AI News for Developers: Sept 29

Published October 1, 2026 — covering AI developer news from September 24 to October 1, 2026. This is a recurring weekly roundup; claims below are attributed to the companies and outlets cited so you can verify them yourself.

If you build with AI tools for a living, the week of September 29, 2026 was the busiest one in months. OpenAI ran its biggest DevDay ever, Anthropic shipped a model upgrade that undercuts its own flagship, Google DeepMind launched a frontier model to a select few, Meta reorganised its entire AI business around enterprise developers, and xAI put its latest model on AWS Bedrock. This AI news for developers roundup covers what actually shipped, what it costs, and what you should do about it.

Table of Contents

OpenAI’s DevDay 2026: 20+ announcements in one day

OpenAI held DevDay 2026 in San Francisco on September 29 and unveiled more than 20 products and features across ChatGPT, Codex, models, and APIs. For developers, five announcements stand out.

GPT-6.1 Sol: near-flagship coding intelligence at one-fifth the price

The new GPT-6.1 Sol model is OpenAI’s answer to the question every engineering team asks: how cheap can good-enough coding intelligence get? OpenAI says GPT-6.1 Sol nearly matches GPT-6 Astra on agentic coding, computer use, and professional work at one-fifth of Astra’s standard token prices. The API pricing is $2 per million input tokens and $10 per million output tokens, with cached input at $0.10 per million — which OpenAI describes as 95% below standard input pricing. The API model id is gpt-6.1-sol. It launched on September 29 in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, and in the API. If your agent pipelines run on GPT-6 Astra today, Sol is worth a benchmarked trial on your own workload.

Agents API goes public beta — with computer use

The Agents API is now in public beta with hosted execution, memory, tools, and multi-agent support. Crucially, it now supports computer use: agents can operate software through its UI, with OpenAI running the Codex harness. This is the managed alternative to running your own agent orchestration — if you have been hand-rolling multi-agent loops with retries and session state, the beta is worth evaluating against your in-house harness.

Decisions API, Codex upgrades, and plugin extensions

Three more developer-relevant launches. The Decisions API (limited preview) uses the GPT-6 Luna model to classify inputs or pick an action from a finite list of answers you define — a neat fit for routing and classification without burning a frontier model. Codex gained cloud execution (tasks keep running with your laptop closed), a refreshed CLI with two-way voice and an /agents view, Code Review with automatic cloud reviews, and Codex Security Cloud, which scans repositories for vulnerabilities and prepares fixes. And ChatGPT opened its sidebar to developer-built plugin extensions: dedicated sidebar entries, interactive panels, file viewers, support for the proposed MCP Events spec, and a redesigned submission flow. Amazon Bedrock Managed Agents, which run OpenAI-powered agents inside AWS, was also announced.

Also announced: GPT-6 Astra Ultrafast, Pro 500, and ChatGPT Space

Ultrafast is a paid speed tier reporting up to 8× faster token generation in Codex (up to 300 tokens per second) and 6× in the API; Astra gets it now, Sol follows. OpenAI also announced Pro 500 in the US at $500 per month, with Ultrafast and 25× the ChatGPT Plus usage allowance. On the product side, ChatGPT Space is a shared team workspace with a Meetings feature that takes notes and saves personalised summaries and follow-ups. Sign in with ChatGPT lets users bring their plan to 16 launch partners.

Anthropic: Sonnet 5.5 ships, Claude Code cloud sessions go GA

Claude Sonnet 5.5: half the price of the flagship, nearly matching it

Anthropic released Claude Sonnet 5.5 on September 28, 2026 — the second model in the Claude 5.5 family after Opus 5.5 shipped the week before. API pricing holds at $2 per million input tokens and $10 per million output tokens (with $0.20 cache reads and $2.50 cache writes), but Anthropic says output generation is more than 30% faster than Sonnet 5 and most tasks cost up to 30% less because the model uses fewer tokens. On Anthropic’s internal Terminal-Bench 4.0 coding evaluation, Sonnet 5.5 scored 70.6%, ahead of Opus 5.5’s 66.4% on the same settings. It is also the first Sonnet model to ship with the cyber safeguards and fallback protections previously reserved for Anthropic’s most capable models. GitHub added the model to Copilot the same day, generally available to Pro, Pro+, Max, Business, and Enterprise users. Anthropic positions Sonnet 5.5 for everyday, well-scoped work — bug fixes, documents, UI refinement — and Opus 5.5 for complex work needing sustained judgment.

