Comparison · verified August 21, 2026
Claude vs GPT-5.6 for Document Analysis
This is the one category where the two families are not close, and the reason is billing rather than intelligence. Anthropic includes the full 1M-token window at standard pricing; OpenAI meters long context at roughly double its headline rate. On document work, that is the comparison.
The second-order effect matters more than the first. Flat long-context pricing lets you put a whole document set in context and skip retrieval entirely — and retrieval is where most document pipelines actually go wrong. Below: the verified numbers, and a test using documents whose answers you already know.
The verified differences
What each one costs
Last verified August 21, 2026
USD per million tokens, standard rates, no discounts applied. Read the three caveats underneath before you put these numbers in a spreadsheet.
| Model | Tier | Input | Output | Notes |
|---|---|---|---|---|
| Claude Fable 5 | frontier | $10 | $50 | The frontier model. Twice the price of Opus 5 for the hardest work. |
| Claude Opus 5 | frontier | $5 | $25 | The default flagship. Thinking on by default. Fast mode available at $10/$50. |
| Claude Sonnet 5 | balanced | $2 | $10 | What most production agent fleets run on. The September 2026 increase to $3/$15 was cancelled. |
| Claude Haiku 4.5 | fast | $1 | $5 | High-volume and latency-sensitive work. |
| GPT-5.6 Sol | frontier | $5 | $30 | The flagship tier. Long-context requests meter at $10/$45. |
| GPT-5.6 Terra | balanced | $2 | $12 | The everyday tier, cut to this rate on July 30, 2026. Long-context meters at $4/$18. |
| GPT-5.6 Luna | fast | $0.20 | $1.2 | The cheapest frontier-family option on the market. Long-context meters at $0.40/$1.80. |
Anthropic changed tokenizers at Claude 4.7
Claude 4.7 and later — including Opus 5, Sonnet 5, and Fable 5 — use a newer tokenizer that produces roughly 30% more tokens for the same text than Sonnet 4.6 and earlier. A $2/MTok model on the new tokenizer is not directly comparable to a $2/MTok model on an older one, or to another vendor. Compare cost per task, not cost per token.
OpenAI meters long context separately
GPT-5.6 publishes a second, higher rate for long-context requests: Sol goes from $5/$30 to $10/$45, Terra from $2/$12 to $4/$18, Luna from $0.20/$1.20 to $0.40/$1.80. Anthropic includes the full 1M-token window at standard pricing on Claude 4.6 and later — a 900k-token request costs the same per token as a 9k one. If your workload is context-heavy, that difference is larger than the headline gap.
Two of Google’s current rates are introductory
Gemini 3.7 Flash and 3.6 Flash are priced at $0.75/$3.75 only through December 31, 2026. Both double to $1.50/$7.50 in 2027. If you are building a twelve-month cost model on Gemini Flash, model the 2027 number, not the one on the page today.
Caching and batching move the number more than model choice
A Claude cache hit costs 10% of the standard input price, and the Batch API takes 50% off both directions; the two stack. A workload that reuses a large system prompt can land below a nominally cheaper model that you are calling uncached. Model selection is usually the third-largest lever, after caching and after not sending the context at all.
What the evidence says, and who is saying it
Almost every performance number in public circulation traces back to a vendor running its own harness. That does not make the numbers useless, but the attribution belongs next to the claim.
Verifiable from published documentation
- Claude 4.6 and later include the full 1M-token context window at standard pricing, with caching and batch discounts applying across it.
- Claude raised the per-request media limit from 100 to 600 images or PDF pages in March 2026.
- GPT-5.6 publishes a separate long-context rate at roughly 2x the standard input price and 1.5x the output price.
These are documented product facts, not benchmark results. They are the ones that determine your bill on document work, which is why they lead.
What neither vendor measures for you
- Extraction accuracy on your document format — scanned contracts behave nothing like clean filings.
- Whether the model notices the absence of a clause, as opposed to summarising the clauses present.
- How each behaves on a document that contradicts itself, which real document sets do constantly.
The third one is the most under-tested and the most consequential. A model that silently picks one side of a contradiction produces a confident answer that is wrong in a way no summary metric will show you.
Test on documents where you already know the answer
Document analysis has the best evaluation property of any AI application: you can grade it exactly. Somebody in your organisation has already read these documents and written down what they say. Use that.
- Pick 15 documents your team has already reviewed by hand. Contracts with a completed redline, filings with a written summary, reports someone already extracted figures from. The human output is your answer key.
- Write the 10 questions you actually ask of every document. Termination clause, liability cap, auto-renewal, governing law — whatever your real checklist is. Not interesting questions; routine ones.
- Include three documents where the answer is "not present." This is the test most people skip and the one that separates the models. A model that invents a liability cap because the question implied there should be one is unusable for diligence, no matter how good its summaries are.
- Include one document that contradicts itself. Real document sets do. Check whether the model surfaces the conflict or silently resolves it. Silent resolution is the dangerous failure, because the output looks identical to a correct one.
- Verify every citation by hand on the first run. Open the document, find the passage, confirm it says what the model claims. Do this once properly and you will know whether you can trust the citations thereafter — which determines how much review the pipeline still needs.
- Price it whole-corpus and retrieval-based, both ways. On Claude, whole-corpus is often cheaper than building and running a retrieval layer once you count engineering time. On GPT-5.6, the long-context meter usually pushes the other way. Run the arithmetic rather than assuming.
Building a RAG pipeline covers the retrieval path when the corpus genuinely will not fit, and hallucination prevention covers the grounding setup that keeps "not present" answers honest.
The honest bottom line
Pick Claude for almost any serious document work. Flat pricing across a 1M-token window, passage-level citations, 600 pages per request, and a compliance surface built for regulated review. The cost advantage on long documents is structural rather than promotional, and it compounds on every request.
Pick GPT-5.6 if your documents are short enough that long-context metering never triggers, if you are doing heavy numerical work where code interpreter has a longer track record, or if you are already committed to the ecosystem and the volume does not justify a second integration.
Either way, do not skip the "not present" tests Every document pipeline that has embarrassed someone failed the same way: it answered a question the document did not answer. That is a property of how you evaluated, not of which vendor you chose, and no comparison page can fix it for you. See how to evaluate your agents.