OpenAI and Anthropic both cut prices last week, and the spread between the cheapest and most expensive models on the market is now about a hundredfold. The right choice depends far more on the task than on the leaderboard.
Picture a distributor in the Fox Valley that gets four hundred order emails a day. Somebody wires them into an AI model to pull out five fields: purchase order number, part number, quantity, ship date, and delivery address. It works on the first try. The team picks the most capable model they have heard of, because that seems like the safe choice, and moves on.
Six months later that workflow is still running on a model built to write software, reason through legal arguments, and run multi-hour research tasks. It is reading purchase orders. It does the job well. So would a model costing a small fraction as much, running faster, and, for a job this narrow, likely making about the same number of mistakes.
That is the most common model selection error we see, and it rarely looks like an error. The biggest model almost never fails the task. It just rarely turns out to be the right fit for it.
Last week reset the menu
On September 22, Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, 20% below Opus 5, the model it replaces. OpenAI released GPT-6 Sol and GPT-6 Luna the same day. Sol costs $2 per million input tokens and $10 per million output. Luna costs $0.10 and $0.50.
Lay out the full ladders and the range is striking. Anthropic’s Claude lineup runs from Haiku 4.5 at $1 input and $5 output up to Fable 5.1 at $10 and $50. OpenAI’s runs from Luna up to GPT-6 Astra, its top model. Google’s Gemini 3.8 Flash sits at $0.75 input and $3.75 output under introductory pricing through December 31, rising to $1.50 and $7.50 in January.
Then there are the open-weight models you can download and run on your own hardware: Meta’s Llama 4, Mistral’s lineup, Alibaba’s Qwen, DeepSeek, and others. Licenses matter here. Several ship under permissive Apache 2.0 or MIT terms, while Llama 4 uses Meta’s own community license. Running your own model also trades per-token bills for hardware, hosting, and the engineering time to keep it patched and monitored. That is a real cost, and it belongs in the comparison.
From Luna’s input price to Fable’s, that is a hundredfold spread. No single model is the right answer across that range, and the vendors know it. OpenAI positions Luna for summarization, extraction, and straightforward questions. Anthropic’s Haiku tier is priced for exactly that kind of high-volume classification and routing work. Every major lab now sells a ladder, because customers should climb it only as far as the task requires.
Where each one tends to earn its place
A fair comparison starts with real strengths. OpenAI remains the most widely deployed: in a16z’s most recent CIO survey, 78% of enterprises were running OpenAI models in production. Anthropic posted the largest share gain of any frontier lab, reaching 44% production use, and CIOs credited Claude with strong gains in coding, writing, reasoning, and analytical work. Google competes hardest on price at the fast tier. Open-weight models win when data cannot leave your building, or when a narrow task justifies tuning your own copy.
Some buyers, particularly in regulated work, also weigh where a model was developed and where it is hosted. That is a legitimate governance question, and it deserves a written answer before a pilot starts rather than after.
None of those strengths settle your decision. Benchmarks measure general capability, and your workflow is specific.
Sort the work before you pick the model
Most business AI work falls into two buckets.
The first is routine: pulling fields from documents, classifying tickets, tagging records, summarizing a call, drafting a standard reply. The instructions are clear, the right answer is checkable, and volume is high. Small and mid-tier models handle this well.
The second is judgment: multi-step reasoning, agentic work that plans and uses tools, complex code changes, analysis where the model has to weigh conflicting information. This is where frontier models like Claude Opus 5.5, Fable 5.1, and GPT-6 Astra justify their price.
In most workflows we see, the routine bucket is far larger. The architecture that follows is simple. Send routine work to a small model, and escalate to a larger one only when the task is genuinely hard or the small model signals low confidence.
Test on your work, not a leaderboard
Pick three candidates across at least two vendors: one small, one mid-tier, one frontier. Run all three against a set of real examples from the workflow, including the messy ones. Measure three things: accuracy against your answer key, cost per completed task, and time to response.
Cost per completed task is the number that matters, not price per token. A cheaper model can end up costing more when it needs extra calls, produces more tokens, or fails often enough to require retries. Sometimes the frontier model wins on total cost. Sometimes the small one wins by a mile. You only learn which by running the test.
Here is what a result can look like. On a batch of two hundred real order emails, a frontier model and a small model might land at nearly the same field-level accuracy, with the small one answering in a fraction of the time. The frontier model pulls ahead only on the thirty or so emails where a customer changed the order partway through the thread. That result does not argue for one model. It argues for two: the small one as the default, and the frontier one for the messy thirty.
Build so you can switch
Model menus change fast. Anthropic replaced Opus 5 after about two months. OpenAI replaced its mid and low tiers at half the price in the same week. In June, Anthropic suspended access to its Fable 5 and Mythos 5 models for about three weeks to comply with U.S. Commerce Department export controls, restoring access July 1 after the controls were lifted.
The lesson is structural. Put a thin routing layer between your application and any one vendor, keep your test set current, and rerun it whenever a new model ships. A switch then becomes a configuration change backed by evidence rather than a rebuild.
What this means in Wisconsin
The 2025-2026 Wisconsin Manufacturing Report found that, for the first time, a majority of the state’s manufacturers are embracing AI. It also found they struggle to get started, struggle to earn a return, and that larger companies report more difficulty implementing it than smaller ones. Meanwhile, 61% said AI will be another piece of technology among many, against 30% who expect it to fundamentally change businesses like theirs.
That 61% has the right instinct. AI treated as a tool gets matched to a job. A plant scheduling assistant, a quote intake workflow, or a quality report summarizer rarely needs the most powerful model on the market, and choosing a smaller one can be the difference between a pilot that pays back and one that gets quietly shut off.
The honest tradeoff
Starting with a frontier model is reasonable. It tells you fast whether the task is solvable at all. The mistake is stopping there. Once the workflow works, step down the ladder one tier at a time and watch the answer key. Stop when accuracy drops. Most teams are surprised how far down they get.
Small models also ask more of you. They need clearer instructions, tighter inputs, and better tests. That discipline pays off regardless of which model you end up running.
We help companies in Madison, across Wisconsin, and throughout the continental U.S. match AI models to the work in front of them. That means building the test sets, running the comparisons across Claude, GPT, Gemini, and open-weight options, and designing the routing so the next model change is a decision instead of a scramble. Bring us the workflow and we will tell you honestly which model it needs, including when the answer is the cheap one.
