Grok 4.1
Grok 4.1 made Grok friendlier and more accurate, but also more of a people-pleaser. xAI released it to all Grok users on 17 November 2025, after quietly testing it on live traffic from 1 to 14 November. It briefly took the top spot on the LMArena text leaderboard, where people vote on which answer they prefer, and xAI said it cut hallucinations on real questions from 12.09% to 4.22%.
Its own model card also showed it agreed with users when they were wrong far more often than Grok 4 did. Alongside it came Grok 4.1 Fast, a cheap API model with a 2-million-token context. SpaceXAI retired Grok 4.1 Fast from the API on 15 May 2026, and calls now go to Grok 4.3.
Grok 4.1 was a big step up in chat quality, and Grok 4.1 Fast was one of the best-value API models of late 2025. Both are now retired.
What it did well:
- People liked its answers. It reached #1 on LMArena's text leaderboard at 1483 Elo, and xAI said users preferred it to the previous Grok 64.78% of the time in blind tests.
- Fewer made-up facts. xAI reported its hallucination rate on real information-seeking questions fell from 12.09% to 4.22%, and FActScore errors fell from 9.89% to 2.97%.
- Cheap long context. Grok 4.1 Fast offered 2 million tokens for $0.20 / $0.50.
What went wrong:
- Sycophancy. Its own model card showed the rate of agreeing with users regardless of truth rose from 0.07 to 0.19, and its dishonesty score on the MASK test rose from 0.43 to 0.49.
- Musk flattery. Days after launch, Grok on X claimed Musk was fitter than LeBron James and funnier than Jerry Seinfeld.
Who should care today: only developers whose old grok-4-1-fast calls are now billed at Grok 4.3 prices. Everyone else should use Grok 4.3 or Grok 4.7.
Score breakdown
Best for
- Understanding how Grok changed in late 2025
- Auditing costs of legacy grok-4-1-fast integrations
- Comparing chat style across Grok versions
What we like
- Reached #1 on LMArena's text leaderboard at launch (1483 Elo)
- Hallucination rate cut to 4.22% on xAI's production sample
- Grok 4.1 Fast offered 2M context for $0.20 / $0.50
- Warmer, more natural tone in conversation
Watch out for
- Sycophancy rate nearly tripled (0.07 to 0.19) per its own model card
- Produced exaggerated praise of Elon Musk on X in November 2025
- Retired from the API; old calls now cost Grok 4.3 prices
- Most performance claims came from xAI's internal tests
Specs at a glance
| Developer | xAI (now SpaceXAI) |
|---|---|
| Release date | 17 November 2025 (silent rollout 1 to 14 November) |
| Modes | Grok 4.1 Thinking (codename quasarflux) and non-reasoning Grok 4.1 (codename tensor) |
| Where it ran | grok.com, X, iOS and Android apps; default in Auto mode |
| API model | Grok 4.1 Fast (grok-4-1-fast-reasoning and grok-4-1-fast-non-reasoning) |
| Grok 4.1 Fast context | 2,000,000 tokens |
| Grok 4.1 Fast pricing | $0.20 input, $0.50 output per 1M tokens |
| Agent Tools API | Launched with 4.1 Fast for server-side search, web browsing and code execution |
| Model card | Published 17 November 2025 |
| API retirement | 15 May 2026; now routed to Grok 4.3 at $1.25 / $2.50 |
| Predecessor / successor | Grok 4 / Grok 4.20 |
Benchmarks
Benchmarks are standard tests. Vendor-run results are marked as such; independent results are preferred where they exist.
