Grok 4.3 quietly cut prices 40-60%, then shipped a voice cloner. The pricing math is now hard to ignore.
xAI shipped Grok 4.3 on April 30 without much fanfare. New rates are $1.25 in, $2.50 out per million tokens, with a 1M context window and the fastest reasoning output speed on any frontier-tier leaderboard. The accompanying Custom Voices API gets less attention than it deserves. Here is what the numbers actually look like next to Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and the cheap-Chinese-frontier crowd.
Photo by Claudio Guglieri on Unsplash
What you pay for Grok 4.3
| Tier | Input / 1M | Output / 1M | Cached input |
|---|---|---|---|
| Standard (≤ 200K context) | $1.25 | $2.50 | $0.20 |
| Long context (> 200K) | $2.50 | $5.00 | $0.40 |
| Batch API discount | 20-50% off standard rates across input, output, cached, reasoning tokens | ||
Context window: 1M tokens. Tool surcharges: $5 per 1k Web/X Search calls, $10 per 1k File Attachments, $2.50 per 1k Collections Search calls. PerxAI docs.
The price cut, in context
Grok 4.20 launched at $2.00/$6.00 per million in March. Grok 4.3 lands at $1.25/$2.50. That is a 37.5% input cut and a 58% output cut, in just over a month, on a model that also gained measurable agentic ability. The blended price (3:1 input:output) drops from $3.00 to $1.56 per million, which is the second-cheapest blended rate in the top 10 of the Artificial Analysis Intelligence Index. Only Kimi K2.6 ($1.71) is in the same neighborhood, and Grok 4.3 outputs roughly 6x faster.
Stack it next to the closed-model field and the gap is wider than the spec sheet hints. On a workload that splits 3:1 input-to-output (the AA leaderboard's blended scenario), here is the per-million cost:
| Model | Input / 1M | Output / 1M | Blended (3:1) | Speed (tok/s) |
|---|---|---|---|---|
| Grok 4.3 | $1.25 | $2.50 | $1.56 | 192 |
| Kimi K2.6 | $0.60 | $2.50 | $1.07 | 33 |
| DeepSeek V4-Pro | $1.74 | $3.48 | $2.18 | 34 |
| Gemini 3.1 Pro | $2.00 | $12.00 | $4.50 | 121 |
| Claude Opus 4.7 | $5.00 | $25.00 | $10.00 | 51 |
| GPT-5.5 (high) | $5.00 | $30.00 | $11.25 | 61 |
| Grok 4.20 (predecessor) | $2.00 | $6.00 | $3.00 | 86 |
Speeds and blended prices from Artificial Analysis (snapshot 2026-05-02).
Two things stand out. Against Opus 4.7 you pay a quarter of the input price and a tenth of the output price. Against GPT-5.5 the gap is even wider on output, where Grok 4.3 is roughly 92% cheaper. The speed adds to the picture: at 192 tokens per second of generation, Opus 4.7 streams at about 27% of that rate and GPT-5.5 at 32%. Latency-sensitive applications get the cost win and the wall-clock win at the same time.
What it costs across four real workloads
Per-token rates are abstract. Monthly bills are not. Here is what each workload looks like at standard pricing, no caching, no batch.
| Workload | Grok 4.3 | Gemini 3.1 Pro | Opus 4.7 | GPT-5.5 |
|---|---|---|---|---|
| Coding agent: 5M input, 500K output / day | $37.50/mo | $240/mo | $1,125/mo | $1,200/mo |
| Daily research summary: 1M in, 100K out / day | $45/mo | $96/mo | $225/mo | $240/mo |
| Mid-volume RAG: 10M in, 1M out / day | $450/mo | $960/mo | $2,250/mo | $2,400/mo |
| High-throughput pipeline: 100M in, 10M out / day | $4,500/mo | $9,600/mo | $22,500/mo | $24,000/mo |
Coding agent row uses the typical 5M-input / 500K-output ratio that Cline and Aider reported in 2025. All numbers assume 30-day months. Plug your own ratios into the cost calculator.
