xAI Grok 4.3 — always-on reasoning, 1M tokenow, ceny -50%

xAI has released Grok 4.3 — a model with an always-on reasoning architecture, 1M token context window, and API prices cut by more than 50%. Launched on May 1, 2026, the release also includes Custom Voices — a voice cloning tool based on a 2-minute sample, currently available in the US only.

Key takeaways

Grok 4.3 applies reasoning to all responses — there is no option to disable reasoning mode
Context: up to 1M tokens; text and image inputs; new agentic capabilities and file creation tools
API pricing: $1.25/1M input tokens, $2.50/1M output tokens — but a new "reasoning tokens" charge applies
Artificial Analysis Intelligence Index ranking: 53 points (7th place); 7 points behind GPT-5.5 (1st place)
Custom Voices: voice cloning from a 2-minute sample, TTS and voice agent integration; US only

Architecture: what 'always-on reasoning' means

Previous reasoning-capable models (OpenAI o-series, Gemini 'Deep Think') let users decide when to activate step-by-step reasoning. Grok 4.3 removes this toggle — the model always processes queries through an internal chain of thought, even for simple questions.

This has consequences. On the positive side: the model is more consistent and less prone to context-omission errors. On the negative side: higher cost per query (reasoning tokens billed separately) and increased latency. Independent tests recorded a 'narcolepsy' phenomenon — the model entered reasoning loops on simple tasks, generating excessive delays.

The model supports text and image inputs, up to 1M token context, code generation and execution, file creation, and video upload. xAI positions Grok 4.3 as a tool for "agentic digital tasks" — not as a general-purpose chatbot.

Pricing: headline cuts, but a new charge type

Grok 4.3 is priced at $1.25 per 1M input tokens and $2.50 per 1M output tokens — approximately 37.5% and 58% lower than previous xAI models, respectively. Compared to the broader market, this places it among the more affordable reasoning-capable models.

However, xAI introduces a new category: "reasoning tokens" — tokens generated during the model's internal thinking process, billed separately. In practice, every query generates additional invisible tokens that increase the real cost. The net budget impact for developers is difficult to estimate without task-specific benchmarks.

Benchmark results: strengths and weaknesses

The Artificial Analysis Intelligence Index (an independent benchmark aggregator) gave Grok 4.3 a score of 53, placing it 7th. This is 4 points above the previous version (Grok 4.20), but 7 points below the leader — OpenAI's GPT-5.5.

Grok 4.3 outperforms Meta's Muse Spark and several other models, but trails Chinese models — Kimi K2.6 and Mimo-2.5 Pro. It performs particularly well on legal and financial tasks. Weaknesses: coding and complex mathematics — areas where the model consistently lags behind segment leaders.

Custom Voices: voice cloning and regulatory questions

Alongside Grok 4.3, xAI announced Custom Voices — a function for cloning a user's voice from a 2-minute audio sample. It integrates with TTS (text-to-speech) services and voice agents. Due to data privacy regulations, Custom Voices is currently available in the US only.

xAI has not disclosed technical details of the voice model architecture or safeguards against misuse (e.g., cloning the voices of public figures without consent). This is a sensitive regulatory area — similar features on other platforms have faced investigations in Europe and the US.

Why this matters

Grok 4.3 is an important signal for the reasoning model market. The 'always-on reasoning' architecture is a bet that consistent precision matters more than flexibility and cost. This is a clear differentiation from OpenAI's and Google's approaches, which preserve a reasoning toggle. If xAI is right, competitors will need to consider a similar move.

However, the model clearly targets a niche: long documents, legal and financial tasks, enterprise agentic workflows. As a general-purpose chatbot — it is not optimized. Reasoning tokens as a separate pricing line item is a novelty that may deter developers seeking predictable costs.

What's next?

xAI announced bi-weekly model updates — a faster release cadence than most competitors
Custom Voices will be evaluated for expansion to other markets — requiring regulatory localization (GDPR, AI Act)
Weaknesses in coding and mathematics are critical for agentic segment positioning — if not addressed, Grok 4.3 will remain a niche analytical tool