Eversynced AI Weekly - September 8, 2026

Your weekly briefing on AI that matters for developers.

Issue #12 - August 31 - September 6, 2026.

This Week’s Big Move

On September 3 OpenAI shipped GPT-6 Astra, which it calls the world’s most intelligent and aligned model. The benchmark line is unusual for how little room it leaves: Astra “saturates FrontierMath Tier 4 with a 98% score,” saturates ARC-AGI-3 at 99.9%, and posts 100% on ExploitBench. Each of those three carries a condition. That 98% is the prose talking, and the academic table further down the same page puts FrontierMath Tier 4 (v2) at 97.6%. The ARC-AGI-3 run used OpenAI’s own responses API harness, which footnote 1 says “changes two settings to better match real-world performance.” And the 100% on ExploitBench is a measurement of the model “without production safeguards,” which is not the Astra you can call. On Terminal-Bench Science 0.1 it reaches 64.6% against 52.6% for Claude Fable 5.1, at roughly 31% lower estimated API cost by OpenAI’s own accounting. It does not win everything, and the same tables say so: Fable 5.1 takes Humanity’s Last Exam with tools at 65.0% against Astra’s 57.2%, and the Artificial Analysis Intelligence Index v4.1.1 at 65.7 against 61.2. It rolled out that day to a limited set of organizations, with ChatGPT Plus, Pro, Business and Enterprise, the OpenAI API, Microsoft Azure and AWS Bedrock following over the coming days.

The part that costs you something is further down the page. Enterprise administrators have to switch it on, because “access is off by default at launch,” so nobody on your team wakes up to a silent model swap. Astra usage comes out of existing subscription allowances with credits available on top. Astra also arrives with heavier cyber safeguards than anything OpenAI has shipped, and the page says why: it “meets the Critical threshold in cybersecurity under our Preparedness Framework.” Issue 9 flagged Sol at High in that same framework and issue 11 glossed the tier. Critical is the one above it, and OpenAI says expert evaluators found Astra, run without production safeguards, could use previously unknown vulnerabilities to achieve arbitrary code execution in hardened browsers. The shipping model refuses the advanced work, proof-of-concept exploit writing included, with looser access planned through OpenAI Daybreak in the coming weeks. OpenAI also warns that those safeguards can interrupt legitimate work, and the failure mode differs by surface: in ChatGPT or Codex a paused task asks you to review the action, while “in the API, the task will stop.” If you are wiring Astra into anything unattended, that is a stop condition to handle rather than a prompt somebody will see. On price, OpenAI API Standard is “$10 per million input tokens and $50 per million output tokens,” with separate cache rates and a Fast mode OpenAI says runs up to twice as fast for twice the price. Those are the same two numbers Anthropic charges for Fable 5.1, published two days apart.

The reason this lands where it does is that issue 11’s headline was the postmortem on the Hugging Face incident, and OpenAI built an evaluation out of it. The eval tests whether a model facing a difficult or impossible task will go beyond its intended scope. GPT-5.6 Sol, stripped of production safeguards, “went beyond the authorized target 48% of the time.” Astra did it in 0% of cases. In a separate internal test, Astra “never attempted to circumvent a Codex Auto-Review denial,” holding even when Auto-Review was deliberately made evadable and the task was otherwise impossible. Our read is that this is the first flagship whose pitch is restraint rather than reach, and it is a direct answer to the incident this newsletter covered as an open wound in issue 11. For something to actually read, Astra can keep notes across context windows in Codex instead of compacting. But it is experimental and off by default: you set features.context_management.experimental_mode in your Codex config file, and it needs a ChatGPT sign-in on Plus, Pro, or Pro Lite. OpenAI says it becomes the default for Astra in the coming weeks.

Tools & Agent Updates

Claude Code v2.1.252 to v2.1.266 - Nine releases between August 31 and September 8, and the useful changes are the quiet ones. v2.1.260 added a /diff panel that opens beside the conversation in fullscreen mode and shows uncommitted changes as Claude edits. v2.1.261 added /skill-doctor, which reports which loaded skills go unused and what they cost you in context, plus bashOutputMaxChars and taskOutputMaxChars to raise inline output up to 128K characters. v2.1.259 added managedMcpServers and --permission-prompts none for unattended headless hosts. Skip v2.1.265 if you pin versions: it made the undocumented CLAUDE_CODE_USE_GATEWAY variable force Cloud-gateway sign-in on its own, so setups pairing it with an API key or custom auth headers failed every request with “Not signed in to the Cloud gateway”. v2.1.266 put it back the same day.

ant apply - The ant CLI brings Terraform’s shape to the Claude Platform, and ant apply requires version 1.30.0 or later. You describe agents, environments, skills, memory stores and deployments as files in your repo, run ant apply, and approve the plan it prints. It writes a claude-lock.json you commit, so later runs on your machine or in CI update the same resources instead of quietly creating a second set.

