Grok 4.3
Last reviewed
Jun 2, 2026
Sources
12 citations
Review status
Source-backed
Revision
v1 · 2,110 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 2, 2026
Sources
12 citations
Review status
Source-backed
Revision
v1 · 2,110 words
Add missing citations, update stale details, or suggest a clearer explanation.
Grok 4.3 is a frontier large language model developed by xAI, released through the company's API on April 30, 2026 and rolled out to general availability the following week.[1][2] It is the fourth public model in the Grok 4 family, arriving after Grok 4 and Grok 4.1, and xAI positioned it less as an attempt to top every leaderboard than as a cost-efficient workhorse: a one-million-token context window, always-on reasoning, agentic tool use, and pricing well below the previous generation.[1][3] Its launch was paired with Custom Voices, a separate text-to-speech and voice-cloning suite.[1]
Grok 4.3 is a text-and-image model that returns text, with reasoning turned on by default for every request.[4][1] xAI's developer documentation lists a context window of 1,000,000 tokens, function calling, structured outputs, and reasoning as standard capabilities, and prices the model at $1.25 per million input tokens, $0.20 per million cached input tokens, and $2.50 per million output tokens.[4] The independent evaluation firm Artificial Analysis scored it at 53 on its Intelligence Index, placing it a few points above xAI's prior internal release and roughly level with mid-tier models from competing labs while costing far less to run.[3]
Rather than chasing a single headline benchmark, xAI leaned into breadth and price. Several reviewers read the release as the company stepping back from the "best model in the world" framing it used for Grok 4 and instead optimizing for the kind of everyday agentic work, document generation and tool calling, that most users actually do.[1][2] The model ships with a December 2025 knowledge cutoff.[5]
Grok 4.3 first appeared on April 17, 2026, when a build labeled "Early Access" showed up in the model selector on grok.com and in xAI's mobile apps for subscribers on the top-tier SuperGrok Heavy plan, then priced at $300 per month.[6][5] There was no blog post or press release at the time; early testers noticed the new option and began posting results before xAI said anything publicly.[6]
The public, OpenAI-compatible API opened on April 30, 2026, the date Artificial Analysis records as the model's release.[2][3] xAI then moved Grok 4.3 to general availability during the week of May 4, 2026, expanding consumer access in stages across its subscription tiers and the social network X.[5][1] On May 6, 2026, the company emailed API customers a notice titled "Grok 4.3 release and xAI API model retirement," confirming the model's general availability and announcing that eight older Grok endpoints would be shut off on May 15, 2026.[7] The retired endpoints included grok-3, grok-4-0709, grok-4-fast and its non-reasoning variant, grok-code-fast-1, grok-imagine-image-pro, and the two grok-4-1-fast reasoning and non-reasoning models, with developers directed to migrate hard-coded model IDs to Grok 4.3.[7]
Grok 4.3 sits at the top of xAI's general-purpose lineup as of mid-2026, succeeding the Grok 4 flagship from 2025 and the Grok 4.1 refresh that followed it. Artificial Analysis benchmarked Grok 4.3 against an interim xAI checkpoint it refers to as "Grok 4.20," reporting that the new model roughly matched that checkpoint's intelligence while cutting the cost of a full benchmark run by about 20 percent.[3] In Artificial Analysis terms, the headline gain was efficiency and agentic skill rather than raw capability: Grok 4.3 landed about four points ahead of the interim build on the Intelligence Index, just above Claude Sonnet 4.6 and the Muse Spark model on the same chart.[3]
The release sat within a crowded frontier field. By this point rivals had shipped GPT-5.5, Gemini 3, and Claude Opus 4.5, and Artificial Analysis noted that on its agentic economic-value benchmark Grok 4.3 still trailed the top OpenAI configuration by a wide margin even after a large jump of its own.[3] xAI's pitch was therefore comparative value: close-to-frontier behavior at a fraction of the token cost.[1]
xAI disclosed little about Grok 4.3's internals. The company did not publish a formal model card detailing parameter counts, the training corpus, or the hardware used, and the public developer page restricts itself to interface-level facts: context length, supported modalities, pricing, rate limits, and the regions where the model is served.[4] Independent coverage described Grok 4.3 as a freshly pre-trained model in the same scale class as the prior xAI checkpoint, paired with a revised post-training recipe, but those characterizations were not confirmed in an official technical report.[3]
What xAI did make explicit is the reasoning behavior. Where earlier Grok releases let callers toggle a "thinking" mode or pick an effort level, Grok 4.3 treats chain-of-thought reasoning as always on: the model deliberates before answering on every request and the behavior cannot be switched off.[1][5] Callers can still tune how much effort it spends through a reasoning-intensity setting with low, medium, and high levels.[8] Because so few architectural details were released, claims about the use of a mixture-of-experts design or specific reinforcement learning methods cannot be verified from primary sources and are omitted here.
In the API, Grok 4.3 accepts text and images and produces text.[4] xAI extended the model's practical reach in two directions beyond that baseline.
