Hey from Tobias,
Grok 4.5 shipped this week, and it does one thing Opus 4.8 can't: it runs up to 6x cheaper on real agent work, while landing just behind Claude's best model on the benchmarks I track. If you're paying frontier prices inside Claude Code, that number is hard to ignore.
Two frontier models actually launched in seven days. Here's where they land against Fable 5 and Opus 4.8, and two new tools I put on the site to help you compare.
Grok 4.5 Undercuts Opus 4.8 by up to 6x
The numbers first. Grok 4.5 costs $2 per million input tokens and $6 per million output. Opus 4.8 costs $5 and $25. On a real agent task, Grok also burns about 60% fewer output tokens, so the bill can come in up to 6x lower than Opus 4.8 for the same job.
On the agent benchmarks I track it lands just behind Fable 5. On one of them, AutomationBench, it actually beat both Fable 5 and Opus 4.8. It was trained alongside Cursor, so if you already use Claude Code, switching over is almost nothing. Same agent loop, a fraction of the cost, available through Grok Build.
The honest catch, at least for me: I can't test it yet. Grok Build isn't available in my region. I'm currently in the EU, where new AI tends to arrive late while AI Act compliance gets sorted out. So this is the one launch this week I'm reporting without hands-on time. The moment it reaches me, I'll run it through the same tests I run on everything else.
Let Fable 5 Plan, Let Grok 4.5 Build
Here's the setup I keep seeing on X this week, and it's a smart one. You run two models. Fable 5 sits on top as the advisor and guard: it reads the task, writes the plan, and reviews the result. Grok 4.5 does the actual building underneath.
The math is the whole point. Fable 5 lists at $10 per million input tokens and $50 output, the priciest model on this list. You don't want it grinding through every step. Let it touch only the plan and the final check, then hand the heavy work to Grok 4.5 at $2 and $6. You keep Claude's judgment as the guard and pay Grok prices for the grunt work.
Most agentic tools support this. In OpenClaw you set one model to plan and another to execute inside the same session. In Hermes Agent you use a delegate step, so Fable 5 hands the job to Grok in its own window and gets the result back to check.
GPT-5.6 Lands Just Under Fable 5 Too
OpenAI shipped GPT-5.6 the same week, in three versions: Luna, Terra, and Sol. Sol is the one that matters. It scores one point behind Fable 5 on the Artificial Analysis Intelligence Index, sits above Opus 4.8, and tops the coding agent index outright.
So in one week, two labs put out frontier models that both land right under Claude's best and match or beat Opus 4.8. Fable 5 still leads the hardest coding tests. The gap at the top is getting thin, and the price gap is getting wide. That's the good news for you: more real choice at every budget.
You can see exactly where all four stack up on my live Intelligence Index chart, updated as new scores land.
See the live benchmark chart →Two New Ways to Compare Your Stack
I keep madebyagents.com current with the newest models and benchmarks so you can play with the metrics and find the right stack for your project. Two new comparisons went live this week, each with a wizard that asks a few questions and points you to a stack.
• AI Inference Engines. Compare the engines that actually run your models, then let the wizard match one to your setup.
• AI Desktop Apps. Find the right local AI app for your machine, from a plain side-by-side or the guided wizard.
Match the model to the job.
Tobias
