Near-Flagship AI at One-Fifth the Price: Why Model Routing Is a Core AI Skill
Dr. Navot Akiva
2026-10-26
Using a flagship AI for simple tasks is like using a Ferrari to deliver mail. Discover how model routing cuts inference costs by up to 80%—and how to build a smart routing architecture.
Using a massive AI model to answer simple customer service queries is like using a Ferrari to deliver the mail. In September 2026, the AI industry handed us a perfect case study of that idea.
Three Models in Four Weeks
On September 3, OpenAI introduced GPT-6 Astra, its new flagship model, at a standard API price of $10 per million input tokens and $50 per million output tokens. On September 22, it expanded the family with two cheaper models, GPT-6 Sol and GPT-6 Luna. A week later came GPT-6.1 Sol.
According to OpenAI, GPT-6.1 Sol performs close to Astra on agentic coding, computer use, and professional work, while costing one-fifth as much for standard input and output tokens: $2 and $10 per million. On one coding benchmark it matched Astra at roughly one-fifth the cost; on a computer-use benchmark it came within 2.1 points at roughly one-seventh the cost.
Independent testing points the same way. Artificial Analysis measured an average cost per task of $0.72 for GPT-6.1 Sol at maximum effort, compared with $3.26 for Astra.
Cost Has Become a Design Decision
This pattern goes well beyond one vendor. Every major AI lab now offers a ladder of models: a flagship for the hardest problems, a mid-tier workhorse, and a fast, cheap option for simple tasks. OpenAI's own lineup runs from Luna at $0.10 per million input tokens up to Astra at $10 , a 100x price difference within a single family.
And cost is starting to bite. McKinsey's 2026 State of AI survey found that about one in five organizations already limits its AI use because of operating costs, including token costs. Organizations that send every request to the most expensive model pay a premium for intelligence they often do not need.
What Is Model Routing?
Model routing means deciding, for each task, which model should handle it. A law firm works the same way. A senior partner does not proofread contracts, and a paralegal does not argue before the Supreme Court. Work goes to the person whose skill matches the task, and the firm escalates only when needed.
An AI system can be organized identically. A lightweight router, which can itself be a small model or a set of rules, looks at each incoming request and sends it to the cheapest model that can handle it reliably. In a customer support system, for example, a router might send password resets and order-status questions to the cheapest model, product comparisons to a mid-tier model, and complex complaints with legal implications to the flagship, with a human reviewing the most sensitive cases.
How to Route Like a Professional
- Start cheap, escalate on failure: Send a task to a smaller model first. If the output fails a validation check, retry with a stronger model. Most tasks never need the upgrade.
- Measure cost per task, not per token: A cheaper model that needs three attempts can cost more than an expensive one that succeeds the first time. Benchmark on complete tasks.
- Use caching: Repeated content such as system prompts and long instructions can be cached at a steep discount. GPT-6.1 Sol's cached input costs $0.10 per million tokens, 95% below its standard input price. As one industry analysis put it: don't build your cost model on the output price.
- Build an evaluation set first: You cannot route intelligently without measuring quality. A small set of real tasks with known good answers shows which model is "good enough" for each job.
- Re-test with every release: Three releases in four weeks means today's best routing decision may be wrong next month. Vendors update and retire models constantly, and a repeatable evaluation process is the only real defense against that drift.
Sources
- OpenAI. "GPT-6 Astra: A new generation of intelligence." https://openai.com/index/gpt-6-astra/
- ProPakistani. "GPT 6.1 Sol Brings Astra-Level Performance at 5x Lower Cost." https://propakistani.pk/2026/09/30/gpt-6-1-sol-brings-astra-level-performance-at-5x-lower-cost/
- MetricNexus. "GPT-6.1 Sol: Benchmarks, Pricing, and Near-Astra Results." https://metricnexus.ai/blog/gpt-6-1-sol-benchmarks-pricing
- BrainDetox. "How Far Does One-Fifth of Astra's Price Go: GPT-6.1 Sol Pricing Structure and How to Choose." https://braindetox.kr/en/posts/gpt_6_1_sol_release_pricing_2026.html
- McKinsey & Company. "The state of AI in 2026: On the road to ROI." https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- Tech Jacks Solutions. "OpenAI Releases GPT-6.1 Sol at DevDay 2026." https://techjacksolutions.com/?p=91537









