Grok 4.3 Is Now the Cheapest Frontier Model
xAI's April 30 release prices Grok 4.3 at $1.25 input and $2.50 output per million tokens — undercutting Gemini 3.1 Pro by 79% and GPT-5.5 by 92% on output, while landing between Opus 4.7 and Gemini on agentic Elo.
Grok 4.3 shipped on April 30 at $1.25 per million input tokens and $2.50 per million output tokens. That is a 40% input cut and a 58% output cut from Grok 4.20's $2/$6, and it makes Grok the cheapest frontier model on the helloai.com leaderboard on both axes simultaneously. Gemini 3.1 Pro sits at $12 per million output. Claude Opus 4.7 is $25. GPT-5.5 is $30.
The output number is the one that matters. At $2.50 per million tokens out, Grok 4.3 undercuts Gemini 3.1 Pro by 79%, Opus 4.7 by 90%, and GPT-5.5 by 92%. Any routing decision made two weeks ago that pointed to Gemini or a mid-tier model on cost grounds is now stale. For long-running agentic loops where output dwarfs input — code generation, multi-step tool use, video transcription with reasoning — the savings compound into something that changes the unit economics of entire products.
The benchmark story is competitive without being decisive. Grok 4.3 lands at roughly 1500 Elo on agentic composites, sandwiched between Opus 4.7 at 1503 and Gemini 3.1 Pro at 1493. The 16-agent Heavy mode carries over from 4.20, native video input is new, and the context window dropped from 2M to 1M — still matching the Claude, GPT, and Gemini tiers, but a real regression for anyone who built on the larger window. xAI is pushing all commercial API users to migrate, with no waitlist gating access.
The honest caveat is that xAI's benchmark transparency still trails Anthropic and Google, and Elo numbers at this stage of a release cycle are noisier than they look. Output quality at a 90%+ discount needs real production testing, not vendor charts — particularly for the output-heavy workloads where the savings are largest, because a small quality regression on every token undoes the price advantage fast. The right play this week is an A/B on a non-critical agentic workload, not a wholesale flagship swap.
What's interesting is the pressure this puts on the rest of the field. Google has room to cut Gemini 3.1 Pro and probably will, since their cost structure has always given them headroom they choose not to use. OpenAI just raised GPT-5.5 pricing and has nowhere to go but a cheaper variant or a quiet walkback. Anthropic, fresh off the Opus 4.6 cut, is the most boxed in — Opus 4.7 is the quality leader, but the price gap to Grok is now an order of magnitude, and that is the kind of gap that reshapes default choices regardless of who actually wins on benchmarks.