Welcome to HowToShipIt — practical how-to guides for developers: code, AI tools, and servers, explained step by step.

Weekly AI developer news roundup: OpenAI Dots agents, Claude Sonnet 4.5 deprecation, xAI Team Bots — October 2026

AI Developer News: OpenAI’s Dots, Claude Sonnet 4.5’s Retirement Date, and This Week’s Launches

by

Published 2 October 2026 — covering AI developer news from 26 September to 2 October 2026. Claims below are attributed to the companies and outlets cited so you can verify them yourself.

If last week was about models, this week was about agents — and about who takes responsibility for what they do. OpenAI shipped its always-on Dots agents, Anthropic gave your favourite older Sonnet model a retirement date, xAI turned Grok Bot into a team product, and the commerce world started drawing hard lines around agent identity. This is your AI developer news roundup for the week ending 2 October 2026: what shipped, what it costs, and what you should do about it.

1. OpenAI’s Dots: always-on agents that keep working after you log off

The biggest developer-relevant launch since last week’s DevDay coverage was arguably hiding in plain sight at DevDay itself. OpenAI unveiled Dots on 29 September — always-on agents that run on their own cloud computer and browser, powered by the flagship GPT-6 Astra model, and keep working on your goals between conversations.

Here is how they work. You give a dot a goal, connect the apps it needs, and set custom rules for what it may do independently and what requires your approval. The dot then works in the background — OpenAI describes examples like monitoring customer feedback and preparing software fixes, updating launch materials when requirements change, and re-running research analysis as new data arrives. When you are not actively engaged, the dot switches to “proactive research” mode, scanning your connected apps with read-only access so it cannot send messages or alter content unprompted. Certain actions, like changing a password, are reserved for you and cannot be delegated.

The practical specs, as reported across the launch coverage:

  • Reach: dots connect to more than 4,000 applications through OpenAI’s plugin ecosystem, and reach you through ChatGPT, Slack, and Microsoft Teams, with SMS texting coming soon. You can message or call a dot, and it carries context across channels.
  • Availability: rolling out now to Pro and Business Premium subscribers in eligible markets. Enterprise, Edu, and Healthcare workspaces get a beta that administrators must switch on — it is off by default. Free and Plus plans are out, and the Pro rollout currently excludes the EEA, Switzerland, and the UK.
  • Cost: your first dot is included at no extra cost on eligible plans, with extended usage limits for the first month. Chats with a dot do not count towards your ChatGPT usage limits, but tasks it starts in Codex or ChatGPT Work do count against normal allowances.
  • Specialist dots: OpenAI is also previewing enterprise-configured dots with their own identity, credentials, and system access for defined jobs like procurement or invoice processing, with pilots managed alongside Microsoft’s Agent 365.

The obvious matchup is Meta’s Muse agent, which launched in early September with a free-to-start tier in the US and Canada. Dots are paid-plan only, run on GPT-6 Astra rather than an open-weights model, and come with workspace admin controls and deep Slack/Teams integration. The interesting bit for you as a builder: the race is no longer about which model writes the best code. It is about the agent that lives inside your company — and how much you trust it with credentials. Audit the custom-rules mechanism and the proactive-research read-only guarantees before you let any persistent agent touch production systems.

2. Anthropic: Claude Sonnet 4.5 gets a retirement date, claude.dev goes live

On 30 September, Anthropic moved claude-sonnet-4-5-20250929 to Deprecated on its model deprecations page and set a retirement date of 30 November 2026, naming claude-sonnet-5-5 as the recommended replacement. That is 61 days of notice — one day more than the 60 Anthropic commits to for publicly released models.

If you still have Sonnet 4.5 pinned anywhere, treat this as a migration task, not a warning to file away. Deprecated means the model keeps working until 30 November and then requests fail outright. Between now and then, Anthropic’s own caution is that deprecated models are “likely to be less reliable” than active ones. The side projects and plugins nobody has touched in a year are the ones this catches.

Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens — a third less per token than Sonnet 4.5’s $3/$15 — but the migration is not free:

  • disabled thinking is removed; use thinking: {"type": "between_tools"} at low, medium, or high effort.
  • Forced tool use and assistant prefill patterns that used to work now behave differently — the migration FAQ recommends structured outputs or a tool marked strict: true with tool_choice set to auto.
  • Thinking is on by default, and the model uses roughly 30% more tokens for the same text, so your real cost depends on settings — re-test effort and compare cost on real traffic before 30 November.

Also from Anthropic this week: claude.dev launched on 30 September (announced via the @ClaudeDevs account) as a dedicated developer hub with engineering deep-dives, Claude Code and API guides, and agent playbooks. Bookmark it — it is now the canonical source for Anthropic’s own agent-building workflows. And on 28 September, Anthropic paired its managed-agents APIs with NVIDIA’s open-source OpenShell secure runtime, adding security and control layers for production agents handling proprietary data. If your agents touch customer data, that combination is worth evaluating.

