AI News Roundup — Week of September 21, 2026

AI News Roundup — Week of September 21, 2026

A quieter week at the frontier, but a busy one for pricing pages and product packaging. Below is what changed in the last seven days, with the caveats we think matter before anyone reworks a stack around it.

New models

• Alibaba open-sourced Qwen-Image-2.1 on September 20. A 7B unified image generation and editing model with a real alpha channel for transparent output and support for up to ten reference images in one pass. Weights went to Hugging Face, ModelScope and GitHub with same-day support in ComfyUI, Diffusers and vLLM-Omni. Worth reading the fine print: the Qwen Research License permits research and evaluation only, so "open weights" here does not mean commercial use without a separate licence.
• DeepSeek's V4.1 Flash is now the default target for the older Flash model names. The natively multimodal release retired V4 Flash and V4 Flash Vision Exp, which now route to V4.1 Flash for compatibility. See DeepSeek.
• No new flagship from the three largest labs this week. The current generation still dates to early September: GPT-6 Astra on ChatGPT (September 3), Fable 5.1 on Claude (September 1) and 3.8 Flash on Gemini (September 2).

Pricing & product updates

• DeepSeek reversed its own deprecation. V4 Pro requests were scheduled to route to V4.1 Flash at Flash rates from 04:00 UTC on September 14. Hours before the cutover, DeepSeek said it would keep serving V4 Pro with billing unchanged, citing user demand. A useful reminder that a posted retirement date is not always a retirement date.
• Perplexity moved Comet's "Control browser" behind Computer credits, reported September 15. The feature had been part of the $20/month Pro plan; browser control is Comet's headline capability, so this is effectively a price increase for existing subscribers rather than a new add-on.
• OpenAI cut Voice costs and shipped new API surface. ChatGPT Voice now consumes 1.25 credits per minute rather than 5 in Enterprise and Edu workspaces, with additional Business Voice usage at the same rate. GPT-Live 1 reached general availability in the API at $0.05 per minute, billed per second, with backend model and tool usage charged separately. The Agents API entered public beta.
• Google unbundled Code Assist from new Gemini Enterprise subscriptions. From September 17, new or renewed Gemini Enterprise Standard and Plus subscriptions purchased online no longer include Gemini Code Assist; existing subscriptions keep access through their current term. Separately, Gemini Enterprise pay-as-you-go was extended to all projects tied to invoiced Cloud Billing accounts.
• Anthropic shipped a run of Claude Code releases (2.1.271 through 2.1.275) aimed at cost visibility: managed pricing multipliers of up to 10 for internal chargeback, agent, skill, plugin and MCP server names attached to token and cost metrics, smaller default workflow sizes on Pro plans, and prompt-cache fixes.
• Mozilla added Mistral Small 4 to the Firefox Smart Window beta on September 18, with zero data retention agreed on the Mistral side. Coverage is France and North America, with the UK and Germany expected later this year.
• xAI added cross-session memory to Grok Build on September 16, so project context and past decisions carry between sessions. Relevant to Cursor users, since Cursor is now an xAI subsidiary and co-develops the Grok models from 4.5 onward.

Money

• Temporal Technologies raised $550M at a $12.55B valuation, the largest US round of the September 12–18 window. Temporal's open-source platform for long-running agents puts it squarely in the agent-infrastructure category.
• Factory raised $200M at $5B for enterprise AI software development, and Arcee AI raised $150M at over $1B to keep building open-weight models.
• Profound raised $180M at $1.8B. The pitch — helping brands appear in AI-generated answers — is a category that barely existed eighteen months ago and is now funded at scale.
• Cohere and Aleph Alpha signed a definitive merger agreement (September 16). The combined company keeps the Cohere name with headquarters in Berlin and Toronto and Heidelberg retained as a research site. Terms were not disclosed; the combined entity is reported at roughly $20B, and the deal remains subject to regulatory approval.

Tool shake-ups worth knowing

• Reve ends image generation on September 27 at 12:00 Pacific, following an investment from OpenAI that also saw members of its research team join OpenAI to work on multimodal research. Reve remains an independent company and the investment size was not disclosed. Existing users can sign in until October 31 to download their images, and subscribers who paid September fees are due full refunds by mid-October.
• OpenAI's retirement calendar is dense this month. The Videos API and sora-2 go on September 24, the last legacy completion snapshots on September 28, and gpt-3.5-turbo and gpt-4 on October 23. GPT-5.5 leaves ChatGPT, ChatGPT Work and Codex on October 14; the API is unaffected, and Codex users are directed to GPT-5.6 Sol.
• AWS Mechanical Turk closes on September 30, ending a service that predates the current wave by close to two decades and still sits in many data pipelines.
• Gemini Code Assist's removal from new Enterprise bundles belongs here as much as in the pricing section — teams renewing online may lose tooling they assumed was included.

Figures, dates and valuations here come from third-party reporting and vendor release notes and move quickly; check the vendor's own pricing and status pages before acting on anything above.