Stop picking a model.
Start assigning lanes.

On July 30, 2026, OpenAI cut the price of GPT-5.6 Luna by 80%. Luna now runs at $0.20 per million input tokens and $1.20 per million output tokens. That is a 25x price gap inside a single product line, and you cross it by changing one string in your API call.

The builder's guide OpenAI published on August 13, 2026 makes the implication explicit. The old advice was to grab the flagship at the highest reasoning setting and let it eat. That advice is dead. The guide reports that GPT-5.6 Sol at "low" reasoning effort outperformed GPT-5.5 at "high" reasoning with the harness held constant, and that Luna and Terra now handle long contexts and tool calls that used to require a flagship.

So the question changed. It is no longer "which model is best." It is "which model handles this specific request, at this latency budget, for this price." Here is the framework for answering that, the napkin math that proves it matters, and what to build this weekend.

The Three-Lane Rule

Every request your app receives belongs in one of three lanes. The job is assigning lanes, not picking models.

Lane 1 is the Fast Lane. High volume, low stakes, latency a human can feel. Classification, routing, moderation, summarizing a support ticket, pulling fields out of a PDF. That is Luna work at $0.20 in and $1.20 out.

Lane 2 is the Work Lane. Real business logic that has to be right but is not novel: multi-step workflows, mixed text and code, most production traffic. Terra dropped to $2.00 input and $12.00 output on July 30, 2026, from $2.50 and $15.00. That is the default lane for the boring middle of a product.

Lane 3 is the Frontier Lane. High stakes, expensive to get wrong, genuinely hard reasoning. Sol, and only when it earns the slot. On Agents' Last Exam, an evaluation of long-running professional workflows across 55 fields, Sol set a new high of 53.6 and beat Claude Fable 5 in adaptive reasoning by 13.1 points. That capability is real. It is also priced like it.

The rule in one sentence: assign the lane, then assign the model, never the reverse. Teams that skip the first step run Fast Lane traffic on Frontier Lane pricing and then call it a margin problem.

Building the Router Without Blowing Up Your Latency Budget

Napkin math makes this concrete. Say an app handles 1 million requests a month at roughly 1,000 input tokens and 500 output tokens each.

Run everything on Sol: 1,000 input tokens is $5,000, 500 output tokens is $15,000, so $20,000 a month. Run everything on Luna: $200 plus $600, so $800 a month. Now split it the way real traffic actually splits, say 10% Sol, 30% Terra, 60% Luna. That is $2,000 plus $2,400 plus $480, or $4,880. You just saved $15,120 a month by writing a classifier.

That is a 4.1x cut on the single largest variable cost, and nobody on the outside can tell. The tighter you scope each lane, the faster the savings compound.

Now the part most tutorials skip. Routing is not free. According to a 2026 production analysis of LLM routing patterns, rule-based routing adds about 1 millisecond, BERT-scale classifier routing adds 10 to 30 milliseconds, and embedding-based semantic routing adds 20 to 50 milliseconds just to generate the embedding. Cascading adds more latency every time an escalation triggers. An ounce in pre is worth a pound in post, so start with rules.

Rules mean if-statements. Request under 200 tokens with no tool call goes to Luna. Anything touching billing or medical content goes straight to Sol and skips the classifier entirely. Everything else goes to Terra. That is 80% of the benefit for 1 millisecond of overhead and zero extra infrastructure.

Speed is the other axis, and it is where the family model gets interesting. OpenAI's Ultrafast mode runs at up to 14x standard speed, which changes what you can put in front of a user in real time. Google is pushing the same lever from the other direction. Gemini 3.7 Flash hits 340 tokens per second, second fastest on the Artificial Analysis index, while scoring 56 on the Artificial Analysis Intelligence Index for ninth place overall. Fast and ninth-smartest is a perfectly good Fast Lane engine.

Think of the model family as a crew, not a hire. Sol is the 500 IQ contractor you call for the load-bearing wall. Terra is the reliable full-timer. Luna is the intern who never sleeps and costs $0.20 a million. Nobody pays contractor rates to sort the mail.

