Quick answer

September 2026 ended with 25 model releases from 16 providers over the month, by our count of announced models. The final week alone: GPT-6 Luna and Sol and Xiaomi MiMo V2.6 Flash and Pro (Sept 22), ChatGPT Voice plugins (Sept 23), Gemini 3.8 Live Avatar, Mastercard Agent Pay, and Amazon seller agents (Sept 24), Perplexity Search Fast and Qwen audio price cuts (Sept 24–25), Anthropic's 950-agent swarm (Sept 26), OpenAI's research-agent image leak (Sept 27), and Claude Sonnet 5.5 (Sept 28). The themes: cheaper mid-tier models, voice that acts, and agents that spend.

Our early-September recap covered the flagship launches. This is the rest of the month, which turned out to matter at least as much for what people actually use and pay.

Models

  • Sept 22 — GPT-6 Luna and GPT-6 Sol (OpenAI): the broadly available and coding tiers of the GPT-6 family, beneath Astra
  • Sept 22 — MiMo V2.6 Flash and Pro (Xiaomi): the phone maker's latest open-weight models, continuing its push into the model market
  • Sept 24 — Perplexity Search Fast: a lower-latency search mode for quick lookups, alongside the Hybrid Compute on Mac launched earlier in the month
  • Sept 25 — Qwen TTS-Next and ASR-Next (Alibaba), launched with price cuts of up to 95% on audio
  • Sept 25 — LLM Gateway Smart Route: automatic routing across providers by cost and quality, one of several gateway products adding model selection
  • Sept 28 — Claude Sonnet 5.5 (Anthropic): unchanged $2/$10 pricing, 30% faster, with a vendor-claimed lead in agentic coding

Products and agents

  • Sept 23 — ChatGPT Voice plugins for email, calendar, and Slack, with backend choice across Astra, Sol, and Luna; ChatGPT mobile Work agents
  • Sept 24 — Gemini 3.8 Live Avatar in 97 languages; "Call for Me" testing on Pixel 11
  • Sept 24 — Mastercard Agent Pay and AgentCard; Amazon agents for sellers
  • Sept 24 — Dataiku Agent Management (GA in October); Alibaba AgentCore and Agent Context, which Alibaba says cuts agent token use by 67%
  • Sept 24 — Ando team chat raised $20 million; Strada launched; both in the agent-collaboration space
  • Sept 27 — Meta Muse expanded to Canada, its first market outside the US

Money and policy

  • Sept 23 — The Ban Artificial Superintelligence Act introduced by Sanders and Casar
  • Sept 24 — OpenEvidence raised $250 million at a $15 billion valuation for its clinical AI
  • Sept 25 — DeepSeek reported at a $1 billion run-rate and raised API prices 2.3 to 4.5 times
  • Sept 26 — Anthropic's roughly 950-agent Claude swarm on reverse transcriptase research
  • Sept 27 — OpenAI disclosed that research agents exposed 53 users' images

What it adds up to

Three things. The mid-tier is where the action is: Sonnet 5.5 and GPT-6 Sol and Luna are the models most products will run, and both vendors held or cut prices. Voice became an interface for action rather than a novelty, with ChatGPT plugins and Google's calling test. And agents gained the ability to pay, which turns "assistant" into "delegate" and makes the controls around them the product. The leak at OpenAI is the counterpoint: more capability, more surface area.

Every benchmark and efficiency figure in this roundup is vendor-reported. Independent evaluations of the September models were still pending at publication.

Related reading

Early September 2026 releases recap Claude Sonnet 5.5 released GPT-6 Luna and Sol released

Bottom line

If you build with AI, re-quote your model costs this week. If you use AI, your assistant can now act by voice and, soon, pay. If you follow policy, note that the first federal ban bill arrived in the same month as the most capable models. September was the month all three moved at once.