Grok 4.3: 1M Context & $1.25/$2.50 Pricing Explained (2026)
Grok 4.3 by xAI (April 30, 2026): 1M token context, $1.25 input / $2.50 output per 1M tokens, GDPval-AA Elo 1500 (up 321 from Grok 4.20). Full specs inside.
Grok 4.3 is xAI's April 2026 flagship, holding a 1 million token context window at Grok 4.20's $1.25/$2.50 per 1M pricing while posting a 321-point GDPval-AA Elo gain to 1500. It reached AWS Bedrock and Azure AI Foundry within two months of launch, but xAI has not published a model card for this version.
Grok 4.3, released by xAI around April 30, 2026, is a flagship model with a 1 million token context window and a GDPval-AA Elo of 1500, up 321 points from Grok 4.20. It keeps Grok 4.20's exact pricing while adding AWS Bedrock and Azure AI Foundry availability.
Provider: xAI · Family: Grok 4
Context window: 1,000,000 tokens
Input modalities: text, image · Output: text
About Grok 4.3
Grok 4.3 is xAI's mid-2026 flagship chat and reasoning model, reaching the direct API around April 17, 2026 in a quiet rollout before Artificial Analysis dated its formal release to April 30, 2026. It sits between Grok 4.20 (predecessor, launched February to March 2026) and Grok 4.5 (successor, launched July 8, 2026) in xAI's lineup, though Grok 4.5 is positioned as a separate, cheaper coding-focused sibling rather than a strict replacement. xAI has not disclosed Grok 4.3's parameter count or whether it uses a dense or mixture-of-experts architecture; no official model card has been published for this version, a gap that persists from earlier Grok releases and that Fortune previously flagged as a recurring pattern for the product line. On Artificial Analysis's current live tracking, Grok 4.3 scores 38 on the Intelligence Index, ranked 55th of 186 tracked models, with an output speed of 94.2 tokens per second and a time-to-first-token around 22.28 seconds, a notably long delay consistent with a model that reasons before producing its first output token. Artificial Analysis's April 30, 2026 launch article separately reported a higher Intelligence Index of 53 at release, 4 points ahead of Grok 4.20 and just above Muse Spark and Claude Sonnet 4.6; the discrepancy between the launch-day figure and the current live score likely reflects a reasoning-effort tier difference or a later recalibration, and both numbers come from the same source (Artificial Analysis) rather than being independently cross-verified. The clearest, most consistent benchmark result is on GDPval-AA: Grok 4.3 scored an Elo of 1500, a 321-point jump from Grok 4.20's 1179, though it still trails GPT-5.5 by 276 Elo points with roughly a 17% expected win rate in that specific evaluation. Other reported gains include tau-squared-Bench Telecom at 98% (up 5 points) and AA-Omniscience accuracy up 8 points, though the model's non-hallucination rate on that same benchmark dropped 8 points, a real tradeoff. xAI has not published GPQA Diamond, AIME 2025, MMLU-Pro, SWE-bench Verified, ARC-AGI 2, or LMArena scores specifically for Grok 4.3 in any source found, so this page does not report them. The model carries a 1,000,000 token context window, matching Grok 4.20's single-model variants and confirmed directly from xAI's own developer docs. No separate maximum output token limit is documented. Modalities are text and image input with text-only output, per xAI's own docs page; this contradicts some secondary blog coverage claiming native video input, which could not be verified against the primary documentation and should be treated as inaccurate until xAI's docs show otherwise. Reasoning is configurable, and the model supports native function calling and structured outputs per xAI's developer documentation. Pricing is $1.25 per 1 million input tokens and $2.50 per 1 million output tokens, with cached input at $0.20 per 1 million tokens (an 84% discount) and a 20% batch API discount off standard rates, all confirmed directly from xAI's developer docs. This list price is identical to what xAI's own docs table shows for Grok 4.20, meaning Grok 4.3 did not change price versus its immediate predecessor at the API level, even though Artificial Analysis's launch article separately describes a 37.5% input and 58.3% output price cut relative to plain Grok 4 (an earlier, more expensive baseline). Rate limits are documented at 37 requests per second and 10,000,000 tokens per minute. Training data cutoff date could not be reliably confirmed: xAI's docs table shows a date that appears identical across every model row, a strong signal it is a stale placeholder rather than a genuine per-model figure, so this page leaves the cutoff unlisted rather than publish an unverified date. Grok 4.3 is reachable through the direct xAI API (model IDs grok-4.3, grok-4.3-latest, and the grok-latest alias) across the us-east-1, eu-west-1, and us-west-2 regions, and reached AWS Bedrock around June 2026 through a Mantle inference engine at an OpenAI-compatible endpoint. Microsoft Azure AI Foundry also lists Grok 4.3 with its own default Content Safety layer, and OpenRouter and Vercel AI Gateway offer the model as well; availability on Cloudflare, Snowflake, or Databricks Mosaic specifically for this version was not independently confirmed. One user-reported issue on AWS re:Post describes a slow vision encoder when running Grok 4.3 on Bedrock Mantle, though the full workaround thread could not be independently verified. The model is closed-weight and proprietary, API-only, with no license for self-hosting. Teams choosing between Grok 4.3 and Grok 4.5 should weigh the tradeoff directly: Grok 4.3 keeps the larger 1M token context window at the same $1.25/$2.50 price point, while Grok 4.5 trades context size down to 500K for a higher $2.00/$6.00 price in exchange for coding-specific tuning and same-day availability inside Cursor and Grok Build.
