Citation and evidence

Grok 4.3

14 min full readUpdated 16 references

This article's verification

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

AI ModelsLarge Language Models

Cite this article

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 part of the Grok 4 family, arriving after Grok 4, Grok 4.1 and Grok 4.20, 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Its launch was paired with Custom Voices, a separate text-to-speech and voice-cloning suite.[1] Grok 4.3 was superseded as the company's flagship by Grok 4.5 on July 8, 2026, which was followed by Grok 4.6 on August 12 and Grok 4.7 on September 21, 2026. Grok 4.3 remained a listed API model as of November 2026.[13][14][16]

Related: GPQA DiamondOpenAIGrokGrok 4

API prices and benchmark results

Compare all listed models

Standard text API rates in USD per 1 million tokens. Named benchmark results are not a universal ranking.

List prices: AI Compute Tracker, owned and operated by the same team as AI Wiki. Routed rates: OpenRouter.

This panel uses the same automatically refreshed snapshot as the front page. Benchmark versions or test configurations may differ or be unknown; scores are not ranked.
ModelInput / 1MOutput / 1MBenchmark resultExample cost
Grok 4.3x-aiDirect provider API$1.25Cached: $0.2$2.5Not measured in this snapshotExample cost$3.75
Sources & rate details for Grok 4.3
AI Compute Tracker (opens in a new tab)Price reviewed Provider pricing referenceSource fetched UTC

AI Compute Tracker reviewed standard (non-batch) provider list prices in USD per 1M tokens, at the short-context or peak-hour tier. Cached input is shown only when recorded in that row; it is not inferred from a discount. Long-context, cache writes, batch, tools, media and other charges are not included. Tracker row notes: Below 200K prompt tokens; at/above 200K $2.50/$5

Example cost: 1,000 requests, each with 2,000 input and 500 output tokens, at the listed standard text rates. No caching, batch discount, long-context surcharge, tools, media or tax; not a capability or availability check. Adjust this workload in the cost planner (choose a model there).

Compare and deploy Grok 4.3

These sourced facts are shared with the comparison and planning tools. Each field retains its own review date.

Grok 4.3

Article updated after the pricing / specification review. Inspect the sources before relying on these fields; a newer edit does not necessarily change the model facts.

Provider API lifecycle
ActiveSource · checked 2026-09-11 · review due
Context tokens
1MSource · checked 2026-07-23

The official model page lists no maximum output limit or committed retirement date. Alias mappings can change; use the explicit model ID.

Grok logo
Grok logo. Source: Wikimedia Commons, Public domain.

A note on the company name: xAI became a wholly owned part of SpaceX in February 2026 and rebranded as SpaceXAI in July 2026; by November 2026 its developer documentation used a further name, SpaceXSI.[16] The Grok 4.5 model card states that "SpaceXAI is a doing-business-as (dba) name of XAI LLC," so the two names refer to the same organization.[15] This article uses "xAI" for the company as it presented itself when Grok 4.3 shipped.

Overview

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; all four figures were still current on the documentation on July 27, 2026.[4][13] The independent evaluation firm Artificial Analysis scored it at 53 on the version of its Intelligence Index in use at the April 2026 launch, placing it four points above the latest version of xAI's previous public model, Grok 4.20, and roughly level with mid-tier models from competing labs while costing far less to run.[3] That figure is not comparable to the index as it stood in July 2026: on Intelligence Index v4.1, read from Artificial Analysis on July 27, 2026, Grok 4.3 (high) records 37.6, with a coding index of 42.2 and an agentic index of 24.1, and the page marks the model deprecated in favour of Grok 4.5.[2]

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][12] The model ships with a December 2025 knowledge cutoff.[1]

Development and release

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]

Place in the Grok line

Grok 4.3 sat at the top of xAI's general-purpose lineup from May to early July 2026, succeeding the Grok 4 flagship from 2025 and the Grok 4.1 refresh that followed it, and was in turn succeeded by Grok 4.5 on July 8, 2026.[13] Artificial Analysis benchmarked Grok 4.3 against xAI's previous public model, Grok 4.20 (the 0309 v2 version), 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 Grok 4.20 on the Intelligence Index, just above Claude Sonnet 4.6 and the Muse Spark model on the same chart.[3]

