Eversynced AI Weekly - October 5, 2026

Your Monday briefing on AI that matters for developers.

Issue #14 - September 28 - October 4, 2026.

This Week’s Big Move

OpenAI held DevDay on September 29 and posted more than 20 announcements. The headline for developers is computer use in the Agents API, the hosted Codex harness OpenAI put in public beta on September 10. The API can now drive a browser, and per the recap it brings “Codex’s multi-agent capabilities, tool search, tool calling, and context compaction into your application.” OpenAI and Amazon also launched Bedrock Managed Agents, which you can use to build OpenAI agents that “run entirely in AWS.”

Computer use is available “through the API and in Codex and ChatGPT Work on Pro 500 and Enterprise.” The browser runs in an OpenAI-hosted environment, and the user has to approve each new website origin, with your app collecting the decision. The docs add a warning: “Origin approval does not enforce confirmation before individual actions.” So if a purchase or a destructive change needs a human yes, OpenAI’s advice is to restrict the hosted browser to resources that can’t perform those actions, or to run a browser you control.

MCP Events is smaller news, but it changes what an MCP server can do. ChatGPT Work chats now support MCP Events, so a plugin can subscribe to events like message.created on a chat channel and act when one arrives. It requires MCP 2.0 (protocol version 2026-07-28, the spec issue 8 covered when sessions went away), plus persistent subscription storage on your server and outbound HTTPS to a callback URL. OpenAI’s guide calls the underlying spec a draft, and the MCP Triggers and Events Working Group that owns it still lists its Events SEP as “Ideating.” Our read is that this is the first big client to support webhook-delivered MCP events, so a server that only answers tool calls is now missing a feature. Start with the guide’s events/list and events/subscribe examples.

Tools & Agent Updates

Claude Code v2.1.284 to v2.1.289 - Six releases from September 28 to October 3. v2.1.284 made Sonnet 5.5 the default Sonnet on the Anthropic API and started interactive sessions in Auto Mode, on every plan and provider, when no permission mode is configured. Issue 10 expected that provider-wide default near the first week of September, and it landed September 28. v2.1.285 capped background Bash at 30 minutes by default, and v2.1.288 limited that cap to unattended sessions. v2.1.289 fixed Bash deny and ask rules that missed commands like TZ="$HOME" rm -rf build when the sandbox auto-allows them.

GitHub Copilot computer use - In public preview since October 1 in Copilot CLI and the Copilot app, on macOS and Windows. Run /computer on, and Copilot asks before it takes control of an app. GitHub’s examples include legacy and GUI-only apps with no API, CLI or MCP integration. On October 2, Copilot code review gained REST and GraphQL support with an effort level per request, and the same post confirms Balanced became the default effort on September 28, as GitHub announced on August 28.

Pi 1.0 - The MIT-licensed coding agent CLI from earendil-works reached 1.0 on October 1. Codemode, the tool that lets the model write a sandboxed script calling Pi’s other tools, got much leaner: with the default tools and codemode on, a GPT-5.6 request drops from about 5,300 prompt tokens to about 3,300.

MCP TypeScript SDK 2.3.0 - The October 2 client release changes how redirects work. HTTP transports and OAuth helpers now follow a redirect only if it stays on the same origin and keeps the method (a 307 or 308, or any redirect of a GET). On a transport, anything else fails that request with an error naming the target, though the session survives. The OAuth helpers move on to the next well-known URL during metadata discovery, and other OAuth requests fail with an error giving the status.

One List Price, Three Labs

Anthropic, OpenAI and Google released or announced models between September 28 and 30, and all three put the headline price at $2 per million input tokens and $10 per million output (Google’s is introductory):

• Claude Sonnet 5.5 - Released September 28 at Sonnet 5’s price, with cache reads at $0.20. Anthropic says it costs “up to 30% less per task” because it uses fewer tokens, and its table puts Sonnet 5.5 at 70.6% on Terminal-Bench 4.0 against 10.3% for Sonnet 5.
• GPT-6.1 Sol - Released September 29 at GPT-6 Sol’s $2 and $10, one-fifth of GPT-6 Astra’s standard prices, with cached input at $0.10, half GPT-6 Sol’s cached rate. It’s in the API as gpt-6.1-sol and in ChatGPT Work and Codex, but “not yet available in Chat.”
• Gemini 4 Argon - Announced September 30 at an introductory $2 and $10, rising to $4 and $20 later. Google raised the output limit from 64K tokens to 1M. It’s rolling out first to trusted cyber defenders through the Fairwind Program, the gate issue 12 explained.

The headline price no longer separates them, so read one line down. Sol’s cached input costs half of Sonnet 5.5’s cache reads, and issue 13 argued that cache hit rate moves an agent’s bill more than the headline does. Past that, what differs is cost per finished task, and neither Anthropic nor OpenAI measures it on equal terms. Anthropic’s table compares Sonnet 5.5 with GPT-6 Sol, which OpenAI upgraded to GPT-6.1 Sol a day later, and on Terminal-Bench it reports GPT-5.6 Sol instead because OpenAI never published a GPT-6 Sol score. OpenAI’s GDP.pdf result is against “Opus 5.5 with fallbacks.” And under its Terminal-Bench 4.0 chart, Sonnet 5.5’s page says that at medium effort it far exceeds Sonnet 5’s best score for less than a tenth of the cost per task, at the same list price. Our read: the only cost-per-task number you can trust comes from your own workload.