Pricing
$1.25 per 1M input tokens, $2.50 per 1M output tokens, $0.20 per 1M cached input tokens (84% discount), confirmed directly from xAI's developer docs. A 20% batch API discount applies off these standard rates. Pricing matches xAI's own table for Grok 4.20, so this release did not change list price versus its immediate predecessor.
Key Features
- 1M Token Context at Unchanged Pricing: Keeps Grok 4.20's 1 million token context window while holding pricing at $1.25 input / $2.50 output per 1M tokens.
- Large GDPval-AA Agentic Gain: GDPval-AA Elo rose 321 points to 1500 versus Grok 4.20's 1179, the release's clearest verified benchmark improvement.
- 84% Cached Input Discount: Cached input tokens cost $0.20 per 1M versus $1.25 standard, an 84% discount for repeat-context workloads.
- AWS Bedrock and Azure AI Foundry Availability: Reached AWS Bedrock via a Mantle inference engine and Microsoft Azure AI Foundry within about two months of launch.
- Configurable Reasoning with Function Calling: Supports configurable reasoning effort, native function calling, and structured outputs per xAI's developer documentation.
Pros
- GDPval-AA Elo jumped 321 points to 1500, the clearest verified agentic-benchmark gain versus Grok 4.20.
- 1 million token context window at the same $1.25/$2.50 per 1M pricing as its predecessor, no price increase for the larger window.
- Live on AWS Bedrock and Microsoft Azure AI Foundry within about two months of its April 2026 launch.
Cons
- No official xAI model card published for this version, unlike Grok 4, 4 Fast, and 4.1.
- Artificial Analysis's own launch-day and current Intelligence Index scores disagree (53 vs 38), with no public explanation for the gap.
- AA-Omniscience non-hallucination rate dropped 8 points even as accuracy rose 8 points, a documented quality tradeoff.
Benchmarks
- artificial analysis intelligence index: 38
- artificial analysis speed tokens per sec: 94.2
Frequently Asked Questions
What is Grok 4.3 and who built it?
Grok 4.3 is a large language model built by xAI, reaching the direct API in a quiet rollout around April 17, 2026, with Artificial Analysis dating its broader release to April 30, 2026. It sits between predecessor Grok 4.20 and successor Grok 4.5 in xAI's lineup, though Grok 4.5 is a separate, cheaper coding-focused sibling rather than a strict replacement. xAI has not disclosed the model's parameter count or whether it is dense or mixture-of-experts, and no official model card has been published for this version. The clearest verified improvement is a 321-point GDPval-AA Elo jump to 1500 versus Grok 4.20's 1179. It carries a 1 million token context window and is priced at $1.25 per 1M input tokens and $2.50 per 1M output tokens, unchanged from Grok 4.20. The model reached AWS Bedrock and Microsoft Azure AI Foundry within about two months of launch.
How much does Grok 4.3 cost per 1M tokens?
Grok 4.3 costs $1.25 per 1 million input tokens and $2.50 per 1 million output tokens, confirmed directly from xAI's developer documentation. Cached input tokens cost $0.20 per 1 million, an 84% discount versus fresh input. A 20% batch API discount applies off these standard rates for asynchronous workloads. This pricing is identical to what xAI's own docs list for the predecessor Grok 4.20, so Grok 4.3 did not raise or lower list price at the API level despite the benchmark gains. As a worked example, a daily coding agent processing 1 million input tokens and 200,000 output tokens costs roughly $1.75 per day at list price before caching savings. By comparison, sibling model Grok 4.5 costs more per token ($2.00 input / $6.00 output) despite a smaller context window, making Grok 4.3 the better per-dollar choice for large-context, budget-conscious workloads.
What is Grok 4.3's context window and max output?
Grok 4.3 supports a 1,000,000 token context window, confirmed directly from xAI's developer documentation and matching the context size of Grok 4.20's single-model variants. No separate maximum output token limit is documented; requests appear to draw from the same 1 million token budget as input and history. Long-context recall above 500K tokens was not independently benchmarked in sources reviewed. This context window is double the 500,000 tokens offered by the newer sibling model Grok 4.5, making Grok 4.3 the better choice for long-document analysis or large-codebase work within the current Grok lineup. Rate limits are documented at 37 requests per second and 10,000,000 tokens per minute. Document handling for PDFs or multi-file inputs beyond the base text-and-image modality was not confirmed in the documentation reviewed.
