api · research
The cheapest API for structured extraction
A dated cost comparison for turning 1,000 documents into JSON, with input, output and local-compute arithmetic shown.
· 2 min read
Evidence: List prices checked 7 August 2026; cost calculation only, with no accuracy benchmark
Review summary
- Verdict
- For the priced workload, GPT-5.6 Luna has the lowest hosted token bill; quality still needs a task-specific evaluation.
- Best for
- Builders estimating the first model bill for schema-constrained extraction
- Not for
- Choosing a production model without measuring extraction accuracy
“Cheapest API” is incomplete until the workload has a size. Here is one concrete estimate: 1,000 documents, averaging 2,000 input tokens each, producing 200 output tokens of JSON each. That is 2 million input tokens and 0.2 million output tokens. Prices and model names were checked on 7 August 2026 and will change.
This is a price calculation, not a quality benchmark. Token counts differ by tokenizer and content, retries cost money, cached-input pricing is excluded, and a malformed extraction is not a useful result.
The arithmetic
Using published standard API rates:
| Option | Input calculation | Output calculation | Token bill |
|---|---|---|---|
| OpenAI GPT-5.6 Luna | 2 × $0.20 | 0.2 × $1.20 | $0.64 |
| Anthropic Claude Haiku 4.5 | 2 × $1.00 | 0.2 × $5.00 | $3.00 |
| Local through Ollama | no API token charge | no API token charge | $0 API bill |
The local row is not “free.” It moves the bill to hardware, energy and engineering time. If a machine draws 70 extra watts for one hour, the run uses 0.07 kWh; multiply that by your electricity rate and add the machine’s amortised cost. For a small batch, setup time is often larger than every token charge in the table.
The verdict
GPT-5.6 Luna is the lowest published hosted token bill among these named options for this input and output mix. That makes it the first candidate to evaluate, not the automatic production choice. OpenAI’s current model comparison lists structured-output support, and Ollama supports supplying a JSON schema for local output. A schema constrains shape; it does not guarantee that the extracted values are true.
Build a 100-document gold set and score exact fields. Track valid JSON rate, field-level precision and recall, latency, retries, and human correction minutes. Then calculate:
cost per accepted document = (API + retries + review labour) / accepted documents
A $0.64 run with 20% silent field errors can be much more expensive than a $3 run that needs almost no correction.
When to revisit the table
Recalculate whenever a provider changes a price, your average document grows, output becomes more verbose, or you add a second-pass verifier. Publish the exact date and workload next to the number; otherwise a price comparison becomes stale while still looking precise.
Sources: OpenAI model comparison, Anthropic model pricing, May 2026 (PDF), and Ollama structured outputs.
Product links
Check the current product, price, plan and regional availability on the seller's page.
- Open the OpenAI platform — Official OpenAI link
- Open the Anthropic Console — Official Anthropic link
- Visit Ollama — Official Ollama link
Take it further
- The economics of automation — Why one job gets automated and the near-identical one next to it does not. (3 lessons, 65 min, 3 free)
- The eval you actually need — Stop reading leaderboards. Build the harness that measures your task, on your inputs. (4 lessons, 8 min, 4 free)
- Coffee break: the summariser you can prove is wrong — Twenty minutes to build it, twenty more to find out where it fails. The second half is the point. (2 lessons, 40 min, 2 free)