Eversynced AI Weekly - September 28, 2026

Your Monday briefing on AI that matters for developers.

Issue #13 - September 7 - September 27, 2026.

This Week’s Big Move

On September 22 Anthropic released Claude Opus 5.5, and the pitch is price. Anthropic says it matches Claude Fable 5.1 on most work and “costs 40% less to run than Opus 5.” Input and output drop 20% to $4 and $20 per million tokens. Cache reads drop 60% to $0.20. Claude Code v2.1.280 made it the default Opus the same day.

But upgrading can break your code. The what’s new page lists four breaking changes from Opus 5: thinking can’t be turned off, forced tool use returns a 400, thinking blocks are tied to the model and the conversation, and the old computer_20251124 tool is rejected on the Claude API and Google Cloud. The default effort also drops from high to medium. Anthropic measured the 40% at default settings, so part of it comes from that. Compare bills at the same effort.

The bigger news is two API betas. You can now add or change a tool mid-conversation without losing the prompt cache, and you can trigger compaction whenever you choose. Together they nudge you toward agents that only append to the conversation. So read the compaction guide before you build your next agent loop.

Tools & Agent Updates

Claude Code v2.1.267 to v2.1.283 - Sixteen releases from September 9 to 25. v2.1.277 reads AGENTS.md when a project has no CLAUDE.md. v2.1.283 adds a managed availableModelsMatch setting: set it to "exact" and new models stay blocked until you list them. It also adds /doctor prompt-audit, which checks CLAUDE.md files, skills, agents and commands written for older models. v2.1.275 fetches npm plugins with npm pack --ignore-scripts. An update to issue 12: v2.1.277 removed the TaskOutput tool, so the taskOutputMaxChars setting we covered and the TASK_MAX_OUTPUT_LENGTH environment variable no longer have any effect.

Cursor Projects - Shipped September 10 in beta, rolling out to all users. A coordinator agent plans the work and “doesn’t write code itself,” handing tasks to subagents and bringing results back for you to check. It runs on its own cloud computer, keeps a set of files that sync across every cloud and local machine its agents use, and can watch a Slack channel, run on a schedule, or follow your PRs without a prompt.

Cursor Rollouts and Security Review - Two bots for Teams and Enterprise, launched September 23. Rollouts writes a monitoring plan as a PR comment, then checks the deploy against your logs, metrics and traces per environment. Depending on configuration, it can open a revert PR, but it “does not merge or roll back on its own today.” Security Review posts one comment per PR on exploitable bugs and skips drafts. For 10 days from launch, Cursor includes Rollouts credits for roughly 50 changes on Teams and 500 on Enterprise.

JetBrains Air - JetBrains announced one agent line on September 22: Air in the IDEs, Air Teams for coordination, and Air Governance, formerly JetBrains Central, for policy and cost. All three are available now, with more products promised “as the system develops.” Third-party agents plug in through Agent Client Protocol and its registry, and Air doesn’t require customers to use Junie, JetBrains’ own agent.

The Cache Line Is the Price Now

Two labs repriced on the same day, and the traffic data explains why:

• Opus 5.5’s cache reads - The headline cut is 20%. The cache-read cut is 60%, and Anthropic’s own parenthetical explains why that’s the one to read: cache reads “make up the majority of agentic and coding work costs.”
• GPT-6 Sol and Luna - OpenAI’s September 22 tiers come in 50% below their GPT-5.6 promotional pricing, at $2 and $10 per million tokens for Sol and $0.10 and $0.50 for Luna, with 90% discounts on cached input.
• Better prompt caching for GPT-6 - OpenAI’s companion post. Cache discounts now apply to eligible prefixes reused within a 30-minute window. You can change reasoning effort between responses with a configuration_update without breaking the cache, and OpenAI says to restrict tools with allowed_tools rather than removing definitions.
• Vercel’s AI Gateway index - Published September 17. Open-weight models processed 56% of August’s gateway tokens and took 14% of spend, and the price per token fell 23.2% that month.

Our read is that cache hit rate now moves your bill more than the headline cut does. Both labs shipped changes that keep the cache alive through edits that used to break it: Anthropic’s inline tool additions work “without losing the prompt cache,” and OpenAI’s effort switch works “without breaking cache,” alongside its advice to use allowed_tools instead of removing definitions. That turns architecture choices into billing choices. An agent that rewrites its system prompt or swaps its tool list mid-run gives up cache pricing on everything after the change. One that only appends keeps it. And Vercel’s numbers point at why. Token volume already went to open weights, so the frontier labs are fighting over the smaller share of work that still justifies their prices, and caching is how they make that share cheaper.

