Auto-Routed Models, Server-Side Threads, Phone-Ready Voice: Your Defaults Are Moving

Four defaults moved between November 2 and 6

Four announcements from four unrelated organizations landed in the first week of November 2026. Only one of them changes your production behavior without a deploy: from December 1, Google routes two Gemini 2.5 model strings to newer models with a different API contract. The other three are opt-in, but each moves a default that teams rarely revisit. OpenAI made outbound phone calls a model-side feature. Anthropic offered to keep conversation state on its servers. Nous Research showed how cheaply the values of an open-weight model can be rewritten.

Google: on December 1, two Gemini 2.5 strings point at different models

On November 2, Google deprecated gemini-2.5-pro and gemini-2.5-flash, and removed both from the Google AI Studio model picker. From December 1, 2026, API requests to gemini-2.5-pro route automatically to gemini-3.8-flash, billed at Gemini 3.8 Flash rates. Requests to gemini-2.5-flash route to gemini-3.5-flash-lite, billed at Flash-Lite rates.

The routing is the part teams will miss, because the target has a different contract. The changelog lists the breaking changes: validation of function calling and thought_signature, thinking_level replacing thinking_budget, and the removal of sampling parameters and prefilled model turns. If your agent loop sets temperature or prefills a model turn, expect a rejected request or a changed output on December 1. The changelog does not say which.

Note also that the Pro string lands on a Flash-tier model. The invoice may go down. Whether quality on your hardest prompts holds is something only your own evaluation set can answer.

Google applied the same mechanism twice more that week. On November 5, Gemini Omni 1.2 Flash became generally available: video generation up to 15 seconds, conversational editing with per-turn extensions up to 30 seconds total, subject consistency across up to 6 reference images, video output 30% cheaper ($12.00 per 1M tokens) and doubled paid-tier rate limits. Requests to gemini-omni-1.1-flash route to 1.2 on December 3, at 1.2 rates. On November 6, gemini-embedding-2-preview started routing to gemini-embedding-2, which Google says shares the same schema, pricing and embedding space, so no re-indexing is needed. Also on November 5, function responses started accepting audio, video and Office files, so a tool can return a clip or a spreadsheet straight to the model.

What to do this week: grep every Gemini model string in your repositories, replace the deprecated ones explicitly instead of letting the alias decide, and run your eval set against gemini-3.8-flash and gemini-3.5-flash-lite before December 1. The migration guides are linked from each changelog entry.

OpenAI: the Live API now places calls, and image models emit SVG

OpenAI's API changelog has two entries that matter to builders. On November 2, outbound SIP calling became available to all organizations through the Live API: you connect gpt-live-1 to a phone number through your SIP provider. On November 4, OpenAI released GPT-Image-2.6 Sunburst and GPT-Image-2.6 Flare, with adjustable complexity, SVG generation and animated GIF generation, plus markup-guided editing for supported GPT Image 2.5 and 2.6 models.

For voice teams, the new piece is the direction of the call. OpenAI's realtime docs already carry a Telephony and SIP section; this entry is specifically about outbound calls. Outbound calling for every organization means appointment reminders, callbacks and follow-up calls can start from the model side, with your SIP provider as the only telephony component. You still own the trunk, the caller ID and the consent rules for automated calls in each country you dial.

We have not benchmarked it, so we make no claim that it is faster or cheaper than a dedicated voice-agent platform. It does add a column to the comparison in our ElevenLabs, Vapi and Retell breakdown. The numbers worth collecting are end-to-end latency on a real call, cost per minute at your volume, and behavior when the caller interrupts.

The image release matters more day to day than its changelog line suggests. SVG output is text: it diffs, versions and passes through a linter. If you generate icons, diagrams or chart skeletons, you can review a vector result in a pull request, which you cannot do with a raster file. Animated GIF output removes a conversion step for teams that chained an image model and a separate animation tool.

Anthropic: the server keeps the thread

Anthropic's Claude Platform release notes for November 2 describe message threads for the Messages API, in beta behind the message-threads-2026-08-12 header. The API keeps the conversation on the server, so each later request sends only the new messages. A thread expires 5 minutes after its last request and is deleted no more than 1 hour after its last use. Anthropic presents the benefit as smaller requests and lower time to first token on long conversations and agent loops. Threads are available on the Claude API and on the AWS, Google Cloud and Microsoft Foundry distributions listed in the notes.

Two consequences for production. First, the 5-minute expiry suggests the feature targets active agent loops, not conversations a user resumes the next day; it matches the default prompt-cache lifetime we covered in our prompt caching piece. An expired thread cannot be resumed, so your application still needs its own copy of the history. Second, a customer data-retention commitment now has one more server-side object to account for, even a short-lived one. Read your platform's retention terms before routing regulated conversations through threads.

A smaller entry from November 3: text generated by Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 now carries Anthropic's text watermark on the platforms listed in the notes. The watermark documentation states that it adds no tokens or hidden characters and needs no change to request or response handling. The exception to know: on Sonnet 4.6, a request that sets temperature, top_p or top_k to a non-default value might produce text without the mark. If a compliance process relies on detection, pin sampling settings to their defaults on that model.

Nous Research: the values of an open model are one fine-tune deep

On November 6, Nous Research published Arete-27B on Hugging Face under Apache 2.0. It is a supervised fine-tune of Alibaba's Qwen3.8-27B: about 52.8 million tokens, two epochs, 1,202 steps, with teacher answers from an open model (Inkling) on twenty thousand prompts across seven value categories. The model card states the goal plainly: to "reinforce western values" on a Chinese-trained model and cut refusals, deflections and state-framed answers on politically sensitive topics.

On Nous's own 803-prompt evaluation, graded by an LLM judge (Grok 4.7), the fact score rose from 2.32 to 2.98 out of 5, and answers containing propaganda fell from 22% to 0.9%, with no meaningful regression reported on the general-capability and agentic benchmarks they ran. These are self-reported results on a purpose-built test with a model judge, and the card says it does not claim to have removed all such behavior.

The practical point holds whichever side of the values question you sit on. The worldview an open-weight model ships with sits in a post-training layer, and a single SFT run on a 27B model is enough to shift it measurably. If you deploy an open model in a product that answers questions on law, politics or history, the provenance of the base model belongs in your risk assessment. Any derivative, this one included, needs its own evaluation on your prompts and in your languages before it replaces a base model.

What we're checking before December 1

• Model strings. Every gemini-2.5-pro, gemini-2.5-flash and gemini-omni-1.1-flash reference, with an eval run on the routing target before December 1 and December 3.
• Sampling and prefill. Any Gemini call that sets sampling parameters or prefills a model turn, since the routed targets drop both.
• Beta headers. Each one is a dependency with its own lifecycle; message-threads-2026-08-12 should sit behind a flag with a stateless fallback.
• Voice routing. If you run outbound calls through a third-party platform, list what a direct SIP connection would remove and what it would leave you to build.

Every model string and beta header in your code has an expiry date, and that date is usually published in someone else's changelog. If you want a second pair of eyes on yours before December 1, get in touch.