Overview
Grok 4.3 is xAI's April 30, 2026 API release — pushed live with no announcement post and following a quiet beta that first appeared in the grok.com model picker on April 17, 2026. Compared with grok-4.20-multi-agent-0309, this version is better described as a model-positioning, reasoning-control, and cost-efficiency update rather than a confirmed modality reset: the model is positioned as xAI's newest default API model, with 1M context, function calling, structured outputs, configurable reasoning, and official $1.25 input / $2.50 output pricing.
For developers running agentic pipelines or high-volume API workloads, Grok 4.3 is now xAI's recommended default. For Grok chat users, third-party launch coverage and early user reports describe new artifact-style outputs such as documents, spreadsheets, PDFs, and PowerPoint files, but availability may vary by plan, region, and staged rollout. A separate Custom Voices / voice-cloning capability covered by VentureBeat is documented under xAI's Voice API docs and is not part of the core grok-4.3 model.
What's New
API Model Profile
Official xAI docs position grok-4.3 as the recommended default for chat and coding workloads, with a 1M-token context window, function calling, structured outputs, and configurable reasoning. xAI's public docs do not currently list native video-file understanding as a confirmed grok-4.3 capability; video generation is documented separately under the Grok Imagine API.
Aggressive Price Cut
API pricing is officially listed at $1.25 per million input tokens, $0.20 per million cached input tokens, and $2.50 per million output tokens for grok-4.3; xAI also notes different pricing for requests above the 200K context threshold. Artificial Analysis described the launch pricing as 37.5% lower on input and 58.3% lower on output versus its Grok 4.20 0309 v2 comparison snapshot; avoid implying that current xAI docs still price Grok 4.20 above Grok 4.3. xAI positions Grok 4.3 as its recommended default for chat/coding and agentic tool calling; cost savings should be framed as workload-dependent because token pricing, cached-input pricing, higher-context pricing, and separate tool-call fees all affect the final bill.
Slides, Spreadsheets, PDFs, and PowerPoint Outputs
Third-party launch coverage and early users report downloadable artifacts from Grok chat, including slide decks, spreadsheets, PDFs, and PowerPoint files. Treat this as a chat-surface rollout detail rather than a confirmed grok-4.3 API capability unless xAI documents it officially. When it lands for an account, it converts a class of "draft, then export, then format" workflows into a single chat turn.
Voice Cloning Suite
VentureBeat's launch coverage flagged a new Custom Voices / voice-cloning capability documented separately in xAI's Voice API docs — keep it scoped as an adjacent xAI release rather than a grok-4.3 model feature. Capability and pricing details should be verified directly in the Voice API documentation before production use.
Configurable Reasoning Effort
Grok 4.3 exposes a reasoning_effort parameter with none / low (default) / medium / high settings, letting callers trade latency and cost against deeper deliberation. none disables reasoning for near-instant, non-reasoning-style use cases. This is the same mental model OpenAI and Anthropic have converged on for reasoning models, and makes effort an explicit knob rather than an implicit "always-on" assumption.
Performance Benchmarks
Third-party benchmarks position Grok 4.3 as a small but consistent step ahead of Grok 4.20, with the bigger story being cost efficiency rather than raw intelligence:
| Benchmark | Grok 4.3 | Grok 4.20 | Notes |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 53 | 49 | +4 points |
| GDPval-AA (agent ELO) | 1500 | 1179 | +321 ELO |
| τ²-Bench Telecom | 98% | — | Agentic customer-support task |
| IFBench | 81% | — | Instruction-following |
| Output speed | Verify current AA value | — | Artificial Analysis speed metrics are volatile; refresh from the current Grok 4.3 model/provider page before citing |
| Context window | 1M | 2M | Reduced from 4.20 |
Numbers above are from Artificial Analysis's launch coverage. xAI has not published a standalone benchmark page for 4.3, so domain-specific scores (GPQA, AIME, SWE-Bench, LiveCodeBench) should be verified independently for hard-bound use cases.
Migration Guide
Most developers can switch their API client to grok-4.3 and ship — but two changes are worth checking before flipping the model ID in production:
- Context window halved. Grok 4.3 supports a 1M-token context vs. 4.20's 2M. Workloads that previously fit a single 1.5M-token corpus into one call (long codebase analysis, large legal document bundles) will need chunking or retrieval. If your prompts comfortably stayed under 1M, no change is needed.
- Reasoning effort is now explicit. Add
reasoning_effort: "none" | "low" | "medium" | "high"to your request payload. xAI documentslowas the default reasoning effort and also supportsnone,medium, andhigh; teams porting from 4.20 should benchmark each effort level on representative traffic before promoting. - Tool-call pricing still bills separately. Web search, X search, file search, code execution, and RAG calls retain their own per-call rates. Re-run cost models on a representative 24-hour traffic sample, since the input/output discount is largest on text-heavy turns and smaller on tool-heavy turns.
