Best LLM for legal contract drafting in 2026: GPT-6.1 Sol

By LLMTest Lab · Nov 11, 2026 · 5 min read uselegaldraftingcost
On this page

On this page

  1. What we asked
  2. How we tested
  3. The cap decided most of the result
  4. What the judge liked and disliked
  5. Which model to pick
  6. Subscription vs API

The short answer: for drafting NDAs, SOWs, offer letters and clause redlines, GPT-6.1 Sol was the best pick in our test. It won 13 of the 30 decisive matchups, and 7 of its 8 drafts ran to the end inside a 2,800-token output cap. Claude Haiku 5.5 is the cheap pick at about $0.0035 per draft. None of this is legal advice, and a lawyer should read anything you sign.

Model Decisive wins Drafts that finished Avg cost per draft Price per 1M tokens (in/out)
GPT-6.1 Sol 13 7 of 8 $0.0185 $2.00 / $10.00
Claude Haiku 5.5 7 5 of 8 $0.0035 $0.10 / $0.50
Gemini 4 Carbon 6 4 of 8 $0.0294 $2.00 / $10.00
DeepSeek V4 Pro 3 4 of 8 $0.0080 $0.96 / $1.91
Claude Opus 5.5 1 1 of 8 $0.0556 $4.00 / $20.00

Prices come from OpenRouter's model list on the test date. "Finished" counts drafts that ended on their own (finish reason stop) with real contract text; a draft that stopped after 84 words of reasoning does not count.

What we asked

Eight tasks, each written the way a founder would ask without a lawyer on call:

  1. A mutual NDA between a Delaware and a Texas company, under 900 words.
  2. A fixed-fee Statement of Work with milestones, deemed acceptance and a risk note.
  3. A California offer letter where the founder also demands a non-compete, weekend invention assignment and unpaid overtime. The model has to refuse or fix those.
  4. A customer-side redline of a one-sided indemnity, one-month liability cap and warranty disclaimer.
  5. A limitation-of-liability clause with carve-outs and a data-breach super-cap that must not contradict each other.
  6. A 1099 contractor agreement for a New York designer whose hours and tools the startup wants to control.
  7. A GDPR processor addendum, with article numbers only where the model is sure.
  8. A review of a termination clause with built-in contradictions.

Several prompts hide a trap on purpose. The best answers catch it instead of obediently drafting.

How we tested

This post comes from one tournament run on 2026-11-11. We used the five models above, called through OpenRouter with exact refs openai/gpt-6.1-sol, anthropic/claude-haiku-5.5, google/gemini-4-carbon, deepseek/deepseek-v4-pro and anthropic/claude-opus-5.5. All five answered every one of the 8 prompts.

Every pair of models was compared on every prompt (80 matchups). anthropic/claude-sonnet-5.5 judged each pair twice with the order swapped, and a win only counts when both orders agree. That happened in 30 of the 80 matchups; the other 50 were ties or split verdicts. The rubric weighed legal soundness and trap-catching first, then completeness, drafting precision, invented citations (penalized heavily) and concision. The judge is a model, not a practicing lawyer, so treat the ranking as drafting quality, not legal correctness. Total spend was $3.94 of the $5 cap. The full outputs, judge reasoning and costs are saved with the post in our runs folder.

One caveat shapes everything below. We capped output at 2,800 tokens, and reasoning models spend part of that on thinking. That is a real condition for anyone calling an API with a tight max_tokens, but it is not the same as an uncapped run.

The cap decided most of the result

Opus 5.5 was cut off mid-sentence on 7 of 8 tasks. Its NDA stopped at "The Recipient is responsible", before compelled disclosure, term and governing law. Judges then had little to compare. Haiku 5.5 hit the cap on three prompts but, on the others, produced the longest drafts (2,162 words on the contractor agreement). DeepSeek V4 Pro's NDA was 84 words of visible text after burning all 2,800 tokens on reasoning, and it took 250 seconds on the last task.

GPT-6.1 Sol was the only model that finished nearly everything while staying under $0.03 a draft. Its one truncation was the 1099 contractor agreement.

What the judge liked and disliked

On the termination-clause review, Sol's list of problems got this verdict: "It catches the key traps: the contradictory termination rights, the penalty-like acceleration, the over-broad sole-remedy language, the missing cure period, and the survival gaps."

Haiku won the indemnity redline against Opus because it delivered all three clauses: "Response B's strengths: it handles all three clauses with detailed, usable drafting. It includes a mutual cap with a 12-month fees floor, a data super-cap, uncapped indemnities, a consequential-damages carve-out, and a warranty set with remedies and a disclaimer."

The California offer letter shows the trap-catching difference. Of Opus 5.5's draft, the judge said: "It states flatly that the position is exempt, which glosses over the trap: salary alone does not establish exemption, and the duties test also applies." Sol's letter ran through at-will terms, acceptance and signature. Telling a founder that "salaried" does not settle overtime is the part a template never does.

Which model to pick

  • Default: GPT-6.1 Sol. Best finish rate, best win count, a mid price. Use it for first drafts and clause reviews.
  • On a budget: Haiku 5.5. About a fifth of Sol's cost per draft. Give it a generous max_tokens, because it writes long, and read the tail of every output.
  • Not at a tight cap: Opus 5.5. We did not test it with a higher limit, so we cannot say how it does uncapped, and at $0.0556 per capped draft it is the most expensive way to get half a contract.
  • Gemini 4 Carbon and DeepSeek V4 Pro: usable, but each had four unfinished or empty drafts here.

Whatever you pick, ask for the "what I changed and why" note explicitly, check that it arrived, and have counsel review. For a related look at tone-sensitive writing, see our business email tests. If you cannot trust a model not to make up statutes, read our explainer on LLM hallucinations, and our guide on how to choose an LLM covers matching a model to the job.

Subscription vs API

API cost per capped draft (measured) Monthly plans
Claude Opus 5.5 $0.0556 Anthropic sells Free, Pro and Max chat plans
GPT-6.1 Sol $0.0185 OpenAI sells Free, Go, Plus and Pro ChatGPT plans

We do not quote plan prices here because they change often; check each vendor's plans page before you decide. The arithmetic is simple. Divide the plan price by the per-draft figure above to get the number of drafts at which a flat plan beats the API. Even at Opus's $0.0556, a flat plan of $20 would equal about 360 drafts a month, far above what a solo founder drafts, so the API is the cheaper route unless you also use the chat interface for other work.

Related articles

LLM hallucinations explained: causes, types, production fixes
Why LLMs hallucinate, the three error types, prompting patterns that reduce them, and the one production eval that catches failures before your users see them.
Best LLMs for business email writing in 2026: Claude leads
Best LLMs for business email writing in 2026: ranked by tone control, length discipline, and per-email API cost for solopreneurs.