One point of continuity is worth flagging, because it runs against the usual pattern of successive releases. Grok 4.5 did not enlarge the context window; it cut it, to 500,000 tokens. SpaceXAI's own model page for Grok 4.5 lists 500k, while grok-4.3 remains at 1M on the same pricing table, so Grok 4.3 is still the longer-context model of the two.[13][14] Artificial Analysis's listings show the same split, 500,000 tokens for Grok 4.5 against 1,000,000 for Grok 4.3.[2]

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]

Architecture and training

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 launched with chain-of-thought reasoning always on: xAI's launch documentation said the model "thinks before responding" and offered no way to switch the behavior off.[1][4][5] xAI later made reasoning configurable; by mid-June 2026 the model page listed settings of none (no reasoning at all), low, medium, and high, and as of November 2026 it lists none, low, medium, high, and xhigh, with low as the default.[4] 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.

Capabilities and modalities

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] Coverage described it as the first xAI model able to take video directly rather than requiring a separate transcription or frame-extraction step beforehand.[5] Accounts of how this works differ: Codersera and AI CERTs describe the vision encoder ingesting video directly rather than sampling frames, while one API reseller's guide says video frames are tokenized and billed as input tokens. xAI's developer documentation lists the API modalities as text and image only.[5][6][8][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] At launch the standalone TTS service exposed 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] Voice pricing has since moved: the developer documentation on July 27, 2026 listed text to speech at $15.00 per million characters, with the realtime voice mode still at $0.05 per minute and $3.00 per hour.[13] By November 2026 the models page listed speech-to-speech at $0.08 per minute, with text to speech unchanged at $15.00 per million characters.[13]

Benchmarks

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. Two dates matter throughout: April 30, 2026, when Artificial Analysis published its launch write-up, and July 27, 2026, when the model page was read for this article. Artificial Analysis moved to Intelligence Index v4.1 after the launch and to v5 in October 2026, and index versions are not comparable across a version change.

BenchmarkGrok 4.3 (high)Source and date
Artificial Analysis Intelligence Index (launch-era version)53About 4 points above the latest version of Grok 4.20; AA launch article, April 30, 2026 [3]
Artificial Analysis Intelligence Index v4.137.6AA model page, read July 27, 2026 [2]
Artificial Analysis Coding Index v4.142.2AA model page, read July 27, 2026 [2]
Artificial Analysis Agentic Index v4.124.1AA model page, read July 27, 2026 [2]
GDPval-AA (agentic economic value)1500 Elo at launch; 1083.8 on the current boardAA reported 1500 at launch, up 321 Elo over Grok 4.20 0309 v2 and 276 Elo behind GPT-5.5 (xhigh); the stored value on July 27, 2026 is 1083.77, so the Elo pool has been recomputed and the two are not comparable [3][2]
τ²-Bench Telecom (tool use)98%A 5-point gain over the prior checkpoint; AA, April 30 and July 27, 2026 [3][2]
IFBench (instruction following)81%Roughly flat versus the prior checkpoint; AA, April 30 and July 27, 2026 [3][2]
GPQA Diamond90%AA model page, read July 27, 2026 [2]
Median output speed104.0 tokens/secAA model page, read July 27, 2026; AA serving numbers drift, so this figure needs its date [2]
Cost to run the Intelligence Index$395 at launch; $303 on July 27, 2026AA launch article gave $395, about 20% cheaper than the prior checkpoint; the model page gave $303.34 when read on July 27, 2026 [3][2]

Expanded article table

Artificial Analysis measured Grok 4.3 as using about 44 percent more output tokens to complete the Intelligence Index than the Grok 4.20 0309 v2 checkpoint, but concluded that it "uses a similar number of tokens to models like Minimax M2.7 and remains less verbose than other leading models."[3] The canonical token count on its model page was 83.2 million output tokens for a full index run when read on July 27, 2026.[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]

Availability, context window, pricing, and access

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 requests whose prompt reaches 200,000 tokens billed at double the standard rate for input, cached input, and output alike.[4] The API is OpenAI-compatible and the model is also offered through third-party routers, including OpenRouter.[5][1]

The table below reflects the developer documentation as read on July 27, 2026.