- Treat video input as unconfirmed for now. xAI's public docs do not currently document native video-file input as a
grok-4.3API capability. Existing image and text endpoints work unchanged; for video generation, use the separately documented Grok Imagine API rather than the coregrok-4.3model.
Availability & Access
- API:
grok-4.3is the recommended default model on x.ai's API as of April 30, 2026 - Grok chat: Available through the xAI API and rolling out across Grok chat plans; SuperGrok, X Premium+, and SuperGrok Heavy are current subscription routes, with model availability confirmed in the in-product picker
- Beta entrypoint: Third-party reports place the early Grok 4.3 chat beta in mid-April 2026, but treat the exact April 17 / SuperGrok Heavy claim as qualified unless verified from official Grok release notes
- Free tier: xAI typically grants free-tier users limited access to flagship models — confirm in the in-product picker, since limits change frequently
Pricing & Plans
| Plan | Monthly Price | Grok 4.3 Access |
|---|---|---|
| API (grok-4.3) | $1.25 input / $2.50 output per 1M tokens | Direct API access, pay-as-you-go |
| Free | $0 | Varies by rollout — check picker |
| X Premium | ~$8 | Varies by rollout |
| SuperGrok | $30 | ✓ in chat |
| X Premium+ | ~$40 | ✓ in chat + X platform perks |
| SuperGrok Heavy | $300 | ✓ + Heavy-tier priority and Grok 4 Heavy |
Tool-call surcharges (web search, X search, file search, code execution, RAG) bill separately on top of the per-token rate.
Pricing note: xAI now publishes an official
grok-4.3model/pricing page listing $1.25 input, $0.20 cached input, and $2.50 output per 1M tokens, with higher-context pricing applying above 200K tokens. Use xAI docs as the source of truth and keep Artificial Analysis/OpenRouter as third-party benchmark references.
Best For
- API teams running agentic pipelines who want lower cost-per-call without dropping below Grok-tier intelligence
- Developers building AI chatbots or assistants that need long-context reasoning, tool calling, structured outputs, and lower-cost text/image workflows
- Operators with high-volume, instruction-following workloads (customer support routing, content moderation, structured extraction) who care about IFBench-style metrics
- SuperGrok and X Premium+ users who want one-shot slide / spreadsheet / PDF generation in chat without an external automation layer
- Power users comparing flagship models — see the best AI assistant roundup for cross-vendor positioning
- Teams migrating away from heavier 2M-context loads who can chunk down to 1M and capture the price drop
FAQ
How do I access Grok 4.3?
Two paths. API: call the grok-4.3 model ID at x.ai/api — pay-as-you-go at $1.25 input / $2.50 output per million tokens. Chat: subscribe to SuperGrok ($30/month), X Premium+ (~$40), or SuperGrok Heavy ($300) and pick Grok 4.3 from the model selector on grok.com or the mobile apps.
How is Grok 4.3 different from Grok 4.20?
Three concrete changes: (1) native video input arrives for the first time, (2) API pricing drops ~40% on input and ~60% on output, and (3) the context window shrinks from 2M to 1M tokens. Reasoning effort is now an explicit low/medium/high parameter rather than implicit. Architecturally, xAI has not described 4.3 as a multi-agent system the way 4.2 was characterized — it reads more like a tuned, cost-optimized successor than a structural rebuild.
Why did the context window drop from 2M to 1M?
xAI has not publicly explained the change. Practically, most production agent workloads run well under 1M tokens, and the price-per-token cut compensates for the smaller window in those cases. Workloads that genuinely needed 2M (full-codebase analysis, large discovery bundles) will require chunking or staying on 4.20 until xAI publishes a path forward.
Is video input available on the chat surface or only the API?
xAI's public docs do not currently document native video-file input as a grok-4.3 API capability. Treat video understanding claims as unconfirmed unless they appear in official xAI API documentation or in-product release notes. For video generation, use the separately documented Grok Imagine API rather than grok-4.3.
How does Grok 4.3 compare to Claude Sonnet 4.6 or Gemini 3 Pro?
Artificial Analysis places Grok 4.3 just above Claude Sonnet 4.6 on the Intelligence Index (53 vs ~49). The bigger differentiator is cost: at $1.25/$2.50, Grok 4.3 undercuts most flagship-tier models on raw token pricing. Claude and Gemini may still be preferable for some workloads depending on recall behavior, multimodal requirements, enterprise controls, and benchmark results — avoid asserting vendor advantages without a cited, current benchmark source.
Is there a free trial for Grok 4.3?
There is no dedicated free trial. API users pay only for tokens consumed. Chat users can evaluate through whatever free or paid access is visible in the Grok picker; subscribing to SuperGrok is a paid evaluation, even if canceled before the next billing cycle. xAI has previously granted free-tier users intermittent access to flagship models, but availability is not guaranteed.
