Yesterday the money moved local (Ollama 65M Series B + Muse Spark 1.1). The day before that, your coding agent looked like an attacker on Sophos EDR. Today the independent numbers landed — and so did a federal must-patch for agent builders who leave flows exposed.

If you route models for agent harnesses, this is the Sunday stack brief: what Sol actually scores, what it costs vs Fable, and why your visual agent platform just became core infrastructure security.

Sol owns the Coding Agent Index

Artificial Analysis published independent pre-release-style numbers that matter more than launch-day marketing: the Artificial Analysis Coding Agent Index pairs models with real agent harnesses across DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA.

Headline for agent builders:

On the broader Intelligence Index, Sol (max) sits at 59 — one point under Claude Fable 5 (max) — at roughly one third the cost per task (~1.04 vs Fable-class spend in AA’s framing). Terra and Luna land at 55 and 51 with deeper cost cuts.

OpenAI’s own GPT-5.6 product post aligns with the same story: Sol as coding flagship, Programmatic Tool Calling in the Responses API to keep intermediate tool spam out of the model loop, and ultra as a multi-agent parallel mode for hard jobs. Pricing stays tiered: Sol 5/30, Terra 2.50/15, Luna 1/6 per million input/output tokens, plus OpenAI’s first cache-write premium (1.25x input) with 90% cache-read discount still in place.

Practical read: stop treating “frontier” as a single SKU. For volume coding loops, Terra/Luna (or Grok 4.5 / Muse Spark from earlier this week) are the Pareto play. For hard judgment and long-horizon agent runs, Sol at max is now the independent coding-agent leader — not a rumor, a scored harness result.

Enterprise embedding: M365 Copilot defaults to 5.6

The other OpenAI beat that matters for non-dev surfaces: GPT-5.6 is now the preferred model in Microsoft 365 Copilot — Word, Excel, PowerPoint, Chat, and Cowork — per OpenAI’s July 9 partnership note and Microsoft’s Tech Community rollout. Millions of knowledge workers just changed models without changing apps.

That is the same agentic spillover Anthropic is chasing with Claude Cowork expanding beyond coding into general office work. The coding-agent wars are not only about SWE-bench. They are about who owns the default inside the tools people already open at 9 a.m.

Langflow IDOR is on CISA KEV — patch or rotate keys

While the benchmarks were digesting, the security side of agent platforms got federal: CVE-2026-55255 (Langflow IDOR on /api/v1/responses) landed in CISA’s Known Exploited Vulnerabilities catalog on July 7, with federal agencies ordered to mitigate by July 10.

Help Net Security and GitHub advisory GHSA-qrpv-q767-xqq2 spell out the failure mode cleanly:

• Pre-1.9.2, an authenticated user can invoke another user’s flow by ID — no ownership check
• Langflow flows routinely embed LLM API keys, cloud credentials, DB secrets
• Sysdig saw in-the-wild abuse from late June; operators injected “leak api keys” style prompts into hijacked flows
• Paired with the earlier RCE path (CVE-2026-33017), same instances got dual-path abuse in the same week

This is the first time an AI agent-building platform sits on the same must-patch footing as core OS/network CVEs in the KEV narrative. Treat visual agent orchestrators as production infrastructure:

# Minimum hygiene if you run Langflow
# 1) Upgrade to 1.9.2+
# 2) Lock who can reach the UI/API
# 3) Rotate every provider key that ever lived in a flow
# 4) Assume multi-tenant SaaS = higher blast radius for IDOR

If you self-host agent graphs for demos, side projects, or internal tools, assume the keys are already gone until you prove otherwise. Rotate OpenAI/Anthropic/Google/cloud tokens that were stored in flow configs. This is the same class of problem as agentjacking via error monitors — agent platforms are becoming credential stores with a UI.

Governance tooling is racing the exploits

Same news window: enterprise agent security vendors are shipping like the market finally noticed agents have tool access. From the AI Agent Store week digest: Codenotary AgentMon 3 (runtime policy learning from agent behavior), Automox MCP Server 2.2 (governed endpoint ops), First Recon AI Security Runtime (inline inspect + audit of human/agent/tool traffic). Token-cost playbooks claim 30–75% savings via tool-call minification and thinking budgets — the boring side of agent ops that keeps monthly bills from eating the roadmap.

Signal: capability (Sol/Fable/Grok/Muse) and capital (Ollama) shipped earlier this week. Security and cost control are the lagging layers catching up over the weekend. If your org is still treating Langflow/n8n/Dify as “just a prototype UI,” CISA just disagreed.

Looking ahead: Gemini 3.5 Pro ~July 17

Still not general-availability confirmed as of this writing, but third-party trackers keep pointing at a mid-July Gemini 3.5 Pro window (often cited around July 17) after the June slip. DeepSeek V4 stable timing chatter sits nearby. Do not re-architect routers on rumor dates — do keep a Gemini seat in the eval matrix so you are not surprised when Vertex/API IDs flip from preview to GA.

Stack notes for Sunday

1. Re-score your coding router with AA numbers. Sol (max)/Codex for hard agent loops. Terra/Luna or Grok 4.5 / Muse Spark for volume. Fable still the intelligence-index co-lead if your suite favors Claude Code.
2. Budget for cache-write pricing on GPT-5.6. Explicit cache breakpoints matter more when writes cost 1.25x.
3. Patch Langflow to 1.9.2+, rotate keys, restrict network. Same hygiene for any agent graph that stores provider secrets in flow JSON.
4. Expect M365 Copilot behavior changes if your org is on preferred-model GPT-5.6 — prompt playbooks from last week may overspend or under-ask.
5. Keep the local escape hatch from yesterday’s Ollama post. When cloud keys leak or policy gates slam, open weights on funded infra still ship code.

Useful kit if you are dual-running cloud agents next to local loops: a used NVIDIA A6000 / 6000 Ada for overnight open-weight harnesses, a solid USB-C docking station that survives multi-laptop agent farms, and a mechanical keyboard that does not die mid-refactor. For cloud seats, Cursor remains a short path to Grok 4.5; Codex maps cleanly to Sol; Claude Code still owns the Fable/Opus lane.

The launch week was about who shipped. The capital week was about the local floor. This weekend is about measured cost-performance and not leaving your agent graph on the public internet with yesterday’s keys. Score it. Patch it. Route it.

Affiliate disclosure: Amazon links above use our Associates tag (spaghettistor-20). We only link tools we would actually buy for agent harness work.