What This Means for You

What Matters

• Re-run your effort sweep before switching to Opus 5.5. The default is now medium, and the model tends to think more per turn at any given level, so a setting copied from Opus 5 measures something else. Start from the migration guide and remove any thinking disable settings first.
• If you manage Claude Code for a team, pin your models. v2.1.280 moved Pro and Team Standard to Opus by default. Set the managed availableModelsMatch setting to "exact" (added in v2.1.283) if you want releases blocked until someone approves them.
• Remove taskOutputMaxChars from your settings. It stopped working in v2.1.277, and so did the TASK_MAX_OUTPUT_LENGTH environment variable.
• Cursor users: OpenAI’s shutoff notice from issue 11 still proposes November 12. Before October ends, find the saved modes that pin a GPT model.

What’s Noise (for now)

• Every cross-lab benchmark from September 22. OpenAI published Sol and Luna the same day as Opus 5.5, so none of its Claude comparisons use Opus 5.5: most run against Opus 5, and four also add a Fable model, 5.1 or 5. Its AutomationBench chart says the Fable 5.1 datapoint “omits the cost of the Opus 5 fallbacks, which occurred on ~40% of tasks.” Anthropic’s own table notes that when safeguards intervened, tasks were finished by Opus 4.8 or Opus 5. Neither sheet compares current against current. Run your own.
• Gemini 4 timing. Per 9to5Google, what Google DeepMind’s chief said is coming “as soon as possible” is an early post-training output. There’s no date and no Google page. Don’t hold a migration for it.

Our Take

Kiro Crew and OpenAI show the same problem from opposite sides. Kiro shipped with no blog post, just a changelog line: Kiro Crew v0.6.0 made Claude Code and Codex selectable backends on September 5, behind Developer Mode. OpenAI announced without shipping: the Codex config reference still listed Astra’s notes feature as “off by default” on September 27, 24 days after OpenAI said it would become the default “in the coming weeks.” So blog posts are a weak signal either way. If a missing feature is what keeps you from switching tools, don’t wait for a post. Read the changelog, or spend ten minutes trying it.

That quiet changelog line caught us too. Issue 11 said Kiro Crew didn’t help Claude Code or Codex users. Two other open calls settled. Issue 8 filed Qwen3.8-Max as noise, and that call split. The weights did ship: the open-weights base of Qwen3.8-Max, Qwen3.8-2.4T-A95B, is on Hugging Face, but under a custom license, while Qwen3.8-27B is Apache 2.0. And Claude Code’s weekly limits landed where issue 11 said, 25% above pre-promotion levels from September 14.

Quick Links

• State of agent skills - skills.sh hit one million skills in seven months, and 375 of them account for 62% of installs
• Grok 4.7 in GitHub Copilot - SpaceXAI’s latest model for coding and knowledge work at $2 and $6 per million tokens, rolling out gradually to paid Copilot plans
• engrams - Cortex open-sourced its platform for running coding agents in your own cloud, with your own Anthropic or OpenAI key
• Managed Agents sessions in your terminal - Two minor versions after the ant apply issue 12 covered, ant CLI 1.32.0 adds ant beta:sessions connect to follow a session and approve pending tool calls

Glossary

Effort parameter - The dial that sets how much a Claude model thinks before answering, from low to max. On Opus 5.5 it replaces turning thinking off, and its default dropped a notch, which is why a price comparison needs it held constant.

Cache read - Input the provider already processed on an earlier request with the same prefix, billed at a steep discount. Agents resend their instructions, tools and history on every call, which is why Anthropic says cache reads are most of an agentic coding bill.

Append-only conversation - An agent loop that only adds messages to the end and never edits earlier ones. It keeps the cache warm and, on Opus 5.5, keeps thinking blocks valid.

Thinking block - The reasoning a model returns alongside its answer, which you pass back in tool-use loops. Opus 5.5 binds each block to the model and conversation that produced it, so on accounts created since August 31, editing anything earlier can turn a replay into a 400.

Agent Client Protocol (ACP) - A standard connection between an editor and an agent’s full harness. It’s how JetBrains Air runs third-party agents and how Kiro Crew now drives Claude Code and Codex.