Gemini 4 Neon: $3 Per Million Tokens, #3 Overall

    Google shipped Gemini 4 Neon on November 18, 2026, priced at $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million, which Google calls a 95% discount.

    The model sits below Gemini 4 Carbon and above Gemini 4 Hydrogen in Google’s lineup, carries a 1 million-token context window. And accepts text, images, audio, and video as input. Google’s own reporting puts it at #3 overall with a score of 61 on the Artificial Analysis Intelligence Index at its highest thinking level. And #1 on Zapier’s AutomationBench. It’s live now in the Gemini API in Google AI Studio, Gemini for business, Google Antigravity. And the Gemini app (Google’s announcement).

    That’s the datasheet. The parts that matter if you wire models into pipelines for a living are the cached input price, the four thinking levels. And the fact that Google deleted two other models the same day. The launch coverage is mostly repeating the benchmark table, so I’ll take the three things it skips.

    The Middle Slot Is Where Production Actually Runs

    Per GSMArena, Neon is the mid-tier model: below Carbon, above Hydrogen.

    Google pitches it at complex software engineering, knowledge work. And science workflows, and Indian Express gives the practical translation: coding, data analysis, research, and other multi-step tasks. That’s a workhorse description, not a research-lab one.

    The architectural claim buried in Google’s post matters more than the pitch.

    Neon is engineered for long-running agent workflows, running standalone or acting as an Orchestrator that directs parallel Hydrogen subagents inside Google Antigravity. Read that as intent: the mid-priced model plans, the cheaper model below it executes. For a one-person operation, that’s the only structure that pencils out. Because you can’t put a top-tier price tag on every tool call inside a loop.

    This is why the middle slot beats the top slot for real work.

    The biggest model wins demo videos. The middle model runs your scheduled jobs and your repo cleanup since it’s cheap enough to leave running.

    Google building orchestration into the pitch tells you exactly where they expect the volume: agents that run for hours, not chat turns.

    Every Number in the Benchmark Table Comes From Google

    The 61 score and #3 rank on the Artificial Analysis Intelligence Index are Google’s own numbers, measured at the xhigh thinking level. And Google itself concedes Neon lands “just behind larger, slower models.” The #1 finishes on Zapier’s AutomationBench (56.3%) and Andon Labs’ Vending-Bench 2 ($16,843), a long-horizon business simulation, also trace to the announcement post. InfoWorld repeats the economics claim with the right framing: the company says Neon completes complex work more economically than larger models.

    I don’t grade launches on the vendor’s own homework, and you shouldn’t either. One line in that table earns attention regardless: Neon ranks #2 on steerability on the Agent Arena Index, which Google defines as adapting cleanly to mid-task user corrections. That’s the metric that maps to actual agent work. An agent that absorbs a correction without restarting the whole task saves you more than two points of benchmark IQ ever will.

    My weighting for a small operation: steerability first, the long-horizon simulation second, the intelligence index last. The first two describe how a model behaves when a Tuesday goes sideways. The third describes a test neither of us runs.

    The $0.15 Cached Input Price Is the Actual Product

    The pricing detail that should drive your decision sits inside a parenthesis. Agents are loops, and every turn re-feeds context: your instructions, the codebase, the tool results so far. Re-feeding a full 1 million-token context costs $3 at the standard input rate and $0.15 cached. Multiply that across a long-horizon run with hundreds of turns. And the cache rate decides whether the workflow ships or gets killed for cost before it finishes.

    That 95% discount isn’t a courtesy, it’s the whole design.

    The four thinking levels are the second lever.

    InfoWorld describes them as four reasoning levels for adjusting computational effort to the workload. And Google says the levels “scale steadily” so spend matches the job. In practice, boilerplate generation runs cheap and the hard debugging session gets xhigh. A model with a spend dial is one you can leave inside a pipeline without watching the meter every hour.

    One caution: these are launch-day numbers from Google’s post. Treat them as current, not permanent, and re-check before you commit a quarterly budget to them.

    Two Models Died So Neon Could Launch

    The same rollout replaced Gemini 4 Argon in the Gemini app’s model picker, per ShiftDelete. And shut down Gemini 3.1 Pro entirely, per Indian Express. If anything you run was pinned to 3.1 Pro, you now have a migration project you didn’t schedule. That’s the standing tax on building against hosted models, and it gets collected whether your roadmap agreed or not.

    Access is broad if you’re already inside Google’s walls: the Gemini API in Google AI Studio, Gemini for business through the API and the model dropdown, Google Antigravity for paid AI Pro and Ultra subscribers. And the Gemini app for Pro and Ultra subscribers.

    Before you explore any of that, audit your own configs for the two dead models. That audit takes ten minutes and prevents the outage that takes a day.

    Verdict: Test the Loop, Not the Leaderboard

    Neon reads like a sensible default for agent work: $3 in, $15 out, a 95% cache discount, a spend dial.

    And a #2 steerability rank on the metric that matters most, all per numbers Google supplied. Notice the verb: reads. Vendor benchmarks tell you what a model does on its best day with its own tests. Your tool-heavy, correction-heavy Tuesday tells you what it does on the average one.

    Run one real task you already do this week through Neon at two different thinking levels, with caching on, and read the bill.

    Compare it against what your current model charges for the same loop.

    If the cached rate holds on your workload, you’ve found your new default.

    If it doesn’t, you’ve spent a dollar to dodge a bad migration.

    That’s the only benchmark that counts.

    Leave a Reply

    Your email address will not be published. Required fields are marked *