ItemDetail
API model IDgrok-4.3 (aliases grok-4.3-latest, grok-latest) [4]
Context window1,000,000 tokens [4][13]
Input price$1.25 per million tokens [4][13]
Cached input price$0.20 per million tokens [4][13]
Output price$2.50 per million tokens [4][13]
Long-context pricingRequests whose prompt reaches 200,000 tokens are billed at 2x on every token type: $2.50 input, $0.40 cached, $5.00 output [4][13]
Batch APIBilled at a 20% discount to standard rates [4]
Rate limits37 requests per second; 10,000,000 tokens per minute [4]
Regionsus-east-1, eu-west-1, us-west-2 [4]
Modalities (API)Text and image in, text out [4]
Video (apps)mp4/mov/webm, up to 5 min, up to 1080p [5]
Knowledge cutoffDecember 2025 [1]
ReasoningConfigurable: none, low, medium, high (always on at launch) [1][4]
Consumer accessFree tier with limits; SuperGrok ($30/mo); X Premium+ ($40/mo); SuperGrok Heavy ($300/mo) [1][5]
Third-party accessxAI API and routers including OpenRouter [5][1]

Expanded article table

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]

Reception

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 jump in output-token usage relative to the prior checkpoint drew some comment, since a chattier model erodes part of the token-cost advantage on long jobs, though Artificial Analysis itself placed Grok 4.3 below other leading models on verbosity.[3]

Limitations and controversies

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: xAI has not documented how the app feature processes or bills video, and the documented API modalities remain text and image.[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]

Continue reading

More to read

Topics:AI ModelsLarge Language Models

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19VentureBeat, "xAI launches Grok 4.3 at an aggressively low price and a new, fast, powerful voice cloning suite," venturebeat.com/...st-powerful-voice-cloning-suite
  2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Artificial Analysis, Grok 4.3 model page, read July 27, 2026, artificialanalysis.ai/...grok-4-3
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18Artificial Analysis, "xAI launches Grok 4.3 with improved agentic performance and lower pricing," artificialanalysis.ai/...ormance-and-lower-pricing
  4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21xAI Docs, "Grok 4.3," read July 27, 2026, docs.x.ai/...grok-4.3
  5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Codersera, "Grok 4.3: xAI's Cheap Frontier Model (May 2026 Guide)," codersera.com/...grok-4-3-launch-guide-2026
  6. ^1 ^2 ^3AI CERTs News, "Grok 4.3 Shows Multimodal AI Models Gaining Native Video Skills," aicerts.ai/...i-models-gaining-native-video-skills
  7. ^1 ^2 ^3Apiyi Blog, "Grok 4.3 Launches on xAI API: Complete Migration Guide for 8 Legacy Models Retiring on May 15," help.apiyi.com/...ease-xai-api-model-retirement-en
  8. ^Apiyi Blog, "Grok 4.3 API major release: 5 major upgrades + 40% price cut," help.apiyi.com/...4-3-api-release-may-2026-news-en
  9. ^ChatlyAI, "xAI Drops Grok 4.3 Beta With Video, Slides & Speech APIs," chatlyai.app/...xai-grok-4-3-beta-video
  10. ^FelloAI, "Grok 4.3 Review: Features, Price, and Verdict (2026)," felloai.com/grok-4-3-review
  11. ^1 ^2WinBuzzer, "xAI Adds Voice Cloning to New Grok 4.3 AI Model," winbuzzer.com/...oices-voice-cloning-launch-xcxwbn
  12. ^1 ^2302.AI (Medium), "Grok 4.3 Review: xAI Gives Up on Being the 'Best' Model, and That's a Good Thing," medium.com/...-and-thats-a-good-thing-b40bfa7505bb
  13. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11SpaceXAI Docs, "Models" (text API, Imagine and voice pricing table), read July 27, 2026, docs.x.ai/...models
  14. ^1 ^2SpaceXAI Docs, "Grok 4.5," read July 27, 2026, docs.x.ai/...grok-4.5
  15. ^SpaceXAI, "Model Card: Grok 4.5," July 14, 2026, revised July 20, 2026, media.x.ai/...4p5-5184fdf9.pdf
  16. ^1 ^2SpaceXSI Docs, "Release Notes," last updated November 2, 2026, accessed November 5, 2026, docs.x.ai/...release-notes

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

6 revisions · v6 · 2,873 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent fact-check of the full article against its cited sources (90 claims, 24 sources fetched); 8 material + 14 minor fixed

Cite this page: AI Wiki. "Grok 4.3." aiwiki.ai, updated 28 Oct 2026, fact-checked 5 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/grok_4_3

Edit article