cd /news/large-language-models/gemini-4-neon-price-benchmarks-and-w… · home › topics › large-language-models › article
[ARTICLE · art-172458] src=digitalapplied.com ↗ pub= topic=large-language-models verified=true sentiment=· neutral

Gemini 4 Neon: Price, Benchmarks and When to Use It

Google released Gemini 4 Neon on November 18 at $3 per million input tokens and $15 per million output tokens, with Artificial Analysis scoring the model 60 on its Intelligence Index at the default high thinking setting and 61 at extra high on November 19. The extra-high setting added nearly seven minutes per task on the same index while returning a first answer token about 16 seconds later than high, and Google bills thinking tokens at the output rate. Neon is available through the `gemini-4-neon` model identifier in the Interactions API or `generateContent`, with a 1,048,576-token shared context window on the direct API versus up to 512K in Antigravity, where Neon becomes the default for new conversations on eligible paid plans.

key takeaway

Gemini 4 Neon scores 60 on the Intelligence Index at its default high thinking level versus 61 at extra high, which adds nearly seven minutes per task, so start at high and reserve extra high for failures you can describe and verify.

read10 min views2 publishedNov 19, 2026
Gemini 4 Neon: Price, Benchmarks and When to Use It
Image: Digitalapplied (auto-discovered)

Gemini 4 Neon makes a persuasive case for leaving the thinking level alone. On November 19, Artificial Analysis scored Google’s new model 60 at its default high setting and 61 at extra high. That last point added nearly seven minutes per task on the same index. Start with high for a difficult code change. Save extra high for a failure you can describe and a result you can check.

Google released Neon on November 18, pitching it for the work between a routine edit and a problem that warrants its most capable Gemini model. At $3 per million input tokens and $15 per million output tokens, its strongest case is a job that takes several attempts on a cheaper model, or one where a reviewer spends more time tracing the evidence than checking the answer. For a team considering a switch, the first decision is where that work will run. Google’s launch announcement

  1. 01High is a sensible first trial.It came within one Intelligence Index point of extra high and returned a first answer token about 16 seconds sooner in the independent measurement.
  2. 02Price all the thinking.Google bills thinking at the output rate. A short final answer can still be the expensive part of a long review.
  3. 03Choose the surface before sizing the input.The direct API has a 1,048,576-token shared window. Antigravity offers up to 512K and its account entitlements differ from the API.
  4. 04Choose a test your reviewer already trusts.A benchmark can suggest the right task. A passing regression check and a reviewable diff can tell you whether the model earned the work.

01 — AvailabilityWhich account gets Neon? #

For an application, the direct model identifier is gemini-4-neon. A paid Gemini API project can use it through either the Interactions API or generateContent. For someone trying prompts in Google AI Studio, Google lists its AI Pro and Ultra plans. Treat those as separate purchases when budgeting. The price of using a model at your desk tells you little about the bill for an application calling it on behalf of customers. Google’s access and migration guide Antigravity gives this choice another dimension. The Antigravity team says Neon becomes the default for new conversations on eligible plans across its app, CLI and IDE extensions. Paid Google AI Pro and Ultra include Neon and Orchestrator (Neon). Pro trial access depends on having an eligible premium device. Enterprise users get Neon through Gemini for business, with Orchestrator excluded. Free and Google AI Plus users have access to neither option. Antigravity’s rollout and plan table

When a colleague says they used Neon to fix a bug, ask which of those modes they selected. In Orchestrator, Neon plans and reviews the work while Hydrogen implements changes and runs tests. If Hydrogen gets stuck, Neon can reconsider the plan. A team that wants to compare Neon with its current model should decide whether it is buying that whole workflow or choosing the model for an existing tool. Each is a valid experiment, with a different result to take back to procurement.

02 — Token costsWhat does one review cost? #

Google’s standard paid rates apply across the full context window. Uncached input is $3 per million tokens, implicitly cached input is $0.15, and output including thinking is $15. Batch is cheaper for jobs that can wait. Its respective input, cached-input and output prices are $1.50, $0.075 and $7.50 per million. Google lists no free API tier for Neon. Gemini Developer API pricing

Suppose a code review consumes 80,000 uncached input tokens across its calls and produces 6,000 billed output tokens, including thinking. The input costs $0.24 and the output $0.09, or $0.33 in total before tool charges. Put the same token volumes through Batch and the token bill is $0.165. A hundred such reviews would be $33 or $16.50. These are worked examples at list price. Your tool transcripts, retries and review comments all change what the next run consumes.

Suppose instead the standard review gets a verified cache hit on 60,000 of its 80,000 input tokens. Those hits cost $0.009. The other 20,000 cost $0.06 and output still costs $0.09, giving $0.159 total. Budget at the uncached rate until you have evidence of reuse. Neon’s cache is implicit. Google documents a minimum prompt length of 4,096 tokens for eligibility, and its explicit cache creation interface is unavailable for this model. Google’s caching guidance

Keep web grounding on its own line. For Google Search, the listed allowance is 5,000 requests per month shared across Gemini 3 and newer models, then $14 per thousand queries, subject to a 1,500-per-day cap. One prompt can trigger several billable queries. Maps grounding also lists 5,000 per month across those models, followed by $14 per thousand. A repository review that uses only local files has a different cost profile from a review that also looks up the documentation for every dependency.