Now the damaging admission, because the routing hype needs one. The xRouter paper from 2026 warns about "the substantial overhead costs of multi-model systems," including multiple API integrations, maintaining router models, and monitoring distributed budgets. It concludes that for many applications, the benefits may not justify those costs versus simply using one capable model. LLMRouterBench found top routing methods delivering up to 4% accuracy gains and up to 31.7% cost reduction, but in controlled settings with clean task separation.

My read on this: the cost savings are real and large, the accuracy savings are mostly a lab artifact, and the failure mode is silent. When your router sends an edge-case query to Luna and Luna answers confidently and wrongly, no alarm fires. It is unclear whether most teams have the evaluation infrastructure to catch that drift before customers do. Route on cost, never on the assumption that cheaper is also smarter.

Three signals inside the same shift

2031

Zoom out five years. The model you pick today is a depreciating asset. The routing layer you build is a compounding one.

Look at the deprecation calendar. OpenAI announced on January 29, 2026 that GPT-5 Instant, GPT-5 Thinking, GPT-4o, GPT-4.1, GPT-4.1-mini and o4-mini would leave ChatGPT on February 13, 2026. GPT-4.5 was retired on June 27, 2026. o3 followed on August 26, 2026. A "best single model" has a shelf life of roughly six to twelve months. An abstraction layer with fallbacks survives every one of those sunsets.

The price data says the same thing from a different angle. Model Price Watch found the blended index for frontier flagships at $4.39 per million tokens on August 1, 2026, down only 3.9% from $4.57 on February 23, 2026. The frontier barely moved. Meanwhile the cheapest model clearing a GPT-4-class capability bar fell from $1.93 per million on March 1 to $0.11 on July 25, a 17x collapse in five months. Frontier prices are sticky. Floor prices are in freefall.

That gap is the asymmetric advantage. Frontier capability stays expensive, so buying it indiscriminately is a permanent tax. Yesterday's frontier becomes almost free, so anything you can push down a lane gets cheaper every quarter without you touching it. Model Price Watch put the flagship basket spread at 21x, from $0.54 per million for DeepSeek V4 Pro to $11.25 for GPT-5.6 Sol.

Amateurs optimize the model. Operators optimize the mix. By 2031 I expect "which model do you use" to sound the way "which server do you use" sounds now: a question about an implementation detail, not a strategy. The durable skill is knowing what a task is worth before you spend on it.

What to Build This Weekend

Build a router. Not a framework, not a platform. One function with three branches.

First, log a week of real requests from whatever you already run. Tag each one by hand into Fast, Work, or Frontier. Fifty examples is enough to see the shape, and you will probably find that most of your traffic is Fast Lane work being billed at flagship rates.

Second, write the rules as plain if-statements. Length threshold, keyword list for high-stakes topics, tool-call present or not. Keep it rule-based so overhead stays near 1 millisecond instead of the 20 to 50 milliseconds semantic routing costs. Default the fallthrough to Terra, not Sol.

Third, build a test harness before you ship. Take 20 requests you know the correct answer to, run them through all three models, and compare. If Luna nails 18 of 20 on a task class, that class is Fast Lane forever. Things will break here, and breaking on 20 saved examples is much cheaper than breaking in production.

Then make it visible. Lovable takes a prompt and ships a running full-stack web app with GitHub and Supabase already wired, which is more than enough to stand up an internal dashboard showing cost per lane per day. If you want the dashboard to match your design system instead of a generic kit, Ugic reads your existing Figma components and generates screens from them. Designing the mobile view, Play lets you prototype directly on an iPhone with real iOS materials.

One last thing worth watching if you sell seats. Rupt monitors for shared logins using identity verification and fraud signals, then routes those users toward paying. Routing your models cuts cost per request. Routing your freeloaders raises revenue per seat. Both are the same skill applied to different traffic.

Get one lane working this weekend. Add the second one next week.