A Fifth of the Price Broke Your Model Routing

OpenAI announced GPT 6.1 Sol this morning: near-Astra intelligence at a fifth of the price. It's sitting at the top of Hacker News for a reason. If your product burns real money on inference, every routing decision you made last quarter just went stale.

This is not a billing note. It's an architecture event. When near-frontier quality gets 5x cheaper, designs everyone called wasteful last month become the obviously correct ones today.

Why this matters

The metric that pays your salary is cost per solved task, not cost per token. A model at 20% of the price that passes your evals 95% as often wins almost any workload, because failures are cheap to detect and retry.

Price drops also compound. OpenAI shipped Dots, always-on agents, in the same news cycle. Agents that never sleep only pencil out when idle inference costs pennies. The cheap tier isn't the fallback anymore — it's the default.

How it works

Write the arithmetic down: cost = attempts × tokens per attempt × price per token. Cut price by 5x and you can buy five attempts for what one used to cost. So stop generating one answer and praying.

• Best-of-n: sample five completions, verify each against tests, schemas, or a rubric, keep the first pass. Verification is code you already trust.
• Cascades: send everything to the cheap tier first, escalate on refusal shapes or low verifier scores. The flagship sees maybe 10% of traffic.
• Background loops: nightly agents that triage, tag, and draft become rational when nobody's watching the meter.

Where this helps

• Support automation: Sol handles the routine 90%, humans or Astra get the weird 10%, cost per resolved ticket collapses.
• Code migrations: generate five candidate patches, run the suite, land the one that passes. Costs what a single flagship call used to.
• Backlog processing: classify and tag years of tickets or docs — the job that was never budgeted suddenly is.
• Always-on ops agents in the Dots style: watch dashboards overnight and open drafted incidents before standup.

Watch out

"Near" is doing heavy lifting in that headline. The quality gap concentrates in the hardest 10% of your prompts — pass rates look identical until they suddenly aren't. Only your own eval set can tell you which model you actually bought.

Also, models drift silently. Another front-page item today is Livenerf, a tracker asking whether Opus 5.5 has been nerfed. Providers rarely announce degradations. Pin dated model IDs where your provider offers them, rerun evals on a schedule, and expect ugly launch-week rate limits.

Try it yourself

The crudest useful eval harness: golden prompts with expected substrings, run against both tiers. Grab exact model IDs from the docs — marketing names and API slugs don't always match.

import os, requests

API = "https://api.openai.com/v1/chat/completions"
MODELS = os.environ["MODELS"].split(",")  # export MODELS="cheap-tier,flagship"

GOLDEN = [
    ("Extract the total from 'Invoice 41: total due $1,240.50'", "$1,240.50"),
    ("Severity of 'primary DB refusing connections'?", "critical"),
]

for model in MODELS:
    hits = 0
    for prompt, needle in GOLDEN:
        r = requests.post(API, timeout=30,
            headers={"Authorization": f"Bearer {os.environ['OPENAI_API_KEY']}"},
            json={"model": model,
                  "messages": [{"role": "user", "content": prompt}]}).json()
        hits += needle.lower() in r["choices"][0]["message"]["content"].lower()
    print(f"{model}: {hits}/{len(GOLDEN)} passed")

Grow the golden set to 50–200 cases before trusting any verdict.

TL;DR

• What changed: GPT 6.1 Sol ships near-Astra intelligence at a fifth of the price, with always-on agents (Dots) launching alongside.
• Why it matters: optimize cost per solved task — 5x cheaper turns best-of-n, cascades, and always-on agents into rational defaults.
• Try today: run your golden prompts against both tiers before touching a single line of routing code.