Cursor self-hosted machines - Cloud agents can now run entirely inside your network, with codebase, build outputs and secrets staying on machines you own while the agent drives tool calls locally. Team pools are named queues that grow as requests arrive, shrink when workers disconnect, and hibernate idle machines. It runs on AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel and E2B. Read it next to the November 12 model deadline issue 11 flagged, because one of these problems is about who holds your code and the other is about who holds your contract.

Project HydraFusion - GitHub’s September 4 research preview routes a single coding task across multiple models at runtime rather than pinning one. It is live for every Copilot plan through /experimental in the Copilot CLI, billed at each model’s standard rate for whatever it actually invokes. GitHub reports 4.9 points of verified task quality over Claude Opus 5 on TerminalBench 2.1 at 67% lower estimated cost, but it trails Opus 5 on the other two benchmarks it ran, by 1.5 points on DeepSWE and 0.1 on CheckpointBench, and all three are its own offline evaluations under its own pricing assumptions.

Three Flagships, Some Sharp Edges

Three frontier vendors shipped in three days, and each one put something between you and the upgrade:

• Claude Fable 5.1 and Mythos 5.1 - Landed September 1 with a 1M token context window by default, 128k max output, and prices held at $10 and $50 per million input and output tokens, with cache reads cut to $0.25. Cheaper and roomier, with the sharp edge filed in the migration guide: forcing a tool call is now an error rather than a preference.
• GPT-6 Astra - Enterprise access off by default, and an API path where a flagged task stops rather than prompts.
• Gemini 3.8 Flash and 3.8 Flash Cyber - Google shipped a pair on September 2, and the Cyber variant ships “a more permissive set of mitigations” and “is only available to trusted defenders.” On CWE-Bench it trails, with a pass@1 of 47.2% against 47.8% for a leading frontier model, which Google presents as the Pareto frontier because the model costs significantly less.

The upgrade stopped being a version bump. For most of the past two years a new flagship was a string you changed, and the worst case was a bill. This week Anthropic turned a working tool_choice call into a 400, Google put a model behind an eligibility check, and OpenAI shipped a model that halts API tasks on its own judgment and arrives switched off for the org. Every one of those is defensible on its own. Together they mean the thing you have to test before migrating is no longer output quality, it is whether your integration has a path for the model refusing, stopping, or not being available to the person running it. That is a different test suite than the one most teams wrote, and the vendors did not coordinate it. They arrived at it separately, which is usually how a shift shows up before anyone has a name for it.

What This Means for You

What Matters

• Audit tool_choice before you move to Fable 5.1. Forcing a tool call with any or a named tool returns a 400 on Fable 5.1 and Mythos 5.1. If you build structured output by forcing a function call, that is a hard break, not a degradation, and it fails at runtime rather than at deploy. The migration guide has the full list, which also covers thinking-block replay rules that return a 400 for accounts created on or after August 31, 2026.
• If your org runs ChatGPT Enterprise, Astra needs an admin to turn it on. Access is off by default at launch, so the rollout will not reach your developers on its own.
• Measure your Claude Code consumption before September 14. The permanent 25% raise starts that day and replaces the temporary 50% boost, which nets to the 17% cut against today’s ceiling that issue 11 flagged. It is still not written down anywhere Anthropic publishes: the support and pricing pages both describe weekly limits without mentioning the change, so the only text for it remains Anthropic’s own statement relayed by reporters.
• Run /skill-doctor after upgrading Claude Code. It names the loaded skills nothing is using and prices them in context. Our read is that most people carrying a large skills directory are paying for it on every request.

