What's new in goHeather in October 2026?

The October 2026 update to goHeather's legal AI software went live on October 8. The August update rebuilt the product around chat. This one is smaller, and it is mostly about the things that make a hundred contract reviews a month less error-prone: defaults you set once, a choice of AI model, a review that keeps going when a provider is busy, and tighter security.

Set your side and review depth once for every contract review

Every AI contract review in goHeather starts with a setup card. goHeather reads the document, names both parties, picks the one it thinks is you, and chooses a review depth. Until now it did that fresh every time, and if it picked the wrong party you had to catch it on the card or you got findings written for the other side.

Settings now has a Review defaults section with two fields. Default party is a short free-text description of who you usually are in a contract, something like "the customer buying software" or "the general contractor". When a review starts, the party that best fits that description is picked for you. Default review depth sets where the four-level depth setting starts, Existential, Issues, Warnings or Everything, instead of always opening on Issues.

You can still change both on any single review.

Pick the AI model for each contract chat

The chat composer now has a model picker. Six models are listed, each with a one-line note on what it is suited to. GPT-6 Astra for complex tasks and deep contract analysis. GPT-6.1 Sol, the default, for medium tasks, drafting and review. GPT-5.6 Terra for lightweight tasks and quick summaries. Claude Opus 5.5 for complex reasoning and careful contract analysis. Claude Sonnet 5.5 for everyday drafting, review and clear writing. Meta Muse 1.3 as an alternative model for drafting and contract analysis.

The choice applies to chat replies. Two independent leaderboards are worth a look before you pick.

Legal Benchmarks scores models on real contract work, marked by practising lawyers, and a task only passes if every criterion is met on two separate runs. On its October 2026 contract workflows leaderboard, Claude Opus 5.5 and Meta's Muse 1.3 (it appears there as Muse Spark 1.3) share first place with a 48.4% task pass rate. Claude Sonnet 5.5 is seventh at 29.0%, GPT-6 Astra and GPT-6.1 Sol are tied ninth at 25.8%, and GPT-5.6 Terra is twentieth at 16.1%. On the same site's data extraction leaderboard, Claude Opus 5.5 is first at 76.7% and GPT-6 Astra is third at 56.7%. Those pass rates are low across the board, and that is the useful finding: even the best model fails about half of real contract tasks when every criterion has to be met, so a review still needs a person at the end of it.

Vals AI's LegalBench leaderboard tests legal reasoning rather than contract output, across six categories such as issue-spotting, rule-application and interpretation. Its scores are bunched tightly. The top twenty-five models all sit between about 85% and 89% accuracy, and of goHeather's six only GPT-5.6 Terra appears, at 85.1% with one of the fastest response times on the list. Vals' own note is that a model's fitness depends on which legal task you put in front of it.

The practical reading: for careful contract analysis and for pulling facts out of a document, start with Claude Opus 5.5 or Muse 1.3. For quick summaries and short answers where speed matters more, GPT-5.6 Terra is fast and cheap. Then test on your own contracts, because the model that scores best is not always the one that writes a clause the way your team would.

And do not get attached to any of them. We will keep adding the newest and best models to goHeather as they come out, and the model leading a leaderboard this month is unlikely to be the one leading it in six months.

goHeather keeps going when an AI provider is overloaded

While we were testing reviews of several contracts at once, a request failed with a "servers are currently overloaded" error. The error came from the AI provider's side, though you would not know that from the screen. Now, when a provider returns that error, goHeather retries the request on another model automatically instead of asking you to try again later. If it works, you will not notice.

A fix for review findings that "could not be found"

Some reviews reported a handful of suggested changes as "can't find in text" when the original clause was plainly in the contract. That matching bug is fixed in this update. Alongside it, you can now reset the changes you have applied and see the full set of suggestions again, each with its High, Medium or Low risk level, rather than losing sight of the ones you already accepted.

Security was improved

This update also includes security improvements across the web app, the API and the document editor, following an audit of the whole platform.

What did not change in goHeather's legal AI

Reviews and drafts still take minutes, not seconds, and still run server-side so you can close the tab. Findings in the AI contract review workflow are still proposals until a person decides them, and the chat still will not move on while findings are undecided. That is by design.

goHeather is an AI, not a lawyer, and it can make mistakes. This is legal information, not legal advice; consult a lawyer for legal advice.