This page compares two budget models on one job: pulling vendor, date, amount and line items from pasted invoices into a table. It tests them through their APIs, covers cost at volume, and ends with a fair way to try both on your own invoices.
Oct 14, 2026 · 9 min read
For a new invoice pipeline, start a pilot with GPT-6 Luna. If a Claude Haiku 4.5 system already runs well, test Luna beside it before you move anything.
Luna lists $0.10 per million input tokens and $0.50 per million output tokens, against $1 and $5 for Haiku 4.51, 2. On a planning profile of 3,000 input and 500 output tokens per invoice, taken from a published extraction-cost analysis9, 100,000 invoices cost about $55 on Luna and $550 on Haiku 4.5 at standard list prices1, 2. Both vendors take half off for batch jobs, so the tenfold gap stays the same.
Haiku 4.5 is also an older model. As checked on October 13, 2026, Anthropic lists it as legacy, recommends Haiku 5.5 instead, and lists its retirement as not sooner than October 15, 2026. That is a minimum and not a scheduled shutdown, but it is a poor starting point for something new2. Anthropic's current Haiku 5.5 is listed from the same $0.10 and $0.50 per million tokens, so the tenfold gap is against the older model and not against Anthropic's current one2.
Accuracy is the open question. The best invoice evidence covers Haiku 4.5 alone, and it shows scrambled totals are hard. A small test with both models favors Luna a little, but it was not about invoices3, 4. A sensible setup uses Luna with reasoning off for extraction, code to check that quantities, prices and totals add up, and a person or a stronger model for the invoices that fail.
You turn pasted invoices into rows for the ledger. Luna costs less and is a current model, and every total still gets checked before posting.
You build the pipeline. Both models can return a fixed table layout, so spend your effort on a hard test set and exact arithmetic checks in code.
You handle many suppliers and many layouts. At 1,000,000 invoices the list-price gap is $550 against $5,500, and your own mix of layouts decides accuracy.
You have a validated system. Run Luna beside it on the same invoices and move only if field and line-item accuracy hold up.
This page compares the two models through their API in one neutral setup, with the same pasted invoice text going to both.
The job is the one an accounts-payable team repeats all day. Take a pasted invoice and return the vendor, the invoice date, the total, the currency and every line item with its description, quantity, unit price and amount. Some invoices are clean. Others lose their columns when copied, carry OCR errors (mistakes from text-reading software), show several totals, or wrap a line item over two rows. Our sources are the official model documentation, one invoice dataset and independent tests1, 2, 4.
We left tools out of the specs table on purpose. Spreadsheet add-ons, file upload and document readers belong to the app around the model, so the same model behaves differently in a chat product, in the API or inside a workspace. The fair test sends identical text to both models. No public benchmark does that for these two models on invoices, which the evidence section covers.
The model facts that shape an extraction job. Tool and file features are left out, since those depend on the app around the model.
Figures from OpenAI and Anthropic documentation, checked October 13, 2026. A token is the small chunk of text a model reads and bills by, and 1M means one million of them. The two vendors count tokens differently, so treat any cross-model cost comparison as a close estimate.
Cost and lifecycle have clear answers. Accuracy does not, so some rows below say a test is required.
Luna looks stronger as the operating choice because of its lower price, current status and encouraging but limited tests. Haiku 4.5 has a narrower case, an existing system that is already validated where changing models adds more risk than it saves. Better-choice calls map to dimensions the sources evaluated, and rows on indirect tests say so.
Generic benchmarks do not cover your suppliers. Build a small set from your own invoices, give both models the same input, and score line items apart from header fields.
Pick three to five groups: clean invoices, text with collapsed columns, text damaged by scanning, invoices with several totals, and line items that wrap or run over several pages. Add credit notes, discounts and invoices where the printed total does not match the visible items.
Use one prompt and one output schema for both models, with identical text. Neither model gets a richer version, and you do not repair an output before it is scored. Run the important examples more than once to catch drift between runs.
Start with reasoning off or at its lowest setting for both, and run them through the API settings production will use. Chat products add their own prompts and tools, so API and chat results can differ.
Count exact matches on vendor, date, currency and grand total. Measure how many fields and line items are found and how many are correct, including description, quantity, unit price and amount. Also track valid schema rate, arithmetic checks that pass, invented line items and minutes of human correction per invoice. For a commercial decision, hide model names from the reviewer.