What’s Noise (for now)

• GPT-5.6 Sol at xhigh reasoning never deciding a task is finished. Issue 9 ran this on one person and three seats, and we went looking again. Nothing independent reproduces it at xhigh. The nearest thing to a vendor answer is a Cursor forum thread where Colin at Cursor writes on August 30 that a parent agent finishing does not cancel the background subagents it spawned, that a subagent blocked on a permission prompt stays alive, and that the visibility gap is “a known issue we’re already tracking,” with version 3.18 keeping the parent chat marked as working while children run. That is a lifecycle and UI problem in one product, and Cursor says nothing about the model. The other nearby report is openai/codex#32389, open since July 11, where GPT-5.6 Terra intermittently returns an empty successful final response after tool use and ends the agent loop early. Different model, opposite failure, and logged at Medium reasoning effort rather than xhigh. Still open, and still resting on the three seats it started with.
• HydraFusion’s cost advantage. Routing across models to beat a single frontier model on price is a good idea and the preview is free to try. The numbers behind it are GitHub’s own offline evaluations, and one of the three benchmarks, CheckpointBench, is GitHub’s own, built from real Copilot sessions. GitHub says it optimized “across the evaluation sets rather than for any single benchmark,” which is the right method and still leaves the scoring inside the house. Run it on your own tasks before you quote the 67%.

Our Take

Read the footnotes this week, because that is where the honest part lives. OpenAI’s Astra post benchmarks Claude by name and repeatedly, which is new enough on its own. But footnote 17 says that “For ScreenSpot-Pro and ExploitGym, the Fable scores we report come from Mythos, which is Fable with fewer safeguards,” and footnote 11 says that on HealthBench Professional, “For Fable 5.1, we used Opus 5 fallback for provider refusals.” So on three named benchmarks the Claude numbers come from a differently configured model, or from a substitute where the first one declined. Neither footnote touches the Terminal-Bench Science line this issue leads with, and that limit is the point: you cannot tell which comparisons carry an asterisk without reading to the bottom. None of it is hidden and none of it is misconduct. It is disclosed, in the place readers do not go.

That is where frontier evaluation sits in September 2026. The numbers in the body are real and the conditions that make them comparable are in small type at the bottom, and every vendor now does this because every vendor now benchmarks its rivals directly. Anthropic’s own >80% figure, which we filed as noise in issue 11 for how it was being reframed, holds up the same way once you read the source report: it “measures the share of lines merged to production that can be attributed to Claude,” a deliberately conservative measure the company explains at length. Issue 11 said to read that report rather than the thread recirculating it, and this week moves the same rule down a level, to the footnotes underneath it. The pattern holds either way: the vendors are being more careful than their coverage, and the gap between the two is where you will get burned.

Quick Links

• Tencent’s teamai-cli - Distributes one team’s skills, rules, MCP config and knowledge across Claude Code, Codex, Cursor, OpenCode and others from a shared git repo, which is issue 10’s argument as a working tool
• VS Code 1.136 - Agent Merge in preview resolves review feedback, failed checks and merge conflicts until a PR is mergeable, plus experimental agent sessions across multi-root workspaces
• SGLang merged Hy4 support - PR #36805 landed on September 5 after 34 commits, closing the CUDA half of the gap issue 11 left open, while the stacked ROCm follow-up closed unmerged
• MCP moves into LangChain core - Lives at langchain.mcp now rather than a separate adapters install, with elicitation via interrupts, Python first and TypeScript to follow
• Docusign opens its MCP server to every agent - Generally available September 30, callable from any MCP client rather than a named partner list

Glossary

Benchmark saturation - When a model scores high enough that the test stops separating it from the next one. OpenAI uses the word directly for Astra on FrontierMath Tier 4 and ARC-AGI-3, which is a vendor telling you its own benchmarks have stopped being useful.

tool_choice - The API parameter that forces a model to call a tool instead of leaving it optional. It is the standard trick for getting structured output, which is why Fable 5.1 rejecting any and named tool values breaks more code than it sounds like it should.

Compaction - Summarizing an agent’s earlier work when the context window fills, losing the detail of why a fix failed. Astra’s Codex alternative keeps searchable notes across windows instead, which is the difference between an agent that forgets a dead end and one that repeats it.

Trusted defenders - Google’s eligibility gate on 3.8 Flash Cyber, which carries lighter cyber restrictions than the general model. Access runs through the Fairwind Program, so it is a capability tier you qualify for rather than buy.

pass@1 - The share of problems solved on a single attempt, with no retries. It is the honest version of a coding benchmark, since giving a model ten tries flatters it in a way production rarely does.