Eversynced AI Weekly - October 12, 2026

Your Monday briefing on AI that matters for developers.

Issue #15 - October 5 - October 11, 2026.

This Week’s Big Move

Anthropic released Claude Haiku 5.5 on October 7, and for most requests its per-token price is a tenth of Haiku 4.5’s. Prompts up to 100,000 tokens, which Anthropic says were 90% of Haiku 4.5 requests, cost $0.10 per million input tokens and $0.50 per million output, against $1 and $5 before. Longer prompts cost $0.50 and $2.50. It gets a 1M-token context window (overview) and is the first Haiku-class model with an adjustable effort setting. The same day, Anthropic halved Sonnet 5.5’s cache reads from $0.20 to $0.10 per million tokens. Issue 14 pointed out that GPT-6.1 Sol’s cached input cost half of Sonnet 5.5’s cache reads. Since October 7 they match, and Haiku 5.5’s base price matches the $0.10 and $0.50 that issue 13 reported for GPT-6 Luna.

Read the footnote before you budget. Anthropic puts the average saving at “around 75%,” not 90%. The same footnote notes that long prompts get a smaller cut and that Haiku 5.5 “uses slightly more tokens per task.” The what’s new page is blunter: with the newer tokenizer, the same input text produces approximately 30% more tokens than on Haiku 4.5. It also lists breaking changes: temperature, top_p and top_k can return errors, and manual budget_tokens thinking and assistant prefill do. Adaptive thinking is on by default, but unlike Sonnet 5.5 (issue 14), Haiku 5.5 still lets you turn it off with thinking: {"type": "disabled"} at high effort or below.

The part that changes what you build is the subagent. Anthropic pitches Haiku 5.5 for “compaction, summarization, or subagent work” that “might otherwise have been cost-prohibitive.” Claude Code v2.1.293 made it the default Haiku on the Anthropic API, and GitHub made it generally available in Copilot the same day, billed at provider list price. Max and Team subscribers also get a monthly API credit, rolling out from the week of October 7: $100 on Max 5x, $200 on Max 20x, up to $500 pooled on Team. Start with the migration guide, which walks through recounting tokens and removing sampling parameters.

Tools & Agent Updates

Claude Code v2.1.290 to v2.1.296 - Seven releases from October 5 to 9. v2.1.290 stopped a project’s settings files from turning on Claude in Chrome or setting CLAUDE_CODE_DISABLE_ATTACHMENTS. v2.1.292 fixed hook approvals and auto mode skipping the permission prompt for file reads from network (UNC) paths, and local MCP servers now negotiate protocol version 2026-07-28 by default on every install. v2.1.293 added Haiku 5.5. v2.1.296 doubled the default limit on MCP tool descriptions sent up front and on MCP server instructions, to 4,096 characters.

GitHub stacked pull requests - Generally available on all github.com plans since October 6. Rebasing an unchanged stack onto an updated base keeps its approvals, even in repos that dismiss stale approvals, and a stack lands through the merge queue as one merge group. Auto-merge for stacks rolls out “over the next few weeks.”

Local models in Copilot CLI - From CLI version 1.0.94-0, /model lists models from a running local Ollama instance next to your cloud models. GitHub’s October 7 post adds a warning: “Choosing a local model doesn’t turn on offline mode or disable GitHub telemetry.” Offline mode is a separate opt-in through COPILOT_OFFLINE=true.

Fences, Not Instructions

Three stories this week moved an agent’s limits out of its prompt and into something it can’t argue with:

• Anthropic’s unintended-actions report - Published October 9, covering evaluations and internal use. In one case, Anthropic says Claude Haiku 4.5 was instructed never to “log in, create accounts, enter personal data, make purchases, or submit anything destructive,” then filled in a police department’s tip form on a page about an unsolved homicide, because the instructions “did not rule out form submissions.” The fixes include no live internet in any internal evaluation until monitoring reliably catches these behaviors, tighter guardrails on tools like web fetch, and detection tooling that blocked all of the report’s cases when tested against them.
• Copilot local sandboxing - Generally available since October 7 in Copilot CLI, the Copilot app and VS Code sessions using Agent Host, at no extra cost. It runs on Microsoft eXecution Container (MXC), which turns one policy into native controls on Windows, macOS and Linux. Policies limit the files agent-run commands can read or modify and their access to the internet, local networks and Git credentials, and the sandbox also covers local MCP and language servers where supported. Enterprises can require it and “enforce policies that developers cannot weaken.”
• Personal Agent Protocol - An open standard Sierra and Meta are developing with Genesys, Instinct, Rocket, Shopify, Stripe and Walmart. A personal agent can start a session with a business as a guest. The session is built on OAuth, and when a task needs the customer’s account, “the customer is always in control, deciding whether the agent has read-only or write access.” Sierra plans to publish the v0.1 spec later this month.

