AI Model Roundup: Sep 14-20, 2026
No major lab shipped a new frontier text model this week. The news was in the layers around models: Google's new live-voice models, DeepSeek reversing a plan to retire V4 Pro, Anthropic loosening biology safeguards for verified scientists, and Mistral landing inside Firefox. A startup also claimed a model class that "can't hallucinate", which deserves some skepticism.
Google: Gemini 3.8 Live
On September 15, Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, its new real-time dialogue models. The standard model is built for scale and cost; Extended Thinking adds multi-step reasoning for harder conversations. Both are in the Gemini API and AI Studio for developers, in private preview for Gemini Enterprise, and are rolling into Search Live, the Gemini app, and Workspace. Google hasn't published prices in the announcement; check the API pricing page before planning volume.
DeepSeek keeps V4 Pro alive
When DeepSeek launched V4.1 Flash on September 10, it said all deepseek-v4-pro API requests would route to V4.1 Flash from September 14. That plan is off. DeepSeek's changelog now says it will "continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged."
The reversal matters for anyone who pinned V4 Pro for predictable behavior. Silent re-routing to a different architecture would have changed outputs without changing the model name. Note that deepseek-v4-flash and the vision-exp name are still temporarily routed to V4.1 Flash. V4.1 Flash costs $0.15 per million input tokens (cache miss) and $0.60 per million output off-peak, with peak hours at double (pricing).
Anthropic: Life Sciences Verification Program
On September 17, Anthropic opened a Life Sciences Verification Program that gives verified researchers access to Mythos, Opus, and Sonnet models with more permissive biology safeguards. There are two tiers:
- Standard Use: team-wide access for most life-science work, renewed annually.
- High-risk Use: a project-specific add-on for areas that are otherwise blocked, renewed every six months.
Monitoring moves from real-time blocking to offline pattern review, with 30-day retention for flagged activity. This is the same direction as Anthropic's enterprise cyber programs: capability gated by customer verification rather than by model-level refusals.
Mistral powers Firefox Smart Window
On September 16, Mistral and Mozilla announced that Mistral models will power Firefox Smart Window, Mozilla's browsing assistant. It launches in France and North America now, with the UK and Germany to follow later this year. It is a distribution win for a European lab and gives Firefox users an assistant not built on a US model.
Other launches
- TypeSafe System One and Jev (Sep 15). TypeSafe introduced a model class that returns typed, probabilistic decisions instead of free text, with 70 to 500 ms responses and input priced at $0.042 per million tokens. The "can't hallucinate" claim means outputs can't break the declared type, not that the decisions are correct. Early access only, no independent evaluation.
- Grok Build Memory (Sep 16). xAI's coding agent now carries project conventions and decisions across sessions, applied to new sessions.
Takeaway
The quiet week had one lesson: model names are not contracts. DeepSeek nearly swapped a model under a stable API name, and still does for Flash. Pin model versions where providers allow it, log which model actually served each request, and re-run your evals when a provider announces routing changes.
Missed last week? Read the Sep 7-13 roundup or the Qwen3.8-Max-0902 WebDev analysis.
Related: Open-weight models · Cheapest models · Agentic models