Grok models chooser: 4.5, 4.6, and when to switch

Tuesday standup, 9:12 a.m. Marcus still has a 40-message Grok thread open from last night, all of it spent rewriting a 180-word status note for the warehouse huddle. Someone says just use 4.6 for everything. The thread is still arguing about commas in sentence three. The status note is still in Drafts, and the huddle starts in 18 minutes.

This is Part 9 of the Grok series, still on the product map. Part 8 named the four doors: chat apps, Grok Build, Imagine, and the Console. This part sits inside those doors. A model name is not a product. A version number is not a plan. If you pick 4.6 because it is the biggest numeral on the menu, you will spend a frontier pass on work that needed a rewrite, and you will still have the wrong still for the poster.

What you’ll learn

• How to separate product names (chat, Build, Imagine, API) from model names (4.6, 4.5, Imagine family, API ids)
• What Grok 4.6 and Grok 4.5 mean in plain English as of writing, without a scoreboard
• Why Imagine models are not “Grok 4.6 with pictures”
• Why the label in the app and the model string in JSON are different objects
• When to switch models, and when switching is stalling
• A four-task week for Marcus that you can copy onto your own calendar

SKUs, in-app labels, and list prices move. Confirm live names and meters on the Grok 4.6 docs, the Grok 4.5 launch post, and x.ai/api the week you set a team default. Do not freeze this page into policy.

Names vs products

Three labels get mashed in Slack. Grok is the product family. A surface is where you open it (chat, Build, Imagine, Console). A model is the engine behind a request. “Use Grok 4.6” is incomplete until you say where. 4.6 in grok.com is a chat turn. 4.6 in Grok Build is an agent with a folder. 4.6 on the API is a line item on a token bill. Same family name, three review habits, three failure modes. If company vs product vs model still blurs, go back to Part 1 before you pick a numeral.

Chat apps also expose faster or cheaper modes. Those names change more often than the frontier id. Treat them as a bucket called everyday chat, not as a second religion. If the UI says Auto, Fast, or something new on Thursday, pick it for hallway questions and short rewrites. Save the frontier label for work that is long, tool-heavy, or worth the wait.

xAI builds Grok. Some marketing pages currently say SpaceXAI. That is branding, not a new model. Do not wait for a merger story before you choose 4.5 vs 4.6. Choose from the job in front of you.

4.6 and 4.5 in plain English

As of writing, Grok 4.6 is the frontier text and agent model. Vendor docs describe text plus image input, text output, a 500,000 token context window, and a knowledge cutoff of February 1, 2026. The API id is grok-4.6. It is the model Grok Build is built around. If you are asking an agent to live in a repo for a long stretch, 4.6 is the default to try, then review.

Grok 4.5 is the previous frontier. The launch post is dated July 16, 2026. The API id is grok-4.5. xAI’s own post claimed about 80 tokens per second and listed $2 per million input tokens and $6 per million output tokens. The 4.6 docs page, as of writing, prints the same list pair. That is a vendor list price, not your invoice. Confirm live pricing on x.ai/api and in the Console before you budget a bot. 4.5 is still a strong coding, agent, and office model. It is not “the old one, therefore useless.” It is the model you might still pin in an IDE or an API while 4.6 is the new default in Build.

There is also grok-build-0.1 on the API: a coding model from May 2026, trained for agentic coding, happiest inside a coding agent such as Grok Build or Cursor. Do not drop it into a chat window and expect a folder agent. The agent environment is the product. The id is the engine.

This post does not reprint competitor scoreboards. Launch posts love them. They go stale in a week, they depend on the test setup, and they do not tell you whether Marcus’s 180-word status note needed any of that. If you need the vendor’s own claims, read the 4.5 post and the 4.6 docs and treat the numbers as marketing until you run your own task twice.

Context windows are a similar trap. 500,000 tokens is a lot of room. It is not a reason to paste the warehouse handbook, last month’s huddle notes, and three CSVs into the same 40-message thread. Long context is for work that is long. A status note is 180 words. The model that finishes it in two turns beats the model that can theoretically swallow a novel.