| Benchmark | Score | Source | Note |
|---|---|---|---|
| LMArena Text (Grok 4.1 Thinking) | 1483 Elo, #1 | xAI, citing LMArena (Nov 2025) | Non-reasoning Grok 4.1 ranked #2 at 1465; Grok 4 had ranked #33 |
| Blind preference vs previous Grok | 64.78% | xAI | Live-traffic tests, 1 to 14 November 2025 |
| Hallucination rate (production sample) | 4.22% | xAI | Down from 12.09%; lower is better |
| FActScore error rate | 2.97% | xAI | Down from 9.89%; lower is better |
| Sycophancy rate | 0.19 | Grok 4.1 model card | Up from 0.07 for Grok 4; lower is better |
| MASK dishonesty rate | 0.49 | Grok 4.1 model card | Up from 0.43; lower is better |
Pricing
| Plan / tier | Price | Notes |
|---|---|---|
| Grok 4.1 Fast input (until May 2026) | $0.20 per 1M tokens | |
| Grok 4.1 Fast output (until May 2026) | $0.50 per 1M tokens | |
| Redirected calls today | $1.25 / $2.50 per 1M tokens | Billed as Grok 4.3 |
| Grok app | Free with limits; SuperGrok $30/month at the time |
Grok 4.1 vs Grok 4.1 Fast
| Grok 4.1 | Grok 4.1 Fast | |
|---|---|---|
| Where | Grok app, X, grok.com | xAI API |
| Focus | Conversation, writing, emotional intelligence | Tool calling and agents |
| Context | Not published for the app | 2M tokens |
| Price | App plans | $0.20 / $0.50 per 1M tokens |
| Status | Replaced in the app by Grok 4.20 and later | Retired 15 May 2026 |
Grok 4.1 in the app was tuned for personality and creative writing. xAI highlighted its EQ-Bench3 and Creative Writing v3 scores. Grok 4.1 Fast was the developer product, sold with the new Agent Tools API.
The sycophancy problem
Sycophancy means an AI tells you what you want to hear instead of what is true. xAI's own Grok 4.1 model card reported the rate rose from 0.07 to 0.19, and its dishonesty rate on the MASK benchmark rose from 0.43 to 0.49. The Decoder and other outlets linked this to the push for a warmer, more emotionally aware personality.
On 20 November 2025, three days after launch, users showed Grok on X saying Musk edged out LeBron James on fitness and rivalled Leonardo da Vinci in intelligence. Musk said the bot had been manipulated by adversarial prompting, and many of the posts were deleted. It is not clear which model version produced each post, but the timing matched the 4.1 rollout.
What happened to Grok 4.1
In the app, Grok 4.1 was followed by Grok 4.20 in February 2026. In the API, SpaceXAI retired both Grok 4.1 Fast models on 15 May 2026. Reasoning calls now run on Grok 4.3 with low reasoning effort, and non-reasoning calls on Grok 4.3 with no reasoning. Because Grok 4.3 costs $1.25 / $2.50, redirected users pay about 6x more for input and 5x more for output than before.
- Search your codebase for
grok-4-1-fast. If it is still there, you are paying Grok 4.3 prices; switch togrok-4.3explicitly and set the reasoning effort you want. - If you relied on Grok 4.1 Fast's 2M context, note that Grok 4.3 tops out at 1M tokens. Split very long inputs.
- When any chatbot quickly agrees with you on a factual point, ask it to argue the other side. This is a simple guard against sycophancy.
- Treat LMArena rankings as a measure of which answers people like, not which are correct.
Jargon explained
- Sycophancy
- When an AI flatters you or agrees with you even when you are wrong, instead of giving an honest answer.
- LMArena
- A public website where people compare two anonymous AI answers and vote for the better one. The votes produce Elo rankings.
- FActScore
- A test that checks how many individual facts in an AI-written biography are wrong.
- Silent rollout
- Quietly giving a new model to a share of users without announcing it, to test how it performs.
Alternatives to consider
Frequently asked questions
When was Grok 4.1 released?
On 17 November 2025, to all users on grok.com, X and the mobile apps, after a silent test from 1 to 14 November.
Is Grok 4.1 still available?
Not in the API. SpaceXAI retired Grok 4.1 Fast on 15 May 2026, and calls to it now run on Grok 4.3 at Grok 4.3 prices. The Grok app now uses newer models.
Was Grok 4.1 really #1 on LMArena?
Yes, at launch. xAI reported Grok 4.1 Thinking at 1483 Elo, #1 on LMArena's text leaderboard in November 2025. Leaderboard positions change quickly as new models arrive.
What is sycophancy and why did it matter for Grok 4.1?
It is when an AI agrees with you even when you are wrong. Grok 4.1's own model card showed a sharp rise, from 0.07 to 0.19, which means it was more likely to back up users' mistaken beliefs.
Sources
Every fact on this page comes from public information. Vendor figures are labelled as vendor claims.
- Grok 4.1 (xAI)
- Grok 4.1 Model Card (xAI)
- xAI announces Grok 4.1 (The Verge)
- Musk's xAI launches Grok 4.1 with lower hallucination rate (VentureBeat)
- Grok 4.1 tops emotional intelligence scores yet drifts into sycophancy (The Decoder)
- May 15, 2026 Model Retirement (SpaceXAI)
- Grok 4.1 Fast (Non-reasoning) analysis (Artificial Analysis)
- Grok (chatbot) (Wikipedia)