|
Hey there 👋
For as long as you’ve built with these models, one rule held: the cheap tier was the compromise tier. You reached for it when the budget was tight and “good enough” would do. This week that rule broke.
All week the story was the coworker agent — whose is better, whose shipped next. Claude Cowork went to your phone; two days later ChatGPT Work made it a two-horse race. Today we drop a level and look at the engine underneath them. It just got cheap and posted the best agentic tool-use score on the board — from the vendor half your shortlists weren’t watching. The catch, and the whole point of today’s issue: the same model got worse at telling the truth. Cheaper autonomy is here. The thing worth building this weekend is the guardrail around it.
The Big Thing
The best agentic tool-use model on the board is now one of the cheapest — with a catch
On Wednesday, xAI shipped Grok 4.5 — its first model built specifically for coding and agentic work, trained on real Cursor developer sessions. It’s live today in Grok Build, Cursor (all plans), and the API and console, with EU access landing mid-July.
Start with price, because it reframes everything else. xAI prices Grok 4.5 at $2 / $6 per million input/output tokens ($0.50 cached) — more than 60% cheaper than Opus 4.8 and GPT-5.5. Cheap used to mean you were trading down. Not this time.
Artificial Analysis ranks it #4 on its Intelligence Index (score 54). But the number that matters for builders: it posts the single best agentic tool-use score of any model measured — above every open-weight model and every Gemini model. The cheap tier is now top of the class at doing things.
The cost gap runs deeper than the sticker. On SWE-Bench Pro, xAI reports Grok 4.5 resolving tasks in ~15,954 output tokens versus ~67,020 for Opus 4.8 (max) — a ~4.2x efficiency gap. That token frugality, not just the per-token price, is where the real savings live. It ships with a 500K-token context plus built-in server-side web/X search and code execution. xAI also reports coding wins (SWE Marathon #1, Terminal-Bench 2.1 at 83.3%).
Now the catch, up front where it belongs. Independent measures put its hallucination rate up sharply — roughly 25% to ~54% versus Grok 4.3. It’s great at DOING and shakier at KNOWING. Point it at a task and it executes; ask it to assert a fact and you check the work.
Ship it? Route, don’t replace. Send tool-calling and long-horizon DOING to Grok 4.5 — behind validation gates, with a cheaper, steadier checker model on anything it states as fact. Re-baseline your routing table this weekend: make a cheap-capable tier the default for execution, escalate to a frontier model only when the task earns it, and add an output-verification step before an agent writes to anything real. Cheap + capable + confident + occasionally wrong is the one combination you never point straight at production.
Sources: xAI (x.ai/news/grok-4-5, Jul 8); independent benchmarks via Artificial Analysis, reported by TechTimes (Jul 9). Vendor figures are xAI-reported; independent figures are Artificial Analysis.
Tour de Headlines
🛡️ CISA just made an agent-builder platform federal must-patch infrastructure. On Jul 7, CISA added Langflow — the low-code framework teams use to wire agents into internal systems — to its Known Exploited Vulnerabilities catalog, with a Jul 10 remediation deadline for federal agencies. It’s the first AI-agent orchestration platform ever on that list. The read for everyone else is a posture shift, not a headline: the tool you prototyped an agent in is now infrastructure regulators treat as critical. Weekend action — inventory where your agent-builder platforms touch real systems, patch to the fixed release, restrict access, rotate LLM-provider and cloud keys, and put agent orchestrators under change control.
🛒 Update: Salesforce agents can now transact, not just chat. Building on the Agentforce platform, Salesforce took Agentforce Commerce generally available ahead of peak season — a commerce-specific Shopper, Buyer, and Merchant suite wired natively into catalog, inventory, and orders. These act on real inventory: the Shopper Agent carries one conversation from discovery to checkout to service; the Buyer Agent runs B2B orders over WhatsApp and SMS with no portal login. ChatGPT integration is GA this month, with Google Search AI Mode and the Gemini app later this summer. The tell: an agent acting on live inventory inside an outside AI channel is now a shipped feature, not a bolt-on chatbot.
🧭 Cisco is handing all 90,000 employees an agent — and the router is the story. Starting end of July, every Cisco employee gets a personal AI agent that routes each task to the cheapest model that can do it instead of defaulting to a frontier model, with much of the stack running on-prem for cost and data control. CFO Mark Patterson: “It’s not going to burn a whole bunch of tokens with frontier models.” That’s today’s Big Thing routing thesis proven at 90k scale. It also arrives as Cisco works through ~4,000 layoffs announced this spring — the real adoption-trust test. Copy the router; watch the trust.
|
Sponsor
You’re about to make every agent check its own work. Your sales calls still ship unchecked.
Today’s whole issue is one move: wrap the cheap, capable, sometimes-wrong agent in a guardrail — a validation gate, a checker on the facts. Good. Now point that same lens at the highest-stakes system in your business that still runs with no gauge: the live sales and CS conversation. The discovery call that got misread. The deal where two people “aligned” on different things. The rapport that never formed — and closed-lost with no warning light. RapportScore scores how your team actually connects, on every call and email, so the misfires surface while you can still fix them. You wouldn’t let an agent write to production unchecked. Stop running revenue conversations blind.
See your team’s score →
|
Tool of the Day
✅ Automox MCP Server 2.2
For teams letting an agent touch endpoint patching — now it proposes, you see the blast radius, you approve.
Shipped Jul 7, Automox MCP Server 2.2 turns agentic endpoint management from a black box into a reviewable workflow. New in 2.2: an MCP-Apps visual surface that renders patch queues, blast-radius previews, and RBAC right inside the assistant; agentic “Patch by Severity” policy creation from plain language (say the intent, get a governed policy — no console-first build); and live capability discovery that shows exactly which safety flag gates each tool. It’s a textbook plan → approve → log contract applied to real infrastructure. Deploy note: start in read-only, wire one severity policy behind an approval step, and widen scope only once the approve-before-act surface is earning its keep.
Get the Automox MCP server →
Worth a Click
- Airia shipped Enhanced Cost Optimization — per-agent budgets that treat a whole agent run as one unit, hard and soft inline limits, live loop-detection for runaways, and spend attribution across provider, model, team, and agent. The FinOps guardrail that stops the runaway at runtime, not on next month’s invoice.
- Google’s AlphaEvolve is now generally available on the Gemini Enterprise Agent Platform — a code-optimization agent that pairs server-side LLM exploration with secure client-side execution to beat human-designed baselines (FM Logistic cut warehouse routing ~10.4%). The agent that improves your algorithms, not just answers about them.
- An agent-building startup let its own agent run its fundraise end to end — it fielded questions from 130+ investors, drafted memos, and tracked which slides backers lingered on. A rare, legible proof that an agent can carry a multi-week, high-stakes coordination workflow with a human only on the final call.
The week’s real lesson isn’t a product — it’s a shape. Agentic doing got cheap and got better: Grok 4.5 at $2/$6 with the best tool-use on the board, Cisco routing 90,000 people to the cheapest model that works. And the exact same model got worse at telling the truth. Cheap + capable + confident + occasionally wrong is the most dangerous thing you can point at production. So the weekend build isn’t a new coworker — it’s the guardrail around the cheap one: a validation gate on outputs, a checker on facts, vuln management on the platform, a budget on the spend. Rent the cheap engine; own the guardrails.
Have a re-baseline-your-routing weekend.
— Ace, for The Agent Stack
|
