Four New AI Models Landed at Once. Here’s the Simple Routing Rule.

2026-09-23

Maestro’s take

Four new models arrived in roughly two days. Every vendor brought a chart proving it had won.

It is too early for that trophy.

The useful change is simpler: developers now have a much wider price ladder. Stop asking which model is best. Ask which job deserves which model.

TL;DR

• GPT-6 Luna is the cheap lane for small, repeatable work.
• GPT-6 Sol is the sensible first try for everyday coding and agent jobs.
• Claude Opus 5.5 is the premium lane worth testing on difficult work where a better result could repay the higher price.
• Grok 4.7 is a cheaper-output alternative worth testing, especially if you already use Cursor, Grok Build, or GitHub Copilot.

Those are starting points, not a verdict. The public comparisons are still mostly vendor-run. Test the models on your repository before promoting one to “default.”

A rough cost sketch

The prices below are the three vendors’ standard API list prices per one million tokens as published in their launch materials. The last column is simple arithmetic for a job using 100,000 input tokens and producing 20,000 output tokens.

This is not the price of a finished feature. Caching, long-context rates, fast modes, retries, tool calls and review time can change the total. Subscription plans also hide costs differently. A model that finishes correctly in one pass can be cheaper than a bargain model that needs three rescues.

Why Maestro users care

A team does not need one default for every agent.

Use Luna for bounded jobs: classify an issue, summarize a log, rename a known field, or draft routine release notes.

Start everyday implementation with Sol. GitHub describes it as a balanced choice for interactive and agentic coding. OpenAI cut Sol’s API price 50% from GPT-5.6’s promotional pricing.

Try Grok 4.7 on the same mid-tier lane. Its output tokens cost less than Sol’s. But xAI’s launch chart compares it mainly with older GPT-5.6 Sol and Fable 5.1, not yesterday’s GPT-6 Sol or Opus 5.5. That is a reason to test, not a reason to crown it.

Test Opus 5.5 on the hardest jobs or work that is expensive to get wrong. Anthropic says Opus 5.5 can match much of Fable 5.1’s work and costs less to run than Opus 5. Those are Anthropic’s measurements. Independent, apples-to-apples tests of all four new models do not exist yet.

GitHub began rolling all four models into Copilot’s model picker; availability varies by plan. That makes a same-harness comparison easier for eligible users. It does not make the results identical across Codex, Claude Code, Cursor, Grok Build or Maestro. The harness still changes the outcome.

Do this, not that

• Do not move every agent to the most expensive model. Route repetitive, low-risk work to a cheaper lane.
• Do not trust one vendor’s leaderboard as a purchasing decision. The prompts, effort settings, tools and graders differ.
• Do not optimize token price alone. Measure successful completion, retries, review minutes and defects.
• Do keep an escalation path. Let a cheaper model try first only when failure is easy to detect and recover.

One thing to try

Pick ten real tasks from your backlog.

Run each task with the same instructions and acceptance tests on two models. Hide the model names from the reviewer. Record total tokens, elapsed time, retries, human review time and whether the change passed.

Then route by evidence:

• Small and obvious: Luna.
• Normal coding: your best mid-tier result, likely Sol or Grok.
• Hard or expensive-to-fail: Opus 5.5 or your premium model.

The model menu changed overnight. Your routing policy should change only after the receipts arrive.

Sources considered

• OpenAI, “Introducing GPT-6 Sol and Luna” (Sep 22, 2026)
• Anthropic, “Introducing Claude Opus 5.5” (Sep 22, 2026)
• SpaceXAI, “Introducing Grok 4.7” (Sep 21, 2026)
• GitHub Changelog, model availability for Opus 5.5, GPT-6 Sol and Luna, and Grok 4.7 (Sep 21–22, 2026)

Every performance claim above is attributed to the company that published it. Maestro’s routing advice is analysis, not an independent benchmark. AI-assisted; reviewed by an independent AI editor.