Google ships a frontier model every six days — cheaper each time

Google released Gemini 4 Neon on Wednesday, the fourth Gemini 4 model in seven weeks, and the gap between releases has collapsed from 28 days to six. The benchmark score is the boring part. What is new is that each release costs less per task than the one before it.
Google announced Gemini 4 Neon on November 18 at $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 — a 95% discount — and a one-million-token context window, according to Google's launch post. That price sits below Gemini 4 Carbon, which lists at $4 in and $20 out; Carbon's discounted introductory rate was $2 and $10, which means Neon is launch-priced at roughly Carbon's promo rate without the promo — the pattern we noted when Gemini 4 Hydrogen shipped at $1.50 and $7.50 six days ago. Availability is immediate on the API in Google AI Studio and Gemini for business, in Google Antigravity for paid AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers. A cyber-tuned variant goes to vetted defenders through the Fairwind program over the next few days.
The capability claims are credible but not a leap. Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, third overall behind GPT-6.2 Astra and Claude Fable 5.2 at their maximum settings and one point ahead of Carbon, per Artificial Analysis. On Vals' enterprise benchmark it ranks second at 70.9% and takes first on the Legal Research and Tax Agent benches. Google also claims the top spot on the Artificial Analysis Coding Agent Index at 58.6, 65.2% on Terminal-Bench 4.1, and 46.2% on FrontierCode 2.0 at the default high setting. Arena's new Agent Arena Index, built from more than 4,200 real agent sessions, puts it second at 1,336 points and second on steerability — the measure of whether a model adjusts when a user corrects it mid-task.
Then there is the number that explains the cadence. On SWE-sweep, Neon scores 12.8% under the open mini-SWE harness and 39.0% inside Antigravity — the same weights, roughly three times the score, with Google attributing the difference to the harness rather than the model. Read that alongside the release intervals Arena tracked: Gemini 4 Argon on September 30, Carbon 28 days later, Hydrogen 15 days after that, Neon six days after Hydrogen. If a meaningful slice of each successive jump comes from scaffolding, tooling and context management rather than new pretraining, shipping a new model every six days stops looking like a research miracle and starts looking like a product cadence — and it explains why the price falls each time.
Google's own efficiency numbers point the same way: at the default high setting Neon costs 32% less per task than Carbon, sits on the Pareto frontier on both Vals and Agent Arena, and runs 50% to 80% cheaper per task than top competing frontier models; at xhigh it still costs at least 36% less than rivals at their highest settings. Artificial Analysis independently measured about 17% lower cost per task than Carbon, a smaller gap than Google's figure but the same direction.
The safety side is mostly continuity. Neon refuses 1.98% of Terminal-Bench 4.0 tasks and none on Terminal-Bench Science, takes shortcuts on 2.45% of Google's internal coding cheating evaluation — the lowest of any Gemini 4 model — scores 0.09% on the company's misaligned-action suite, passes 99.2% of its agent honesty evaluation, and resisted every attack in Gray Swan's indirect prompt-injection benchmark when real-time safeguards were on at the highest thinking level. Those are vendor-run evaluations, and the harness effect above is a reminder to treat benchmark deltas as claims about a whole system, not a model.
Early user reports are mixed in the way new releases usually are. One developer ran the same prompt through Neon and Claude Opus 5.5: Neon finished in about 20 minutes for $3.50, Opus in 18 minutes for $7.20, and the developer judged Neon's output the best of the Gemini 4 line while still visibly behind Opus, according to a hands-on writeup at Zhidx. Others complained that the elemental naming — Argon, Carbon, Hydrogen, Neon — gives no clue which model to pick. That is a real cost: a cheap, fast model nobody can identify is a discount nobody redeems.
Intel signed its first outside customer for its 18A manufacturing process, CEO Lip-Bu Tan said at the TomorrowX Summit in Austin on Wednesday, without naming the buyer, CNBC reported. Intel's technology chief Pushkar Ranade clarified on X that the customer signed for 18A-P, the variant that entered risk production in June. Tan also said Microsoft is the first publicly named customer for Intel's most advanced packaging technology, EMIB, a deal tied to Azure infrastructure.
The number that matters more than the logo is yield. Crypto Briefing reports 18A yields have reached 70%, up from what Tan inherited in March 2025, when they were described internally as horrible. A 70% yield on a leading-edge node is the difference between a science project and a product line, and it is the reason Intel can now sell the node externally at all — 18A was originally built for Intel's own chips, with outside customers steered toward 14A, which is not expected to take external work until 2028 or 2029.
The context is ugly enough to keep the milestone in proportion. Intel's stock has lost more than half its value since the end of September, most of the fall coming after the company guided fourth-quarter revenue at least $1 billion below analyst estimates. Analysts including Creative Strategies' Ben Bajarin have speculated the unnamed customer could be Nvidia, Apple or a hyperscaler short of TSMC capacity; reports in May had Apple and Intel closing in on a foundry deal. An unnamed 18A-P customer proves someone signed, not that anyone meaningful is shipping volume — Intel disclosed no volume commitments, no revenue impact and no identity, and a customer that cannot be named cannot be sized.
What to watch: whether Google can keep a six-day cadence without the quality wobbles that usually follow, and whether Intel names the 18A customer before its next earnings call.
Would you pick a cheaper model with a smaller benchmark gap, or pay for the leader? Write to us via our contact page.