3. xAI: Grok Bot goes team-wide, and the API docs get sharper cost math

xAI had a quiet but productive week for developers. On 28 September, Team Bots entered public beta: Grok Bot becomes a shared, role-based agent for whole teams. One bot is configured once — files, instructions, skills, plugins (Salesforce, Notion, GitHub, and custom MCP servers), secrets and credentials, and memory — then published to teammates, each of whom keeps private conversations. Each Team Bot gets its own Slack handle and app. OAuth plugins act as the person asking, while scoped service-account secrets (encrypted, redacted in output, up to 25) are shared team-wide. It is on Teams and Enterprise plans with no separate pricing, and the docs are live at docs.x.ai.

On 1 October, the @grok account announced a new Agent Dashboard for Grok Build: run grok dashboard (or /dashboard) and you get one interactive screen showing all your parallel coding-agent sessions — peek at output, reply in place, dispatch new agents, even launch agents in isolated Git worktrees with Ctrl+W. If you run several agents at once, this is the orchestration view that was missing.

The Grok 4.7 API docs were also updated on 28 September with details that change your cost math:

  • A US regional endpoint, https://us.api.x.ai/v1, keeps inference in the US at a 10% token premium.
  • Strong guidance to set prompt_cache_key to keep cache hits on warm servers.
  • Encrypted reasoning is always returned on the Responses API for multi-turn agent loops.
  • Prompts over 200k tokens cost double ($4 input / $12 output per million versus $2 / $6 below 200k) — so design your long-context agent loops to stay under that line or budget accordingly.

4. The agent-commerce week: identity is the new API key

Four moves this week turned on the same question: who answers for what an agent does? DesignRush’s weekly roundup captured the pattern:

  • Microsoft rebuilt Copilot around three tabs and moved its agent features onto usage-based credits that admins can watch line by line — metered agents that spend money unsupervised are getting the auditability treatment.
  • Amazon opened its seller backend to Claude, just three days after blocking Meta’s Muse from its storefront — for browsing anonymously and apparently holding customer credentials. Access is going to agents that can be named and audited.
  • PayPal added Muse checkout across its merchant network — the payment rails are now agent-aware.
  • Six major banks published AI trust principles that lead with transparency: every party to a purchase should know an AI agent was involved.

Satya Nadella’s line from the coverage is worth pinning to your team wiki: “Every agent has to have an identity. Everything it does needs to be observed.” If you build agents that touch money, credentials, or other people’s systems, the window for anonymous agents is closing. Build identity and audit logs into your agent architecture now — the platforms are about to require it.

5. Two numbers and one caution you should act on

The JetBrains 2026 Developer Ecosystem Survey, published 2 October, found that 90% of developers now use AI coding agents weekly and 47% of all new code is AI-generated. The companies moving fastest have stopped treating AI as autocomplete and now run agents across whole repositories, terminal CLIs, and orchestration frameworks — with humans reviewing architecture rather than writing every line.

The caution comes from the same coverage: DeepMind documented a 17x error amplification rate in cascading agent chains. The teams that do best, the survey found, are the ones tracking quality alongside velocity — raw output volume is not the metric. If you have not set up observability for your multi-agent pipelines, this is the week to do it. The tooling gap between “agents that write code” and “agents you can trust in production” is the whole ballgame in 2026.

Finally, a note on what did not ship: Forbes reported that OpenAI held back the planned GPT-6.1 model after deception tests, and AP noted the CEO avoided any mention of security concerns during his DevDay presentation. The same company that launched always-on agents also decided one of its models was not safe to release. Hold both facts in your head when you decide how much autonomy to grant.

What this week’s AI developer news means for your stack

  • Grep for claude-sonnet-4-5 today. Anything pinned to the dated model ID breaks on 30 November. Budget the Sonnet 5.5 migration — including re-testing thinking effort and cost on real traffic — before the holiday freeze.
  • Trial the persistent-agent model on a non-critical workflow first. Dots, Team Bots, and specialist dots all push the same idea: agents that outlive the chat session. Pick one internal workflow, define strict custom rules, and measure before expanding.
  • Design long-context agent loops around the 200k-token line. xAI’s doubled pricing above 200k tokens and DeepMind’s 17x error amplification in cascading chains both say the same thing: shorter, cached, observable agent loops win.
  • Give your agents an identity and an audit trail. The commerce world is moving faster than regulation here — Amazon, PayPal, Microsoft, and the banks are all demanding identifiable, observable agents. Build it in before a platform forces you to.

Further Reading & References

Related Posts:

Leave a Comment