v1.16.0 — The right agent, on the first ask
Ask for something specific and you should land on the specialist who does it — not on a generalist who then hands you along a turn later. This release rebuilds how an agent gets chosen.
It costs no extra model call per turn.
Your request is read before an agent is chosen
The system now works out what you're actually asking for — what you want produced, for what surface, and what ability that would take — and orders the roster by it.
Before, the roster was just listed, with the general-purpose agent pinned to the top on every turn. A specialist could only win if it held a tool that directly produced the thing you asked for — so every craft and domain specialist without a tool could never win a first turn. Copywriters, translators, tutors, strategists: unpickable by construction. That rule is gone. Being the right expert is enough.
Agents are judged on what they can do, not just what they say
The step that orders candidates used to read only an agent's written profile, never the tools it holds. So "compute the SHA-256 hash of this" was topped by a pet-persona agent scoring well on prose, while the agent holding the hashing tool didn't make the shortlist.
Every agent is now judged on its profile together with its tools:
| You ask for |
You used to get |
You now get |
| A SHA-256 hash |
a dog-persona agent |
the agent holding the hashing tool |
| A QR code for your booking page |
an image generator, which can't make a scannable code |
an agent with a real QR tool |
| Something unzipped |
an agent whose only "tool" was a follow-up suggestion |
an agent with archive tools |
| A passage read aloud |
whoever's description sounded warmest |
an agent that can actually speak |
| Numbers read off a screenshot |
an agent whose model can't see images at all |
one whose model can |
Your own shelf is actually ranked
If you work in a project or a team, its members are your shortlist — and that shortlist was being read, never ranked. On a 28-member project the scoring step matched zero of them, because it filtered on the wrong kind of membership; the model got an unordered wall of 28 names and had to do all the work itself.
It now matches all 28, in order of fit.
Work that needs two specialists now gets two
- "Find three peer-reviewed papers on CRISPR and read their abstracts aloud" used to go to a geneticist who can't speak. Now the researcher finds them and a speech-capable agent reads them.
- "Analyze this CSV, chart the trend, and draft an email to the team" used to collapse into one generalist step. Now it's a two-step plan, every time.
Fewer wrong turns
- A photo you mention is no longer a photo you asked to have made — "draft an Instagram caption for our sunset photo series" stays with your social specialist.
- A greeting no longer decides your afternoon: say "hey there", then "read the text out of this scanned receipt", and the second message re-routes to an agent that can do OCR.
- "Describe the plot of Hamlet" used to trip a hidden "make a chart" signal that removed the general-purpose agent from the running entirely.
Across the routing test set: the first-choice agent is right 58 times in 60, agents that should never have been picked now top the list zero times (down from three), and asking the same question twice gives a different answer 1.5% of the time instead of 6.1%.
Models, and a Max setting that finally does something
96 agents moved up to Grok 4.6 — 88 of them had been on a two-generation-old model. The Flash lane is now Gemini 3.7 Flash, same envelope and price. Web search, X search and code execution stay bound exactly as they were.
Reasoning effort "Max" now means something. Max and High were sending the identical instruction on every Grok agent; Grok 4.6 adds a genuinely deeper tier and Max now reaches it. Grok image agents also stopped showing temperature and top-P sliders that their model ignores.
The full release note lives in the in-app What's New drawer.
Keep the feedback coming in Feature Requests — tell us what to build next.