Commercials · 7 min read · Aug 15, 2026

Four tiers, one rule: route by judgement.

Model releases arrived at a pace in 2026 that makes "which model should we use" the wrong question — it has a different answer every six weeks. The durable question is how you route work between tiers, because that decision survives the next release and it is the one that sets your bill.

Where the tiers sit, as of August 2026

Anthropic released Claude Opus 5 on 24 July 2026 at $5 per million input tokens and $25 per million output — the same price as the Opus 4.8 it replaced, so the improvement arrived at unchanged cost. It also carries an effort setting, letting the same model be run harder for difficult work or lighter to conserve tokens. It was Anthropic's fourth Claude 5 release in under two months.

OpenAI's GPT-5.6 line is explicitly tiered: Sol at $5/$30 for high-end reasoning and complex coding, Terra at $2.5/$15 as the everyday model at roughly half of Sol, and Luna at $1/$6 optimised for speed. Google shipped Gemini 3.7 Flash on 13 August 2026 with an introductory price running to 31 December 2026, alongside Gemini 3.5 Flash-Lite positioned specifically as a low-latency, low-cost subagent for high-volume automation.

Published prices per million tokens in August 2026: GPT-5.6 Sol at $5 input and $30 output, Claude Opus 5 at $5 and $25, GPT-5.6 Terra at $2.5 and $15, GPT-5.6 Luna at $1 and $6.
Published list prices, August 2026. The spread between the top tier and the budget tier is roughly 5× on both sides of the meter.

The rule: route by whether the work needs judgement

The useful distinction is not "hard" versus "easy". It is whether the task has one defensible answer that a competent person would agree on, or whether it needs weighing. Extraction, classification against a fixed taxonomy, reformatting and summarising to a template are the first kind. Reconciling a variance, drafting something a director will send, or deciding which of four explanations fits the numbers are the second.

JobTierWhy
Read a document, pull fixed fieldsBudget / fastOne correct answer; verifiable against the source
Classify against a known taxonomyBudget / fastClosed set, and mistakes are measurable
Draft a summary to a fixed templateEverydayNeeds fluency, not judgement
Explain a variance against contextReasoningWeighs several causes; a wrong answer is expensive
Anything a director will send uneditedReasoningThe cost of being wrong is reputational, not computational

In practice most volume in an operations system is the first kind, and most of the value is in the second. A system that sends everything to the top tier pays reasoning prices for extraction work; one that sends everything to the cheap tier produces confident nonsense on the decisions that matter. Neither is a model problem.

What actually moves the bill

Output tokens. Across every tier above, output costs five to six times input, and reasoning models are exactly the ones that produce more of it. A prompt that asks for working, alternatives and a recommendation costs multiples of one that asks for a decision and a one-line reason — from the same model, on the same task.

This is why the effort setting on Opus 5 and the Flash-Lite class of models matter more than the headline prices. The lever is not which vendor you pick; it is how much thinking you are paying for on work that does not need it. Ontilus sets out the three cost lines in detail in what an AI operating system actually costs to run.

A tiering strategy is worth more than a vendor negotiation. Moving the high-volume half of a workload down one tier typically changes a bill more than any discount you will be offered on the top one.

Why the top tier's price stopped being the interesting number

Opus 5 arriving at Opus 4.8's price is the pattern of the year: capability improves, list price holds or falls, and the tiers below absorb more of the work that used to need the top. For a buyer this has a specific consequence — architect so the tier is a configuration choice per job, not a decision baked into your code.

Groups that hard-coded a single model in 2025 have re-migrated repeatedly since. Groups that put a routing layer in front of it changed a config value. The second group spent less on tokens and considerably less on engineering.

Sources

  • Anthropic — Claude Opus 5 release, 24 July 2026; $5 / $25 per million tokens; effort setting.
  • OpenAI — GPT-5.6 tier pricing: Sol $5/$30, Terra $2.5/$15, Luna $1/$6.
  • Google — Gemini 3.7 Flash release, 13 August 2026, introductory pricing to 31 December 2026; Gemini 3.5 Flash-Lite positioning.

Dealing with this in your own group?

We answer scoping questions before there's a contract in sight — including the ones about cost and data handling.

Questions

Short answers,
in full.

The questions this article gets asked most, answered so each one stands on its own.

Talk to us
Which AI model should we use in 2026?

The question that lasts is which tier each job belongs to, not which model. Route work that has one defensible answer — extraction, classification, templated summarising — to a budget or fast tier, and work requiring judgement to a reasoning tier. Published August 2026 prices span roughly 5×, from about $1/$6 per million tokens at the budget end to $5/$30 at the top.

Why is our AI bill higher than expected?

Almost always output tokens. Output costs five to six times input across every current tier, and reasoning models generate more of it. Prompts that request working, alternatives and a recommendation cost several times one that requests a decision and a short reason, from the same model on the same task.

Is the most expensive model worth it?

For work where a wrong answer is expensive, yes; for high-volume extraction and classification, no. Claude Opus 5 launched in July 2026 at the same price as its predecessor, which is the year's pattern: capability improves at flat prices while lower tiers absorb work that previously needed the top. The saving comes from routing, not from choosing one model.

Should we commit to one AI vendor?

Architect so the model is a configuration choice per job rather than a decision embedded in application code. Four significant model releases arrived in under two months in 2026; teams that hard-coded a single model have re-migrated repeatedly, while teams with a routing layer changed a config value.