First, video. In the grok.com app and apps, a user can attach a video file, up to five minutes long and up to 1080p in mp4, mov, or webm, and Grok 4.3 will reason over it, covering speech transcription, speaker segmentation, object tracking, and motion in a single pass.[5][9] xAI presented this as its first model able to take video directly rather than requiring a separate transcription or frame-extraction step beforehand.[9] In practice the system extracts frames from the clip and bills them as image tokens, so the feature is best understood as a video front end built on the model's image input rather than a distinct native-video token stream; the developer documentation continues to list the API modalities as text and image only.[5][4]
Second, agentic output. Grok 4.3 can run as an autonomous worker with access to a sandboxed computer where it writes and executes code, installs packages, and produces files, and it can generate finished documents directly, including PowerPoint (PPTX), Excel (XLSX), and PDF.[1][10] Reviewers highlighted strong instruction following and tool calling as the model's most improved areas, consistent with xAI's framing of it as built for real work rather than for benchmark maxima.[2][1]
Alongside the model, xAI shipped Custom Voices, a text-to-speech and voice-cloning suite. It can clone a voice from a reference clip of roughly two minutes and reuse that voice ID across xAI's TTS and voice-agent APIs, offers more than 80 preset voices across 28 languages, and gates cloning behind a two-stage consent flow that pairs a spoken passphrase with a speaker-embedding match.[11][1] The standalone TTS service exposes five named voices (Eve, Ara, Rex, Sal, and Leo) at $4.20 per million characters, and a real-time speech-to-speech voice agent (model ID grok-voice-think-fast-1.0) billed at $3.00 per hour, or $0.05 per minute.[1]
Most published numbers for Grok 4.3 come from Artificial Analysis, which evaluates the high-reasoning configuration. xAI did not release a benchmark table of its own. The figures below are drawn from Artificial Analysis coverage; absolute scores it did not publish are left out rather than estimated.
| Benchmark | Grok 4.3 (high) | Notes |
|---|---|---|
| Artificial Analysis Intelligence Index | 53 | About 4 points above xAI's interim "Grok 4.20" checkpoint [3] |
| GDPval-AA (agentic economic value) | 1500 Elo | Up 321 Elo over the interim checkpoint; trails the leading GPT-5.5 configuration by 276 Elo [3] |
| τ²-Bench Telecom (tool use) | 98% | A 5-point gain over the prior checkpoint [3] |
| IFBench (instruction following) | 81% | Roughly flat versus the prior checkpoint [3] |
| GPQA Diamond | ~90% | Reported approximate figure [10] |
| Output speed | 144.6 tokens/sec | Measured by Artificial Analysis [2] |
| Cost to run the Intelligence Index | $395 | About 20% cheaper than the prior checkpoint [3] |
Artificial Analysis also flagged the model as unusually verbose, noting it generated about 88 million output tokens to complete the Intelligence Index against a median near 35 million for comparable systems, which partly offsets the low per-token price on long evaluation runs.[2] Other benchmarks it tracked for the model but did not assign public scores to here include Humanity's Last Exam, SWE-bench-style coding and agentic suites such as Terminal-Bench, and its own AA-Omniscience and AA-LCR tests.[3]
Grok 4.3 is reachable through the consumer Grok apps, the social network X, and the developer API. The model serves a one-million-token context window, with input requests above 200,000 tokens billed at double the standard input rate.[5] The API is OpenAI-compatible and the model is also offered through third-party routers, including OpenRouter.[5][1]
| Item | Detail |
|---|---|
| API model ID | grok-4.3 [4] |
| Context window | 1,000,000 tokens [4] |
| Input price | $1.25 per million tokens [4] |
| Cached input price | $0.20 per million tokens [4] |
| Output price | $2.50 per million tokens [4] |
| Long-context surcharge | Requests over 200,000 input tokens billed at 2x input [5] |
| Modalities (API) | Text and image in, text out [4] |
| Video (apps) | mp4/mov/webm, up to 5 min, up to 1080p; frames billed as image tokens [5][9] |
| Knowledge cutoff | December 2025 [5] |
| Reasoning | Always on; low / medium / high intensity [1][8] |
| Consumer access | Free tier with limits; SuperGrok ($30/mo); X Premium+ ($40/mo); SuperGrok Heavy ($300/mo) [1][5] |
| Third-party access | xAI API and routers including OpenRouter [5][1] |
The pricing was the launch's main talking point. Against the previous generation, xAI cut input prices by roughly 37 to 40 percent and output prices by about 58 to 60 percent, undercutting comparable frontier offerings from other labs.[3][5]
Coverage was broadly positive about the value proposition and notably warmer about xAI's change in tone. VentureBeat framed the launch around the combination of aggressive pricing and the new voice-cloning suite, calling the model's price point a deliberate play for developers.[1] Artificial Analysis credited the release with real gains in agentic tool use and instruction following at a lower cost, while pointing out that it did not lead the frontier on the hardest economic-value tasks.[3] Several reviewers welcomed xAI dropping the "best model" posturing and shipping something practical, with one widely shared review arguing the company was better off for it.[12] The verbosity Artificial Analysis measured drew the most consistent criticism, since it can erode the token-cost advantage on long jobs.[2]
The clearest limitation is transparency. xAI published interface specifications but no technical report, so independent parties cannot verify claims about model scale, training data, or architecture, and the agentic and reasoning improvements rest largely on third-party benchmarks rather than disclosed methods.[4][3] The "native video" description also warrants care: because frames are extracted and billed as image tokens and the documented API modalities remain text and image, the capability is more accurately a video front end than a separate video modality.[5][4]
Custom Voices raised the familiar safety questions that accompany high-quality voice cloning. xAI built a two-stage consent flow to deter cloning a voice without the speaker's involvement, but as one outlet noted, the company did not publish false-acceptance rates, anti-spoofing results, or red-team findings, leaving the system's safety properties unverified by outside researchers at launch.[11] Finally, the abrupt retirement of eight older API models on May 15, 2026, gave developers a short window to migrate any hard-coded model identifiers to Grok 4.3 or another supported endpoint.[7]