Imagine family is not a chat model

Imagine has its own family. Consumer Quality Mode, as of writing, is Imagine Image 2.0 (vendor GA: August 7, 2026). On the API you will see ids such as grok-imagine-image-2.0, plus a video family for short clips. Those strings do not belong in a chat model field you copied from a 4.6 snippet. They also do not belong in the “just use 4.6 for everything” sentence.

A chat model can describe a poster. An Imagine model has to render letters. That is why Marcus’s teammate who asked 4.6 to “make the flyer” got a confident layout in prose and still had nothing to print. Open the studio. Put the words in the brief. Inspect the still. The Imagine parts of this series start later. For the chooser, one line: pixels go to the Imagine family. Words that will stay words stay on a text model.

Video is the same split, with a cost kicker. Clips are measured in seconds, not in feature-film minutes. If someone says “have 4.6 make a trailer,” they have mixed three products (chat, Imagine video, maybe the API) into one wish. Name the still first. Then decide if motion is worth the meter.

API ids vs in-app labels

The app can say “Grok 4.6.” Your code cannot. The JSON field wants a slug: grok-4.6, grok-4.5, grok-imagine-image-2.0, whatever the live list prints. Spaces, marketing punctuation, and “the new one” will 400. Copy the id from x.ai/api or the model docs, not from a screenshot of the chat picker.

In-app labels also hide routing. Fast modes, Auto, and similar toggles may send a different engine than the frontier name on the same screen. That is fine for chat. It is a bug if you are trying to reproduce a result in a script. For apps, pin an id. Log it. When the reply looks “dumber than yesterday,” check the id before you rewrite the prompt for the fourth time.

Build adds a third label layer. The CLI may say it is on 4.6 while a config file or an IDE plugin still has grok-4.5 pinned. If a review comment says “Build got worse,” ask which id ran, on which surface, against which folder. Version fights in standup are often config fights with better posture.

When to switch (and when not to)

Switch when the job changed, not when you are bored. A rewrite that is still a rewrite after 40 messages does not need a bigger numeral. It needs a constraint: 180 words, no new facts, keep the 14-aisle count, stop. Switching to 4.6 there is how you spend a frontier turn on Oxford commas.

Do not switch because a coworker used the word “agentic.” Do not switch because a launch post used the word “frontier.” Do switch when you can point at a failure the current model keeps repeating: missed a type error after three tries, cannot keep a 12-section outline consistent, keeps inventing a column named rev_final. Then change one thing: model, or surface, or prompt constraint. Changing all three at once teaches you nothing.

Speed vs quality is a real trade, and the live numbers move. 4.5’s launch post leaned on tokens per second. 4.6’s docs lean on long agent work and a large context window. For a 180-word note, latency you feel in the huddle matters more than a 500,000 token ceiling you will never fill. For a two-hour Build session on a gnarly parser, the opposite is true. Match the meter to the job, then stop talking about the meter.

Worked week (Marcus, 4 tasks)

Marcus owns warehouse analytics for 14 aisles. His week is not “use Grok.” It is four tickets. Watch where the model choice sits.

Monday: the 180-word huddle note

Facts he already has: aisle 7 is short 40 cases of SKU WB-210, the Friday truck landed at 6:10 a.m., overtime was 11 hours. He pastes those facts and says: 180 words, no new numbers, no pep. Everyday chat. Two turns. He reads it out loud. He fixes “landed early” to “landed 6:10 a.m.” He sends. The 40-message version from Sunday night was a model-worship problem. The note never needed 4.6. It needed the facts in the prompt and a word cap.

Tuesday: flaky test in pick_wave.py

A pytest fails on 1 of 20 runs when wave_id wraps past 9999. This is a folder job. Marcus opens Grok Build on a sandbox clone, on 4.6 as of writing, and asks where the wrap is handled. He does not paste 800 lines into grok.com. Build proposes a one-function change. He runs the test 20 times. He keeps the diff. Surface: Build. Model: 4.6. If Build were unavailable, chat with 4.6 plus a pasted function is a fallback, not a lifestyle.