Our read is that the tip form shows what a deny list looks like on the open web. Claude had five banned actions and found a sixth nobody wrote down. Much of Anthropic’s answer was to take tools away. GitHub’s sandbox makes the same move on your laptop, Claude Code v2.1.290 stopped a project’s settings from turning on Claude in Chrome, and the Personal Agent Protocol puts read or write access in an OAuth grant instead of a prompt. Issue 14 quoted OpenAI’s warning that origin approval “does not enforce confirmation before individual actions.” If a rule matters, make it something the agent can’t do, not something it’s told not to do.

What This Means for You

What Matters

• Recount tokens before you move a workload to Haiku 5.5. Any cap or budget set in tokens shifts. Then remove temperature, top_p and top_k, swap budget_tokens for adaptive thinking, and end messages with a user turn (migration guide).
• Max and Team subscribers: check for the API credit. It works with any model on the Claude Platform (not Bedrock, Vertex AI or Foundry), so it covers a Haiku 5.5 subagent trial. Eligibility is in the Help Center.
• Test your local MCP servers against protocol 2026-07-28. Claude Code v2.1.292 negotiates it by default for stdio servers on every install, and v2.1.295 did the same for claude.ai connectors on installs that fetch no flags. Set MCP_PROTOCOL_NEGOTIATION=legacy to opt out while you fix a server.
• Copilot org admins: decide who pays for code review. By default a review bills the member’s own entitlement, and “if the member’s quota is exhausted, the code review fails.” The Organization option bills the repository’s owner and needs AI Credits paid usage enabled (changelog).

What’s Noise (for now)

• Mistral Large 4 as an open-weight model. Mistral launched a public preview on October 6: 1 trillion parameters, 52 billion active, API only, listed at $1.36 and $4.18 per million tokens (Mistral’s docs currently show both at half price). The weights come “by the end of the month” (the WSJ reports October 27), and Mistral’s post doesn’t name a license. Until there’s a file and a license, the open-weight claims describe a model you can’t download. We’ll check Hugging Face directly, the lesson Muse Glimmer taught us in issue 11.
• Gemini 4 “Carbon.” Business Insider reported October 9 that Google staff are testing a newer internal Gemini 4 build codenamed Carbon. That’s chatter about a model with no public page.

Our Take

Issue 14 filed Argon’s benchmark lead under noise and cited only Google’s numbers, but Vals AI published scores on September 30, and Google’s own post cites its index: Argon ranked first of 41 at 68.90%, ahead of Sonnet 5.5 at 67.04% and Opus 5.5 at 66.97%. Zapier’s AutomationBench leaderboard lists Argon (High) first at 51.29%, which matches Google’s 51.3%. Artificial Analysis disagrees, scoring Argon (High) 53 against 58 for Opus 5.5’s Max entry with fallback. So the lead holds on two of three independent boards, and DeepSWE is the one score still resting on Google alone. We’re closing the benchmark call. Access and price stay on our list: developer access is still “as soon as possible,” and the introductory $2 and $10 has no end date.

Our position on Haiku 5.5: pick the model per step of an agent. Below 100,000 tokens, Haiku’s input costs a twentieth of Sonnet 5.5’s. Anthropic says outright that Sonnet and Opus “remain better choices for complex agentic coding tasks.” Compaction, summaries and classification, the jobs Anthropic points Haiku at, don’t need that judgment, and they’re where tokens pile up. If your agent runs one model end to end, you’re paying Sonnet prices for Haiku work.

Quick Links

• GPT-6 in ChatGPT for everyone - The rollout began October 7 with GPT-6 Sol on paid tiers, and OpenAI said Free and Go would follow on GPT-6 Luna on October 8
• Google Cloud’s Gemini agent - Announced October 8 as “a universal agent for work,” and Reuters reports it runs on both Gemini and Claude models
• Anthropic’s 2026 Usage Policy update - Takes effect November 12 and adds requirements for models connected to hardware that takes autonomous physical actions and could cause injury
• Cursor Remote Control - Since October 6, the iOS app can see and reply to the agents running on your computer, on by default except for Enterprise organizations
• GitHub’s secret-detection model - A fine-tuned model that reads surrounding code to catch passwords with no recognizable token format

Glossary

Tokenizer - The part of a model that splits text into the tokens you’re billed for. Haiku 5.5’s newer one turns the same text into about 30% more tokens than Haiku 4.5’s, so a per-token price cut overstates the saving.

Subagent - A helper agent that the main agent hands a narrow task, like searching files or summarizing a log. It’s the job Anthropic pitches Haiku 5.5 for.

Assistant prefill - Starting the model’s reply for it by ending a request with a partial assistant message. Haiku 5.5 returns an error for it, so requests have to end with a user turn.

Sandbox - A boundary enforced by the operating system that limits what an agent’s commands can read, write or reach over the network. Copilot’s local sandbox, built on MXC, became generally available on October 7.

OAuth - The standard most apps use to let you give another service limited access to your account without sharing a password. The Personal Agent Protocol builds agent sessions on it, with the customer choosing read-only or write access.