The short version: for two years, a model with Flash, Mini, or Lite in its name was a safe bet for the cheap tier — pick by the suffix and you'd land near the price floor. Last week's price war broke that shortcut. Gemini 3.6 Flash now lists at about $1.50 input / $7.50 output per million tokens — a mid-tier price, roughly 10x the models that actually define the floor. The real budget tier is three unglamorously-named models: Qwen3.7 Flash (~$0.03/$0.13), DeepSeek V4-Flash-0731 ($0.14/$0.28), and GPT-5.6 Luna ($0.20/$1.20 after its cut). The lesson isn't "switch to the cheapest" — it's that the name on the model stopped telling you its price band.

The shortcut that used to work#

The suffix system was genuinely useful while it lasted. Google had Pro and Flash; OpenAI had its full models and the Minis; Anthropic had Opus, Sonnet, and Haiku. Within one vendor, the small-name variant was reliably the cheap one, and across vendors the names roughly rhymed — "Flash-class" meant "fast and cheap." You could reach for the suffix without opening the pricing page.

That worked because names and prices were set at the same time, by the same launch. It stops working the moment prices start moving faster than names — which is exactly what happened last week.

What broke it#

Three repricings in the space of a week dropped the floor out from under the naming convention:

Meanwhile, Gemini 3.6 Flash kept a list price set weeks earlier: about $1.50 / $7.50 per million (Fello AI). Nothing wrong with the model — it's capable and deeply wired into the Google stack — but its price is now competing with GPT-5.6 Terra ($2/$12), a tier OpenAI explicitly labels mid. The badge says budget; the meter says mid-tier.

The name is a launch-day marketing tier; the price is a live market number#

This is the durable takeaway, and it will keep being true after these specific numbers go stale. A model's suffix is fixed at launch and rarely changes. Its price is a competitive response that moves whenever a rival cuts. Those two things drift apart, and the drift is widest right after a price war — precisely when you're most tempted to "just grab the Flash one" to save money.

So the first rule is boring and load-bearing: open the current pricing page before you pick. The suffix is a hint about capability, not a promise about cost.

Then read past the pricing page, too#

Even the live sticker doesn't set your bill, because you don't pay per token — you pay per finished job. The metric that matters is dollars-per-completed-run:

cost-per-completed-run = (total spend on a task type) ÷ (runs that actually succeeded)

Three things make the cheapest sticker lose on that metric:

  1. Failure and retry. A model at half the price that fails 30% of the time forces a retry, a longer chain, or a fallback to a pricier model. Two cheap failures plus one expensive success can beat one cheap success that never comes.
  2. Output length. Output is where the money is — Luna charges $1.20 out against $0.20 in, a 6x multiplier. A cheaper-per-token model that rambles can lose to a terser, pricier one on a completed run.
  3. Multimodal you don't use — or do. If your agent is text-only, don't pay for eyes: Luna and V4-Flash are text-only and cheaper for it. If it must read a screenshot or a scanned invoice, Qwen3.7 Flash is the standout — multimodal at the floor, where Gemini 3.6 Flash sees too but at ~10x.

How to pick, now that the name won't do it for you#

  1. Shortlist by capability, not suffix. Text coding/agent loops → DeepSeek V4-Flash-0731 or GPT-5.6 Luna. Needs to see → Qwen3.7 Flash. All-in on one cloud → the native option, priced honestly on its live page.
  2. Route real traffic to the candidate behind a proxy or gateway (LiteLLM, an AI gateway) so you can revert instantly — don't rewrite your stack to run a test.
  3. Measure cost-per-completed-run and quality for a week, per task type. Switch only where the candidate wins on the completed-run number.
  4. Clear the non-price gates first. Qwen and DeepSeek are Chinese-hosted APIs — confirm data-residency, privacy, and procurement before routing customer data, and don't ship on benchmarks you haven't reproduced.

For the head-to-head cost cases we've already run the pairwise math: GPT-5.6 Luna vs Gemini 3.6 Flash (its Luna numbers predate the July 30 cut), DeepSeek V4-Flash vs Qwen3.7 Flash on whether your cheap agent needs to see, and the method piece on why dollars-per-completed-run beats dollars-per-token. For the week's full context, see what V4-Flash-0731 means for founders and the August 3 Founder's Wire.

The memo for a team of one: the era where you could pick a cheap model by its name just ended. Open the live pricing page, ignore the suffix, and let dollars-per-completed-run on your own traffic make the call.