What This Means for You

What Matters

• Check your permission mode after you update Claude Code. Since v2.1.284, interactive sessions with no configured mode start in Auto Mode, and v2.1.285 did the same for claude -p and Python Agent SDK sessions on third-party providers or with telemetry off. Set permissions.defaultMode, or pass --permission-mode for claude -p and the SDK, if you want something else.
• If you run Sonnet with thinking off, switch to between_tools before you move to Sonnet 5.5. On Sonnet 5.5, thinking: {"type": "disabled"} returns a 400. thinking: {"type": "between_tools"} keeps up-front thinking off and works at high effort or below (what’s new).
• Check MCP servers that sit behind a cross-origin redirect. After the TypeScript SDK 2.3.0 client update, those requests fail instead of following the redirect. Point the transport at the final URL, or set redirectPolicy: 'follow' on the transport to get its old behavior.
• Cursor users: treat November 12 as tentative. OpenAI’s help article still only proposes that date, says Cursor may end access sooner, and promises the official termination date “as soon as it is confirmed between both companies.”

What’s Noise (for now)

• GPT-6.1 Astra’s cancellation. The Wall Street Journal reported on September 28 that OpenAI dropped a planned October release after internal safety testing. Reuters (via Cyprus Mail) reports that OpenAI confirmed it, CNBC confirmed it separately, and an OpenAI spokesperson told CNBC other models are coming soon. We found no OpenAI post on it and no date for what comes next. Don’t plan around the next Astra.
• Argon’s benchmark lead. Google reports 77.9% on DeepSWE v1.1 and first place on AutomationBench at 51.3%, for a model almost nobody reading this can call yet. VentureBeat says most customers will wait for broader API access before they can test it.

Our Take

Three of the week’s model stories came with a condition on what you actually get. Argon went to defenders first, and OpenAI shelved GPT-6.1 Astra after safety testing. Sonnet 5.5 is the first Sonnet where higher-risk cyber tasks “visibly fall back to Sonnet 5,” so on security work the model answering you may not be the one you picked. Put the fallback rules in your eval notes next to the price.

Several open calls settled this week. Issue 13’s Gemini 4 advice held: Google announced Argon on September 30 with no date for developers. Docusign’s MCP server hit GA on September 30, as issue 12 said it would (developer docs). And Artificial Analysis now ranks Opus 5.5 and GPT-6.1 Sol in one index, the cross-lab comparison issue 13 said neither vendor sheet offered, though its Opus 5.5 entry runs “with fallback.”

Issue 9 filed a writeup under noise saying GPT-5.6 Sol at xhigh never decides a task is finished, and issue 12 kept it there. Toloka’s July 16 study of 711 enterprise tasks, which predates issue 9, splits the claim. Sol at xhigh scored 4.0 points below GPT-5.5 at xhigh, so the headline holds. But against Sol at medium, xhigh moved success only +0.2 points while creating more extra records (+1.3) and more attempts to reverse values it had already written (+1.0). Sol keeps acting without getting more right, which isn’t the same as never stopping.

One call does need correcting. Issue 11 told you to read Hy4-preview’s Hugging Face sidebar as community numbers, but every entry there cites Tencent’s own model card, so they repeat Tencent’s figures rather than adding a second source. Finally, the Codex notes feature issues 12 and 13 tracked moved backward. The entry for its setting, features.context_management.experimental_mode, now reads “The feature is not currently available” in the config reference, about a month after OpenAI said on September 3 that it would become the default “in the coming weeks.”

Quick Links

• Copilot model deprecations - Claude Opus 4.7, Gemini 3.5 Flash and 3.6 Flash, and Kimi K2.7 Code left Copilot on October 2, and Enterprise admins may need to enable the replacements
• A model guide for the GPT-6 family - OpenAI’s October 2 advice on choosing a GPT-6 model and managing long-running work
• Jev for Python engineers - The latest AI SDK for Python release (uv add ai) adds an experimental evaluate() API for Jev, TypeSafe’s universal classifier
• Barclays scales Claude - The bank expects Claude Code to reach 50% of its developers by the end of 2026

Glossary

Computer use - A model operating software through its interface (clicking, typing, scrolling) instead of an API. Claude’s version left beta on August 19 (issue 10). OpenAI’s Agents API added it at DevDay on September 29, and GitHub Copilot’s CLI and app followed in public preview on October 1.

Origin approval - The Agents API’s check before its hosted browser visits a new website. It approves the site, not each action on it, which is why OpenAI says to restrict the browser when an action needs confirmation.

MCP Events - A draft MCP extension that lets a server push events to a client instead of waiting to be asked. ChatGPT Work chats support its webhook delivery for servers on MCP 2.0.

Cyber fallback - When a safeguard flags a higher-risk security request, a model hands it to an older, less capable one. Sonnet 5.5 falls back to Sonnet 5, so results on security work can come from a different model than the one you asked for.