No public benchmark scores these two exact models on the same invoices. Here is what each related source helps judge.
Luna was released in September 2026, so its task-specific evidence is thin. Vendor figures are not independent, and the tests use different prompts and ways of damaging text. The verdict rests on cost, lifecycle and related tests, and it stays short of conclusive on accuracy.
Use one shared prompt first and tune per model only after the controlled test. The two patterns below differ in how they lay out the rules.
For GPT-6 Luna, start with reasoning set to none, a strict schema and short field definitions. Say exactly what to do when a value is missing, and define the total as the amount payable. Reasoning set to none is the lowest-cost place to start, and the schema fixes the table shape1, 8.
For Claude Haiku 4.5, separate the rules, the examples and the invoice text with XML-style tags, and still use structured output at the API level. Anthropic recommends clear steps and several varied examples when consistency matters, and extended thinking can stay off until a test shows it helps10.
A GPT-6 Luna prompt: a strict schema and an explicit rule for the total
Extract one invoice. Return only the supplied schema.
Never infer a missing value. Use null when it is absent.
Preserve line-item order.
"total" means the final amount payable, not the subtotal,
tax or a prior balance. For each line item return
description, quantity, unit price and line amount.
Set needs_review to true if the totals do not reconcile.A Claude Haiku 4.5 prompt: tagged sections for rules and invoice text
<rules>
Extract only visible values. Use null for missing fields.
Do not combine separate line items.
Select the final payable total.
</rules>
<invoice>
...pasted invoice text...
</invoice>Neither model is safe to post accounting data on its own. The useful question is where each adds cleanup work and what to change in the prompt or the workflow.
One question first: which kind of invoice job do you have. Then follow the branch that matches it.
Use this as a starting point and test on your own invoices before you commit
For a new deployment, pilot GPT-6 Luna. Its lower price and current status win unless your own test shows a clear accuracy gap1, 2. For a very clean, repetitive set of invoices, use Luna with reasoning off and batch processing.
If Haiku 4.5 already runs in production, run Luna beside it on the same invoices. If field and line-item accuracy match, move to Luna. If Haiku wins on important supplier layouts, keep it for now and plan a move to a current Anthropic model2.
For messy totals or wrapped line items there is no automatic winner, so pick the model that wins your supplier-specific test and add reconciliation and review. For strict table output either model works, and Luna has the cost edge unless measured accuracy says otherwise. For high-stakes accounting, require arithmetic checks in code and a human review whichever model you choose.
If you process a handful of invoices a month by hand, Playgram is not the right buy for this job. It earns its place when a team compares models, shares context and checks extraction samples together.
GPT-6 Luna is the practical winner for this budget extraction job, with cost and lifecycle on its side and accuracy still to be confirmed on your own invoices.
The evidence is incomplete. No public test runs both models on the same invoices. Haiku 4.5 has useful but mixed invoice results, and Luna has newer and mostly directional ones. Benchmarks use different prompts and harnesses, vendor figures are not independent, and prices and availability can change quickly2, 3, 4.
The safest final step is to test the shape of your own invoices, starting with your worst suppliers and scoring line items apart from header fields. A fair test needs the same setup for both models: the same invoice text, the same prompt and the same place to run them, so the result reflects the models and not the tool around them. In practice that is harder than it sounds, since most teams end up running one model in one app and the other in a different one, on two separate subscriptions, which tilts the comparison before the first answer comes back. The cleaner the setup, the more the difference you see is really GPT-6 Luna vs Claude Haiku 4.5, and which one happened to be easier to reach that day matters less.
Upgrade as needed, and only pay for what you actually use
Save ~17% with the annual plan
Pro
Perfect for small and medium teams
Unlimited users & infinite memory
Multi-LLM chats
Granular access control to models
EU data residency
Ultra
Best for large, growing teams
Unlimited users & infinite memory
Multi-LLM chats
Unlimited use of DeepSeek V4 Flash
Granular access control to models
Choose US or EU data residency
Enterprise
Get in touch
For organizations with advanced needs
Unlimited users & SSO
Priority Support
Unlimited use of DeepSeek V4 Flash
Granular access control to models
Choose US or EU data residency
30-days money back guarantee