Gemini 3.7 Flash and 3.5 Flash IDs Now Route to Newer Models
This week's two Gemini stories share one warning: a Gemini id you pinned is no longer a promise about the model that answers. Calls to gemini-3.7-flash and gemini-3.5-flash already reach 3.8 Flash and 3.6 Flash, and Google now limits Gemini 2.5 Pro to past users, so we moved it to conditional.
113 announcements scanned · 25 mattered · 1 verdict changed
Verdicts changed
Gemini 2.5 Pro
Gemini 2.5 Pro
Google's Gemini API deprecations page limits 2.5-model access to users who actively used them in the past and tells new projects to use 3.5 Flash-Lite or 3.8 Flash. The model is not deprecated and has no shutdown date, so it stays usable for existing users only. Announced October 23 in our post on Gemini 2.5 Pro access limits.
Act before next issue
Re-test anything pinned to gemini-3.7-flash or gemini-3.5-flash
Google's Gemini API changelog for 8 October deprecates both ids and routes requests to them to gemini-3.8-flash and gemini-3.6-flash. The old ids still answer, but a different model replies, and they can no longer be selected in Google AI Studio. Google also set 23 October as the shutdown date for the deep-research-pro-preview-12-2025 agent.
Details & stepsStart new Gemini work on 3.8 Flash or 3.5 Flash-Lite, not 2.5 Pro
Google now limits Gemini 2.5 model access to users who actively used them before and tells new projects to use 3.5 Flash-Lite or 3.8 Flash. Gemini 2.5 Pro is not deprecated and has no shutdown date, so existing pipelines can keep it. Our rating moves from recommended to conditional.
Details & stepsOn our radar
Microsoft-Decision-1 at $0.042 per 1M input tokens
It returns a calibrated probability for each fixed answer, with output free, on Foundry and OpenRouter. Every benchmark result is Microsoft-reported, and we have not tested it.
Ollama Team is $500 a month flat
Ollama's pricing page lists Team at $500 a month (early access) with unlimited users and $1,000 of shared monthly usage credits, replacing the $25 per seat pricing we reported in August. Pro and Max state their allowance as $60 and $300 of credits a month.
NVIDIA OpenShell pairs with Claude Managed Agents
Anthropic said on 28 September that Claude Managed Agents can use NVIDIA OpenShell as the sandbox where their tool calls run. OpenShell is an Apache-2.0 runtime that blocks everything unless a rule allows it and keeps real credentials away from the agent. It is a 0.1.x release with anonymous telemetry on by default, so pilot it before you rely on it.
OpenClaw Enterprise, a self-hosted agent control plane
The OpenClaw Foundation released it on 29 September as a free, open-source, vendor-neutral platform for persistent agents, with multi-tenancy and hard security boundaries. Its maintainers rate it for internal pilot workloads only until 1.0.
Claude Marketplace is live
Anthropic opened it on 23 September with more than 2,000 connectors and plugins, and customers can spend part of their committed Anthropic spend on partner software. Plugin submissions opened on 25 September to developers on paid Claude plans.
In review, not yet published
None of these drafts has passed our source review, so we report no figures from them yet: the LiteLLM 1.107.0 release compromise, the OpenCode MCP sign-in flaw on Windows and WSL, GitHub Copilot's model deprecations, OpenAI's legacy snapshot shutdowns and its newly listed GPT-5.x and TTS removal dates, Gemini's preview audio model shutdowns, Claude Code 2.1.304 and 2.1.305 to 2.1.309, the OpenAI Decisions API going GA, OpenAI's misalignment report on GPT-6.1 Sol and GPT-6 Astra, Anthropic's report on unintended model actions, MiniMax M3.1 Flash, Mercury 3, Nemotron 3.5 Super VL, Voxtral 3, Cloudflare Clairic, the DHS order on DeepSeek products, Claude Duet, Claude on Google Cloud, Claude Security, Meta Muse share links, Eleven Realtime, DeepInfer, the OpenAI CLI, Agent Skills moving to the Agentic AI Foundation and 1Password for Agents. Each is written up when it publishes.
Agent attacks with no tool named
Norforge's findings on an AI agent campaign against company websites (reports disagree on more than 100 or 1,200 sites), Storm-0940's use of agents in the ConnectWise Cloud Backup compromise, and npm packages said to hijack AI agent traffic. None names a model, harness or tool a reader runs. Held until one does.