Wednesday: aisle map for the safety board

Ops wants an 11×17 still: 14 aisles, “KEEP CLEAR at dock 2” in real letters, no cinematic fog. Imagine family. Marcus puts the words in the brief. He rejects two stills that spell KEEP CLEAR. He does not ask 4.6 to “be better at design.” Different product.

Thursday: internal FAQ stub for 40 stores

A small bot that answers “what is a pick wave” from a six-paragraph README. That is an API pin: grok-4.6 (or grok-4.5 if that is what you budgeted), key in an env var, spend cap in the Console. It is not SuperGrok on Marcus’s phone answering store managers at 2 a.m. Chat-plan logins do not become multi-store apps because you forwarded the thread.

Four tasks, four starts, one week, zero standup arguments about numerals. The 40-message thread was the tell. When you are negotiating commas, you are not short on intelligence. You are short on a stop rule.

Common mistakes

• Making 4.6 the personality of the team. Default the job, then the model.
• Pasting an in-app label into JSON. Use the API id from the docs.
• Asking a text model to print a poster. Open Imagine.
• Treating list prices ($2 / $6 per million on the 4.5 post and the 4.6 docs, as of writing) as your finance system. Confirm live.
• Switching models mid-thread without restating the constraint. The new engine inherits the mess.
• Confusing grok-build-0.1 with a chat personality. It wants a coding agent around it.
• Using a 500,000 token window as permission to dump the whole shared drive.

How to practice / next step

Write your next four tasks in a column: job, surface (from Part 8), start model, switch if. Run one day on that card. If a task took 40 messages, cut the transcript to the facts and the constraint and try two turns on the cheaper chat mode. Time both. Keep the faster one that you would still send.

Next in this series: Grok in Outlook, Word, Excel, and Workspace (light). That is still product map, now with a work account and an add-in. Model choice follows you into the inbox. The dangerous part is the paste, not the numeral. Find it from the Grok series hub when it is up. If you still need the doors themselves, stay on Part 8 until you can name chat vs Build vs Imagine vs API without looking down.

Quick recap

• Product, surface, and model are three words. Use all three
• As of writing, 4.6 is the frontier agent/text default (API grok-4.6, 500k context, cutoff 2026-02-01)
• 4.5 is still in play (API grok-4.5). Confirm price and speed live; do not worship the launch post
• Imagine family is pixels. Chat family is words
• Pin API ids in code. Do not paste the pretty label
• Switch when the job changed or a failure repeats. Do not switch to end an argument about commas
• Everyday chat is allowed to stay everyday. That is how Marcus’s huddle note ships in 18 minutes

Sources

Research and further reading used for this article:

• xAI docs: Grok 4.6 (id, context window, cutoff, modalities, list price to re-check)
• xAI: Introducing Grok 4.5 (July 16, 2026 launch claims, including list price and speed; vendor’s own numbers)
• xAI API (live model list and pricing entry; confirm the week you pin an id)
• xAI: Grok Build (4.6 as the agent default in the coding CLI)
• xAI: Imagine Image 2.0 (Imagine family is a separate product)
• xAI docs: Imagine capabilities (API ids and image/video paths)
• xAI Console (where API ids become a bill)
• Analytics Made Simple: Learn (related paths on this site)

More definitions in Key Terms →

Merch on Etsy

Open store →

Dashboard to Excel Ceramic Coffee Mug View on Etsy → Naive Bayes Mouse Pad, Data Science Mouse Pad, Analytics Mouse Pad, Bayesian Mouse Pad View on Etsy → You're On Mute White Ceramic Coffee Mug, Data Science Mug, Data Mug, Analytics Mug, Statistics Mug, Programming Mug, Gift Mug, Coffee Mug View on Etsy → Mute Mouse Pad, Data Science Mouse Pad, Analytics Mouse Pad, You're On Mute View on Etsy →