Open-Weight Models Are the New Budget Tier: A Price Map for AI Features

If you are building an AI feature and pricing it against frontier-model API rates, you are probably overpaying by an order of magnitude for a large share of your workload. Open-weight models served over API, DeepSeek, Qwen, Kimi, gpt-oss and friends, now price 10 to 50 times below frontier flagships, and for mechanical tasks like classification, extraction, and routine summarization they are frequently good enough. The catch is the word "frequently": the only way to know is a small eval on your own data, because vendor parity claims are marketing until you test them. This week's Hacker News front page carried yet another "frontier-level results at a third of the cost" pitch, and CIO Dive was reporting enterprises straining under AI tool costs, the demand for a cheap tier is real. Here is the price map from the providers' own pages, a worked monthly bill, and a routing rule.

The price map

All prices are per million tokens from the providers' published pricing pages, checked July 24, 2026. One health warning before the table: providers use different tokenizers, Anthropic notes its newer models produce roughly 30% more tokens for the same text than its older ones, so per-token prices across vendors are directionally comparable, not exact.

Discounts change these numbers materially and are worth knowing before you compare: Anthropic's Batch API halves both directions for non-urgent work and cache reads cost a tenth of input price; OpenAI discounts cached input around 90% with batch around half; DeepSeek charges $0.0028, essentially free, for cache-hit input on Flash.

What a real feature costs per month

Our arithmetic, using the list prices above. Take a support-ticket summarizer handling 30,000 tickets a month at about 1,500 input and 300 output tokens each: 45M input, 9M output tokens monthly.

• Claude Fable 5: $450 + $450 = $900
• Claude Opus 4.8: $225 + $225 = $450
• GPT-5.6-terra: $112.50 + $135 = $247.50
• Claude Haiku 4.5: $45 + $45 = $90
• gpt-oss-120B (Together): $6.75 + $5.40 = $12.15
• DeepSeek V4 Flash (first-party): $6.30 + $2.52 = $8.82

That is a 100x spread on the identical workload. If the cheap model summarizes tickets acceptably, a testable question, the frontier bill is a subsidy you are paying for capability the task never uses. Flip it around and the same table says something founders miss: even the expensive tier is cheap in absolute terms for moderate volumes. The point is not "always go cheap"; it is that the cost of checking is an afternoon, against a 100x recurring difference.

Same model, different landlord

A wrinkle worth exploiting: open weights mean anyone can serve the model, and hosts price very differently. DeepSeek V4 Pro costs $0.435/$0.87 on DeepSeek's own API but $1.74/$3.48 on Together, a 4x difference for the same weights. The premium is not irrational: you are choosing jurisdiction, data-handling terms, latency, and reliability along with the price, and for customer data those terms can matter more than the invoice. The practical move is to treat hosting as a market: read the data-use policy of the cheap host as carefully as the price list of the expensive one, and remember that with open weights the exit is real, if a host raises prices or degrades, the same model runs elsewhere, which is a portability no frontier API offers.

A routing rule instead of a model religion

Our framework: sort each AI-touching task in your product into three tiers, and buy each tier separately.

1. Mechanical tasks, classify, extract fields, tag, dedupe, route. Start with a small open-weight model. Build an eval of 100 labeled examples from your real data; if the cheap model clears your accuracy bar, the decision is made. This is where most volume lives and where the 50x savings are usually free.
2. Drafting and summarizing, ticket summaries, first-draft emails, descriptions. Mid-tier open-weight or budget frontier (Haiku 4.5, GPT-5.4-nano) both plausibly clear the bar; test both, and let batch pricing decide ties for non-urgent volume.
3. Agentic and high-stakes work, multi-step tool use, code generation on real systems, anything customer-visible where an error costs trust. Pay for the frontier tier (Sonnet 5, Opus 4.8, GPT-5.6). This is precisely where "equivalent at a third of the cost" claims most often collapse in practice, and where your eval must be hardest before you downgrade.

Then route per task, not per product. A product that uses Fable-class reasoning for its planning step and a $0.14 model for its extraction step is not being cheap; it is being priced correctly. This is the product-side twin of auditing your own AI tool spend, same discipline, applied to COGS instead of overhead, and it is a large part of why the one-person software company pencils out at all in 2026. The same instinct that says pick boring technology applies here with a twist: the boring choice for a mechanical task is now the cheap open-weight model with an eval proving it works.

Act on it this week: pick your single highest-volume AI call, spend one afternoon building a 50-example eval, run the cheapest plausible model against it, and escalate tiers only on failure. And diarize one date, Sonnet 5's introductory $2/$10 pricing ends August 31, 2026, after which it moves to $3/$15. If your margins depend on the intro rate, you want to know that before your September invoice does.

For the build-side view of these choices, see the boring stack; for keeping hosted-model spend honest, the AI subscription audit.