Launches read only as a title or feed summary
OpenAI models and Codex workflows on Google Cloud's Gemini Model Garden (OpenAI's page returned no readable body), Canva Craft, Claude working with Google Docs, Sheets and Slides, and Google Research's Harness2. Held until a primary page shows models, prices or a release.
Deprecations and wind-downs still undated
GitHub's entry on Copilot model deprecations in mid-November was seen as a title only. OpenAI's Help Center still calls the Cursor wind-down proposed, with a 12 November transition period and no termination date. Mistral Large 4 weights are still unpublished.
Policy and research held at headline level
Canada's biggest newspapers suing Anthropic, a Chinese export-control proposal for very large models, Anthropic's AI safety bill pitch, OpenAI's distribution-partner standards, Microsoft's guidance on agents as insider threats and Hirundo's Qwen3-Coder persona research. Nothing read changes a product yet.
Open verdict watches
A move of GPT-4.1 from recommended to caution is staged with OpenAI's deprecation draft and lands only when that post publishes. LiteLLM, OpenCode and the OpenAI Decisions API stay conditional after this week's stories, and GPT-4o stays conditional.
Filtered out
- Recursive's Flint coding model as a news story. At triage only TechCrunch reported it and no Recursive or Hugging Face page could be found, so the story was rejected. A later directory pass read Recursive's own model card and blog and added the model.
- A third post on DeepSeek's weekday-only peak window. It restated three published posts with no new fact, so review kept it in draft on October 23.
- A post on an Eleven v4 Turbo discount. The cited source did not show the promotion and the story duplicated another draft, so review kept it in draft on October 20.
- Re-reports of stories already in our queue: TechCrunch and The Hacker News follow-ups on the DHS DeepSeek order and the OpenCode Console breach, a restatement of OpenAI's October 23 shutdown table, and NVIDIA OpenShell on Hacker News.
- Corporate, governance and customer posts from OpenAI, Anthropic and the Claude Blog, such as Anthropic's public benefit fund, a new LTBT member, the HCLTech partnership and Uber's ChatGPT Work rollout. No product change.
- Funding, deals and consumer news from TechCrunch and The Verge, including OpenAI's raise, reported DeepSeek and Nscale talks, Otter and Tenstorrent rounds and Muse user counts. None moves a tool, model or price.
- Non-AI security items from The Hacker News, including eMudhra CA keys, PHP2Shell, a GitLab CVSS 10 flaw, FortiSIEM, Qilin and Rank Math PRO. Not AI tools, or not a reader setup.
- Index and changelog pages with nothing new for readers: Google's model cards, DeepSeek's change log, Meta, xAI and Mistral news, Anomaly notes and Codex alpha pre-releases with no notes.
- Research, opinion and off-beat Hacker News, Verge and TechCrunch items, demos and launches with no reader-setup change.
Our take: Our verdict on this week: re-test anything pinned to an older Gemini id, and start nothing new on Gemini 2.5 Pro. Google's Gemini API changelog for 8 October deprecates gemini-3.7-flash and gemini-3.5-flash and routes requests to them to gemini-3.8-flash and gemini-3.6-flash, so the old ids still answer but a different model replies. Google's deprecation table also set 23 October as the shutdown date for the deep-research-pro-preview-12-2025 agent. Google now limits Gemini 2.5 model access to users who actively used them before and points new projects to 3.5 Flash-Lite or 3.8 Flash, so Gemini 2.5 Pro moves from recommended to conditional: fine if you already run it, wrong for anything new. Microsoft-Decision-1 lists at $0.042 per 1M input tokens with output free, on Foundry and OpenRouter, and every benchmark result is Microsoft-reported. Ollama Team is now $500 a month flat (early access) for unlimited users, with $1,000 of shared monthly usage credits. The week's bigger stories, including the LiteLLM 1.107.0 release compromise, the OpenCode MCP sign-in flaw on Windows and OpenAI's newly listed deprecation dates, are still drafts in review, so their figures are not in this issue. How we count: two daily scans ran on October 20 and 23 and fetched 113 items (46 and 67). Grouped into stories, that is 7 act, 0 verdict claims, 19 fyi and 17 watch stories, plus 55 noise rows. 25 stories became news drafts and 1 was rejected (a single-source report). Seven news posts were published (4 on October 20 and 3 on October 23) and 1 verdict changed. Two verification runs on October 19 checked 41 directory entries (30 and 11) and drafted 1 news post, and a fix run on October 21 checked 3 more entries and drafted 1 update. The directory grew by 6 models and 6 comparisons.
Neomanex Utility editorial
Get issue 30
Next week's verdict changes and deadlines, in your inbox.