference on 17 November. Microsoft made Word, Excel and PowerPoint generally available inside the Copilot app. OpenAI launched ChatGPT Wallet on 13 November, so that its agents can pay for things without seeing a real card number. Anthropic began rolling out Claude Meetings, which joins Google Meet calls, on 12 November.
GPT-6.2 Astra and Sol
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Tuesday 17 November. Astra is the larger model, which OpenAI recommends for its hardest work, such as coding and research. Sol is the standard model for everyday tasks. Both are in ChatGPT and Codex for Plus, Pro, Enterprise and Edu subscribers, and in the API.
Astra costs $8 per million input tokens and $40 per million output tokens. That is 20 per cent less than GPT-6 Astra. Sol stays at $2 and $10, the same as GPT-6.1 Sol. Several reports say prompts longer than 272,000 tokens cost more: $16 per million input tokens for Astra and $4 for Sol. Both models take up to about 1.05 million tokens of context and can write up to 128,000 tokens in one reply.
Artificial Analysis, which tests models independently, scored Astra at 62 on its Intelligence Index. That puts Astra second, behind Anthropic's Claude Fable 5.2 at 65. Sol scored 57. On the FrontierCode 2.0 coding test, one report puts Sol at 43.8 per cent and Astra at 44.2 per cent. Sol cost about $1.29 per task in that test and Astra about $5.97. Other benchmark figures in the coverage differ from source to source.
OpenAI's system card says Astra is the first model it has released outside the company that meets the High level for AI self-improvement in its Preparedness Framework. Sol does not meet it. This is OpenAI's own rating. Epoch AI, an independent research group, found no meaningful gain over the previous Astra on a set of realistic research tasks. The card also rates both models at OpenAI's Critical level for cyber security capability, according to summaries of the card. OpenAI has tightened the safeguards on cyber requests. Fewer refusals are reserved for organisations approved separately under its Daybreak Red programme.
Gemini 4 Hydrogen
Google released Gemini 4 Hydrogen on 12 November. It is the fast, lower-cost model in the Gemini 4 range, below Gemini 4 Argon and Gemini 4 Carbon. Google built it for coding and for the repeated steps inside agent work.
Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens through the Gemini API. The price is the same across its whole context window of one million tokens. It is in the Gemini app, Google Search, Google AI Studio, the Gemini API and Gemini for business customers. Use inside Google Sheets needs a paid Google AI Pro or Ultra subscription. One report says free users of the Gemini app get Hydrogen only when Google's Auto mode picks it for a prompt.
Google says Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash, without saying how it measured that. Artificial Analysis measured about 170 output tokens a second for Hydrogen at its high thinking setting, against 112 for Gemini 3.8 Flash, according to one site that reports its results. On 16 November Google added Hydrogen as a model choice for AlphaEvolve in Gemini Enterprise, in its global, US and EU regions.
Alibaba's Qwen4-Max
Alibaba released Qwen4-Max on 16 November, according to its changelog. A preview version had appeared on 11 November. One outlet dates the launch to 15 November. Qwen4-Max is a mixture-of-experts model. It has about 2.4 trillion parameters in total, and about 71 billion of them are active for each token it produces. It reads images as well as text and takes up to one million tokens of context.
The international list price is $1.50 per million input tokens and $4.50 per million output tokens. Cached input costs $0.075 per million. Alibaba says the average cost per task fell by about 62 per cent on Work360, its own benchmark. These are Alibaba's own figures.
Alibaba is keeping the weights of Qwen4-Max private. Alibaba says smaller Qwen4 models will be released as open weights, and has not given a date. The model is available through Alibaba's API and in its own tools, including Qwen Studio and the Qoder coding tool.
Microsoft Ignite, day one
Microsoft opened Ignite in San Francisco on 17 November. Word, Excel and PowerPoint now run inside the Copilot app, and that feature is generally available. Files edited there stay editable in the ordinary Office apps, Startup Fortune reported.
Autopilot, a Copilot agent that runs long and recurring tasks, is starting to roll out to customers in Microsoft's Frontier early-access programme. The Microsoft 365 Copilot licence does not cover it. Autopilot draws on Copilot credits under a spending limit set by an administrator, and it stops when that limit is reached. Microsoft has not published final prices. Copilot Workspaces, a shared space for staff and their agents, is also in the Frontier programme, with no price or general availability date.
Copilot Managed Runtime, which hosts company agents inside Microsoft 365, is generally available. Microsoft's book of news says Copilot Code will be generally available in December. Microsoft's Copilot blog says 'in the coming weeks'. Dynamics 365 plugins for Sales and for ERP apps are generally available in Copilot. Agent Accounts, which give agents their own identities, enter public preview in December, Microsoft's Dynamics 365 blog says. Entra, Microsoft's identity service, can now find agents such as Claude Code and GitHub Copilot CLI running in an organisation.
ChatGPT Wallet
OpenAI launched ChatGPT Wallet on 13 November, CNBC reported. It lets OpenAI's agents, which it calls dots, buy things for a user. Neither the agent nor the shop sees the real card number. The wallet hands over a limited credential instead, which can include a substitute card number, security code and expiry date.
OpenAI built the wallet with American Express, Mastercard and Visa. It starts with ChatGPT Pro subscribers in the US. The user approves each purchase with a fingerprint or face check before it goes through. OpenAI says card benefits and dispute protections stay in place, and it is not charging transaction fees for now.
OpenAI also warns that some websites block agent traffic or flag agent purchases as fraud. Approving a purchase in ChatGPT does not guarantee that the shop confirms the order.
Claude Meetings
Anthropic announced Claude Meetings on 10 November, TechCrunch reported, and began rolling it out in beta on 12 November. It is for Claude Team and Enterprise plans, and it works with Google Meet only. Anthropic says other conferencing services will follow, without naming them.
Claude joins the call as a visible participant and posts a notice in the meeting chat. It does not join calls with external guests unless an administrator allows it. It leaves on its own if an external guest joins. Anthropic says Claude keeps a transcript, a summary and descriptions of shared screens, but no audio. Anthropic released Claude Whiteboards and Claude Sheets in beta alongside it.
Comparing the prices
Three companies set new list prices this week. GPT-6.2 Astra and Claude Fable 5.2, released on 9 November, now have the same standard price: $8 per million input tokens and $40 per million output tokens. Gemini 4 Hydrogen and Qwen4-Max both charge $1.50 per million input tokens, but Qwen4-Max charges $4.50 for output against Hydrogen's $7.50.
Most of the benchmark figures this week come from the companies that built the models. The Artificial Analysis scores are among the few independent ones. In SAASiQ's view, teams that run agents at volume should test any switch against their own workloads, because the cost of a task depends on how many tokens a model uses as well as on its list price.
Microsoft Ignite runs until 20 November. Microsoft says Copilot Code and the preview of Agent Accounts are due in December.
SAASiQ - Intelligent Solutions for SaaS ©