On the high-throughput row, switching from GPT-5.5 to Grok 4.3 is the difference between a $24K/month bill and a $4.5K/month bill. At those volumes, the question stops being "which model do I prefer" and starts being "does the quality delta justify a $19.5K/month tax." For most applications outside reasoning-heavy science work, the answer is now: probably not.
Where Grok 4.3 actually lands on the scoreboard
xAI's announcement is brief on benchmarks. Independent measurement from Artificial Analysis, Vals AI, and Andon Labs gives a clearer picture, including the regressions.
| Benchmark | Grok 4.3 | Grok 4.20 | Read |
|---|---|---|---|
| AA Intelligence Index v4.0 | 53 (#10) | 47 | Mid-pack frontier; behind GPT-5.5 (60), Opus 4.7 (57), Gemini 3.1 Pro (57) |
| GDPval-AA (agentic) | 1500 Elo | 1179 | +321 Elo gain; tops Gemini 3.1 Pro, Muse Spark, Kimi K2.5 |
| tau-2 Bench Telecom | 98% | 93% | +5 pts; matches GLM-5.1 |
| IFBench (instruction follow) | 81% | 81% | Flat |
| Vals AI CaseLaw v2 | 79.3% (#1) | ~54% | +25 pts; legal-domain leader |
| ProofBench (math) | 11% | n/d | Weak; flagged as a gap area |
| Vending-Bench 2 | regression | stronger | Andon Labs: 'narcolepsy problems' on long-running sim |
| AA-Omniscience (hallucination) | -8 pts | leader | Grok 4.20 still beats Grok 4.3 on non-hallucination |
Sources: Artificial Analysis, OfficeChai breakdown.
The agentic jump is the headline most reviews are missing. A 321-point Elo gain on GDPval is not a small thing; it puts Grok 4.3 above Gemini 3.1 Pro Preview on agentic workloads despite costing roughly a third as much. Bindu Reddy at Abacus AI posted on X that the model "is as smart as Sonnet 4.6 and 5x cheaper and faster." That is roughly correct on the leaderboards we've checked: AA Intelligence Index has them within 1 point (53 vs 52 max), and Sonnet 4.6 max blended price is $6.00 vs Grok 4.3's $1.56.
The regressions matter too. Andon Labs measured a degradation on Vending-Bench 2, their long-running business-simulation eval, and described it as Grok 4.3 preferring inactivity over taking actions for multiple in-sim days. ProofBench math at 11% is below most reasoning-tier competitors, and the AA-Omniscience non-hallucination score moved 8 points the wrong way. If your workload requires careful, conservative reasoning over long horizons, run an eval before switching.
One operational note: Grok 4.3 is verbose. Running the full AA Intelligence Index costs $395.17 on Grok 4.3 vs a median of around $295 for similar-tier models, and it emits 88M output tokens to do so vs a 35M median. The per-token rate is cheap; the per-eval bill is not as cheap as the per-token rate suggests.
We ran a quick check against our calculator with the AA verbosity numbers plugged in. At 2.5x the median output count, the headline 92% output savings vs GPT-5.5 on a per-token basis collapses to closer to 80% on a per-completed-task basis. Still a real win, just not as dramatic as the rate-card math implies. Worth measuring on your own prompts before committing.
Custom Voices is the underreported part of this release
The Custom Voices launch alongside Grok 4.3 is getting less attention than the price cut, which is a mistake. xAI lets you clone a voice from 120 seconds of reference audio. The clone picks up delivery patterns - the example xAI demoed was a customer-support inflection that survived the synthesis. Up to 30 voices per team, deletable in one click, with an 8-character voice ID for API use.
Pricing for the voice surface is structured separately from chat:
- - Voice Agent API (grok-voice-think-fast-1.0): $3.00/hour, or $0.05/min, speech-to-speech
- - TTS: $4.20 per 1M characters across 5 base voices (Eve, Ara, Rex, Sal, Leo)
- - STT: $0.10/hour for REST batch, $0.20/hour for streaming
Two important caveats. The programmatic voice-clone API is gated to Enterprise customers (web playground access is available to all paid tiers). And xAI excludes Illinois from voice-clone availability outright, due to the Biometric Information Privacy Act. If you operate at scale and have any Illinois users, that geofence is a real product constraint, not a footnote.