Claude Code cloud sessions leave research preview

Anthropic made Claude Code cloud sessions generally available on September 23. A cloud session runs Claude Code remotely on Anthropic-hosted Ubuntu 24.04 VMs (roughly 4 vCPUs, 16 GB RAM, 30 GB disk), so your computer can be shut down while the task continues; you can start sessions from a browser, the Claude apps, or the terminal with claude --cloud. Sessions share your plan’s usage limits and there is no separate VM charge. Pro and Max subscribers also got one-time promotional cloud-session credits — $100 for Pro, $250 for Max — applied before regular plan usage. If you have been waiting to try agent-driven coding that outlives your laptop session, this is your cue.

Claude for Government goes GA

On September 30, Anthropic made Claude for Government generally available to US federal and state agencies, moving out of the public beta that began in July 2026. It delivers Claude’s coding and agentic capabilities through a FedRAMP High authorised environment, with admin spending controls and audit logs for agency authorisation-to-operate processes. Relevant if you sell software into the public sector.

Google DeepMind: Gemini 4 Argon launches to a select few

Google DeepMind launched Gemini 4 Argon on September 30, 2026, initially to trusted cybersecurity testers and US government partners through its Fairwind Program. The headline spec is a 1 million token output limit for long, multi-step reasoning, aimed at complex workflows in coding, enterprise knowledge work, and cyber defence. Sundar Pichai confirmed the release and said Google intends to expand availability to developers and consumers “as soon as we can and as safely as we can.” For now, the practical takeaway is that the next Gemini generation is real and being dogfooded by security teams first — keep an eye on the API waitlist if you build long-horizon agents.

Meta: enterprise platform, new boss, and Muse for small business

Meta made the biggest organisational AI move of the week. On September 28, Mark Zuckerberg announced the Meta Enterprise Platform, which he called “the next major pillar” of Meta’s business, bringing the Muse agent, Meta Business Agent, Muse API, and Muse Code to businesses and developers. He hired former MongoDB CEO Chirantan “CJ” Desai to run it as chief enterprise platform officer, reporting directly to him. On September 29, Meta launched Muse for Small Business: business skills and 15 third-party connectors inside the Muse app — Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe, and Zoom — plus direct access to Facebook and Instagram business accounts, available in the US and Canada. Developers take note: the Muse API and Muse Code are now explicitly part of Meta’s enterprise story, which means new distribution surface for tools you build on them.

xAI: Grok 4.7 hits AWS Bedrock; Grok Bot connects to your bank

xAI’s Grok 4.7, first released on September 21, 2026, arrived on Amazon Bedrock on September 28 with a 500K token context window, text and image input, and four configurable reasoning effort levels. xAI priced it starting at $2 per million input tokens and $6 per million output tokens, with a fast variant at twice the output speed for twice the price. Separately, on September 26 xAI announced a Finance integration for Grok Bot: read-only connections to bank, card, and investment accounts through Plaid, which then lets the agent analyse spending and subscriptions. And Musk announced on September 13 that Grok 4.8 — a 2.5 trillion parameter model on an in-house C++ stack — is completing its main training run before reinforcement learning. One pricing quirk worth knowing: Grok 4.7 charges double ($4/$12) for prompts over 200K tokens, so long-context agent work gets expensive at the far end of the window.

What this means for you this week

Price pressure is real and it favours builders. GPT-6.1 Sol and Sonnet 5.5 both offer near-flagship coding performance at mid-tier prices. If you have not re-benchmarked your agent stack’s default model this month, do it now — the cheapest good model changed this week.
Managed agent infrastructure is the battleground. The Agents API public beta, Claude Code cloud sessions, and Bedrock Managed Agents all push the same idea: let the provider run the harness. If you maintain your own orchestration layer, start measuring it against managed alternatives.
Security tooling shipped alongside the models. Codex Security Cloud, Sonnet 5.5’s cyber safeguards, and Gemini 4 Argon’s security-tester rollout all point to vendors expecting agents to do security-sensitive work. Turn on the security scanning features and verify Plugin4Shell patches across Codex, Claude Code, Gemini CLI, and Copilot.
New distribution channels are opening. ChatGPT plugin extensions, Meta’s enterprise platform, and xAI’s API/Bedrock presence all widen where agent-powered products can reach users. If you have been waiting for a channel to ship an agent product on, pick one and build.

AI News for Developers: What Shipped the Week of September 29, 2026