How does Grok 4.3 compare on benchmarks vs Grok 4.20?
The clearest, most consistently sourced benchmark gain for Grok 4.3 is on GDPval-AA, where it scored an Elo of 1500 versus Grok 4.20's 1179, a 321-point improvement confirmed by Artificial Analysis's April 30, 2026 launch article. That same article reported tau-squared-Bench Telecom improving 5 points to 98% and AA-Omniscience accuracy gaining 8 points, though the model's non-hallucination rate on that same benchmark dropped 8 points, a real quality tradeoff rather than a clean win. IFBench instruction-following held flat at 81% versus Grok 4.20. On the Artificial Analysis Intelligence Index, the launch-day figure (53, 4 points ahead of Grok 4.20) and the current live score (38, ranked 55th of 186 tracked models) disagree, likely due to a reasoning-effort tier difference or later recalibration that neither xAI nor Artificial Analysis has clearly documented. xAI has not published GPQA Diamond, AIME 2025, MMLU-Pro, SWE-bench Verified, or ARC-AGI 2 scores specifically for Grok 4.3, so no comparison on those axes can be made with confidence. Against GPT-5.5, Grok 4.3 trails by 276 Elo points on GDPval-AA specifically, with roughly a 17% expected win rate in that evaluation.
Is Grok 4.3 open source or proprietary?
Grok 4.3 is proprietary and closed-weight; xAI has not released its weights and access is API-only. This is consistent with every Grok release since Grok-1, which xAI open-sourced under Apache 2.0 back in March 2024, a policy that has not applied to any Grok 4-generation model including 4.3. Access runs through xAI's direct API using the model IDs grok-4.3, grok-4.3-latest, and the grok-latest alias, plus managed access via AWS Bedrock through a Mantle inference engine and Microsoft Azure AI Foundry. OpenRouter and Vercel AI Gateway also offer the model; availability on Cloudflare, Snowflake, or Databricks Mosaic specifically for this version was not independently confirmed. No SDK language list was surfaced in the documentation reviewed, though xAI's API follows an OpenAI-compatible convention across the Grok line. There are no open-weight variants of Grok 4.3.
What modalities does Grok 4.3 support?
Grok 4.3 accepts text and image input and produces text-only output, confirmed directly from xAI's developer documentation. This directly contradicts some secondary blog coverage claiming native video input support, which could not be verified against the primary docs and should be treated as inaccurate. The model supports native function calling and structured JSON outputs, and reasoning effort is configurable per request. No audio input or output, and no video input or output, is confirmed for this version. Microsoft's Azure AI Foundry deployment applies its own Content Safety filtering layer by default on top of the base model's modalities. Compared to xAI's separate Grok Imagine product, which handles text-to-video generation, Grok 4.3 itself does not generate images or video directly.
Does Grok 4.3 train on user data?
No Grok 4.3-specific data retention or training-on-inputs policy has been published, and no official model card exists for this version as of this writing, a gap versus Grok 4, Grok 4 Fast, and Grok 4.1, each of which shipped a model card. SOC 2, ISO 27001, HIPAA, and GDPR compliance status for this specific model were not confirmed in sources reviewed. Microsoft's Azure AI Foundry listing does apply its own default Content Safety and governance layer for enterprise customers deploying Grok 4.3 through that channel, which may offer more documented compliance posture than the direct xAI API. Training data cutoff date could not be reliably confirmed: xAI's own docs table shows a date that appears identical across every model row, suggesting a stale placeholder rather than a genuine per-model figure, so this page leaves it unlisted. AWS Bedrock deployments run through a separate Mantle inference engine, and its own data-handling terms may differ from both the direct xAI API and the Azure listing. Enterprise buyers with strict compliance requirements should request xAI's or their chosen cloud provider's current data-handling terms directly rather than assume a policy based on other Grok releases.
Who is Grok 4.3 best for and who should avoid it?
Grok 4.3 fits teams that need a large 1 million token context window without paying a price premium over Grok 4.20, and agentic or tool-calling workflows that benefit from the model's 321-point GDPval-AA Elo gain. It suits enterprises already standardizing on AWS Bedrock or Microsoft Azure AI Foundry, since Grok 4.3 reached both within about two months of its April 2026 launch. Teams that specifically need coding-tuned performance and don't mind a smaller context window should compare it against sibling model Grok 4.5, which costs more per token but launched with same-day Cursor and Grok Build integration. Buyers who require a published model card or verified GPQA, AIME, or SWE-bench scores for procurement should be cautious, since none of these exist for Grok 4.3 as of this writing. Teams sensitive to long time-to-first-token delays should also test latency directly, since Artificial Analysis measured roughly 22 seconds to first token on reasoning-heavy requests. Workloads doing heavy image analysis on AWS Bedrock specifically should also test vision-encoder latency before committing, given the user-reported slow vision encoder issue on that endpoint.