Compared to ElevenLabs Pro at roughly $0.18 per minute of generated audio, $3/hour ($0.05/min) speech-to-speech with reasoning baked in is materially cheaper. If voice-agent workloads are anywhere on your roadmap, this changes the math.
Use it for this. Skip it for that.
For high-throughput coding agents, content generation pipelines, function-calling backends, and anywhere output speed matters, Grok 4.3 is now the cost-quality leader that's not from China. The 192 tok/s output rate is the highest of any top-15 model on the AA leaderboard, and the price is roughly half of any closed competitor in the same intelligence neighborhood. The 1M context, the batch discount, and the cached-input rate combine to make this a real production option, not just a chat alternative.
Legal and corporate-finance workloads are a surprise category. The Vals AI rankings put Grok 4.3 at #1 on both CaseLaw v2 and CorpFin, which is not where most generalist models sit, and at $1.25/M input it's cheaper than the legaltech-specialized models that charge $20-50/M.
Where it falls short: long-horizon agentic work that requires patience and follow-through (per Andon Labs), formal mathematics (the 11% ProofBench is real), and applications that need rigorous non-hallucination guarantees. If you previously chose Grok 4.20 specifically for its honesty profile on AA-Omniscience, the switch to 4.3 is a measurable regression there. And if you need video understanding, Grok 4.3 is still text+image only despite some pre-release coverage suggesting otherwise.
Anthropic-stack agents that need computer use, or applications already tied to OpenAI's ecosystem (Realtime API, Files API, Assistants) won't see their calculus change here. Switching means rewriting integrations, not just changing a base URL. The price gap may eventually justify that work. Today it might not.
Distribution, SuperGrok tiers, and what is gated to enterprise
The model is available on the xAI API direct, on OpenRouter, and inside SuperGrok consumer plans. SuperGrok Lite is $10/month, SuperGrok is $30/month or $300/year, and SuperGrok Heavy is $300/month for power users. X Premium+ at $40/month also gets API-tier access, with 50% off the first two months still running as of May 2.
What stays gated: the programmatic voice-clone API requires Enterprise. Higher rate limits beyond the default 1,800 RPM and 10M TPM also require Enterprise. The Custom Voices web playground is available to anyone with a paid account.
SOC 2 Type II, HIPAA-eligible, and GDPR-compliant per the xAI compliance page, which opens the door for healthcare and regulated-industry workloads that previously had to use Anthropic on Bedrock or OpenAI on Azure.
What this means for the rest of the market
For anyone tracking the LLM API price curve, Grok 4.3 is the third major price reset this calendar year. DeepSeek V4 reset the open-weights frontier in late April. Kimi K2.6 reset the cheap-API frontier earlier in the month. Grok 4.3 has now reset the mid-priced closed-API tier where most production workloads sit. Three cuts in roughly five weeks, all from non-Anthropic, non-OpenAI players.
Anthropic and OpenAI have not matched any of these cuts on their flagship models. Both companies are shipping incremental capability gains (extended thinking budgets, tools, agents) and counting on capability lock-in to defend the price gap. Whether that holds depends on whether teams keep reaching for Opus 4.7 and GPT-5.5 by default once the cost difference is measured in dollars per workload, not cents per token.
The HN thread on Grok 4.3 hit 386 points and 515 comments in 22 hours, which is unusually high engagement for a model release that didn't come with a flashy demo. The signal is loud: people are paying attention to the price curve, not just the benchmark curve.
Sources
- - xAI model docs (pricing, context, batch): docs.x.ai/docs/models/grok-4.3
- - xAI tools and surcharges: docs.x.ai/docs/models
- - Artificial Analysis Grok 4.3 page: artificialanalysis.ai/models/grok-4-3
- - AA leaderboard (competitors snapshot): artificialanalysis.ai/leaderboards/models
- - OpenRouter listing (tiered pricing confirmation): openrouter.ai/x-ai/grok-4.3
- - VentureBeat launch coverage (May 1): venturebeat.com
- - OfficeChai benchmark deep-dive: officechai.com
- - HN discussion (386 points): news.ycombinator.com