xAI Grok models: two Fast modes, large contexts and team rate limits

Fast reasoning and non-reasoning are different endpoints

Use the complete grok-4-1-fast-reasoning identifier when testing multi-step analysis and grok-4-1-fast-non-reasoning for the direct-response alternative. A task that needs a short extraction or a simple rewrite is a useful first comparison: the additional reasoning mode should earn its place through a better accepted result, not through a longer answer. For a multi-step calculation, require intermediate values and independently verify the final number.

Both Fast entries have the same two-million-token context setting in the bootstrap catalog, and both are marked as vision capable. That is capacity parity, not evidence of identical accuracy, time to first token, or total bill. Put the same source material into each test, require the same answer format and record unsuccessful attempts. Do not compare a short non-reasoning response with an unconstrained reasoning essay and call the difference a benchmark.

Grok 4.6 is not the largest configured window here

The Grok 4.6 entry is configured for 500,000 tokens, one quarter of the Fast entries' two-million-token ceiling. Its reasoning flag is enabled, but that does not make it a drop-in capacity replacement for an oversized Fast conversation. Before switching an existing document thread, estimate how much source evidence and history will be included in the next request. A model switch should not silently discard the part of the document that contains the answer.

For current-events research, give Grok 4.6 several dated sources and ask it to separate established facts, disagreements and unresolved claims. Verify that quotations and dates exist in those sources. This is a test of grounded synthesis from supplied evidence, not a demonstration of live retrieval. A confident statement about today's event remains unverified if neither the input nor an executed search contains the supporting material.

xAI limits are per team and per model, not per browser tab

The current xAI rate-limit documentation describes requests per second and tokens per minute as separate dimensions. It says the per-second allowance is derived from the team's per-minute request budget divided by forty-eight, with a minimum of two. Team tiers scale with cumulative API spending. Those mechanics explain why an example quota from somebody else's account is not a dependable concurrency target for yours. Inspect your own team's model limits before planning throughput.

Prompt, completion and reasoning tokens all count toward the documented tokens-per-minute allowance. Cached prompt tokens also count, even when their billing rate is lower. A small number of very large requests can therefore exhaust the token allowance before reaching the request-count allowance. Conversely, a burst of tiny requests can reach the request limit while using relatively few tokens. Track both dimensions when diagnosing a 429 response rather than treating it as evidence that the model is unavailable.

For an interactive chat, wait for the active generation to finish and inspect the error before explicitly trying again. Opening multiple tabs is not a way to create a new team quota. For a separate batch integration, bound concurrency and follow provider retry guidance, but do not replay an interrupted request blindly when it may already have produced billable output. This app preserves interrupted chat output and requires an explicit retry; the guide does not recommend automatic duplicate generations to work around throttling.

A pricing gap should stay visible

Our historical Grok 4.6 estimate is $2 per million input tokens and $6 per million output tokens. The exact model's rate was not reconfirmed in the reviewed provider model page, so the table labels it unconfirmed. Neither Fast variant has a verified price in this snapshot. We do not substitute a newer Grok release's rate or treat an absent row as a free endpoint. Check the exact identifier in your provider account before starting a paid experiment.

At the unconfirmed 4.6 estimate, 100,000 input tokens and 5,000 output tokens would cost $0.23 before other charges. That is a sensitivity example only, not a quote: if the true rate differs, the total differs proportionally. Repeatedly resending a large dossier multiplies its input cost even when each new question is short. Keep a record of the provider's actual usage and bill rather than forecasting from the visible length of the answer alone.

Search and context solve different problems

None of these three bundled xAI entries enables native search in this catalog. That statement concerns this app configuration, not the full set of capabilities xAI may offer elsewhere. If your question needs fresh information, supply credible sources or configure an available search path and verify that it actually runs. A two-million-token window can hold a large archive, but it cannot make yesterday's archive contain today's announcement.

Choose a Fast mode when its tested output and account economics fit the job; consider 4.6 only after checking its smaller window and exact availability. For a long archive, plant verifiable references in several distant sections and demand citations to the supplied section names. Check omissions as carefully as fabricated details. These practical tests are more useful than creating separate public pages for the same Grok Fast specifications with a single reasoning flag changed.

Concrete xAI model checks

Suggested evaluation protocols, not executed benchmark results.