03 — Waiting timeWhat changes at each thinking level? #

The useful part of the Artificial Analysis results is seeing time and quality together. The table shows the same model on the same day, November 19, using the Google API. The score is the Artificial Analysis Intelligence Index v5.0.1. First answer token is how long it took before an answer began. Output speed divides all generated tokens, including reasoning, by the full request duration. The last column is the evaluator’s time per index task, a different measurement again. Artificial Analysis measurements and effort breakdown

Gemini 4 Neon on November 19, 2026. Source: Artificial Analysis.
Effort Index score First answer token Output tokens/s Minutes per task
--- --- --- --- ---
Low 44 13.93 s 97 3.7
Medium 55 20.14 s 101 8.3
High (default) 60 40.64 s 101 14.8
Extra high 61 56.51 s 110 21.5

Notice the medium and high rows. Both average 101 generated tokens per second end to end in this snapshot, yet the user waits roughly twice as long for high to begin answering. For a developer waiting at a terminal, record time to a usable patch. For an overnight job, completion before the review meeting may matter more than the first word appearing quickly.

04 — Coding evidenceRead the score alongside the tools #

Google’s November 18 launch post cites 58.6 on the Artificial Analysis Coding Agent Index. The run used Antigravity CLI with Neon at extra high. Google also reports 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0, specifying default high for the latter. The first figure is an index score. The others come from distinct coding benchmarks. A raw percentage from one suite cannot tell you that Neon is better at one kind of engineering than another. The suites ask different questions and award credit under different conditions. Google’s reported coding results · Coding Agent Index v2.1

The most instructive pair in the release comes from one benchmark. Google reports 12.8% on SWE-sweep with the mini-SWE harness and 39.0% inside Antigravity. Here the model name and benchmark stay fixed while the software directing the agent changes. The result gives a reason to test your own setup. It cannot tell you which tool, instruction or retry accounted for the difference. If your integration hands Neon a file excerpt and gives it one attempt, the Antigravity number is a poor basis for forecasting how many fixes it will finish.

Keep the broader claims in their original setting as well. Google cites 70.9% on the Vals Index for professional work and 56.3% on Zapier’s AutomationBench. Those figures may justify giving Neon a workbook or a multistep task to try. A team whose automation must preserve a particular invoice field or date format should include that requirement in its own acceptance test. The benchmark never saw your field or its downstream consumer.

05 — CapacityOne window for input and output #

The direct API has a unified window of 1,048,576 tokens shared by the input, thinking and answer. A default synchronous response allows up to 65,536 thinking and output tokens. For example, a prompt that already occupies 1,000,000 tokens leaves at most 48,576 for the remaining work. Raising the output setting cannot recover space already occupied by the input. Antigravity has its own limit of up to 512K, so a successful direct API experiment with a large document set can still need trimming when moved to the IDE. Gemini API model specifications

Long decode is for work that genuinely needs a larger generation. In the Interactions API, background execution can run several continuation steps and lets the application poll for a combined result. Google also documents client-managed continuation. An ordinary one-call streaming request still has a 65,536-token hop limit. Allow room for thinking even when the final document seems comfortably below the limit. How long decode handles continuation

OpenRouter uses google/gemini-4-neon. As of November 19 its endpoint metadata listed 943,718 maximum completion tokens on several routes and 65,536 on another. Those describe particular routes. For a direct Google integration, size the request against Google’s synchronous or long-decode contract. For a routed integration, record the chosen endpoint with the result so a provider fallback can be investigated later. OpenRouter endpoint metadata

06 — Source filesGive Neon the evidence it needs #

Neon accepts text, images, audio, video and PDF and returns text. The November 18 Gemini API update also added direct processing for Office and OpenDocument files. A request can select pages, worksheet tabs or cells, time intervals and particular layers of content, including formulas. That matters when the defect lives in a relationship between two files. A cropped screenshot can show that the total is wrong while hiding the formula that produced it. Gemini API release notes for November 18

Imagine an export that shows a different total in a workbook and in the web app. Give Neon the failing export function, the affected worksheet range including its formulas, and the part of a screen recording where the total changes. Ask for the first cell at which the totals diverge and a regression test that demonstrates the difference. A reviewer can check that cell against the workbook. Including more sheets becomes worthwhile if the formula points outside the selected range. Google’s document and spreadsheet inspection options

If the job depends on a recently changed library, supply its current reference too. DeepMind gives March 2026 as Neon’s knowledge cutoff. An agent with fresh files or a search tool can use information published later. For a code review, keep the exact installed version with the evidence, so the reviewer can distinguish an obsolete API suggestion from a bug in the application. DeepMind’s model card and limitations

07 — AdoptionWhich work should move first? #

Take one recently closed issue where a cheaper model needed help. Recreate the starting commit, preserve the tools your team allows, and write down what the reviewer accepted. Run Neon at high and the current model against that starting point. Include the failing example in both prompts. Keep a second example out of the prompts and use it to check whether the fix generalises. A patch that passes the visible case but guesses the wrong rounding rule can look excellent in a transcript.

Next, compare the cost and elapsed time of getting to an accepted result. Record requests, cache hits, billed output, tool calls, retries and any minutes a reviewer spent reconstructing unsupported claims. If neither patch passes, try changing the evidence before paying for more thinking. If the evidence is sufficient and high still misses the same constraint, try extra high on that case. This gives the person choosing a default a concrete exception rule to apply next time.

If your current default is already a Gemini model, the Neon, Hydrogen and Carbon comparison covers the family decision. For a different API provider, use the comparison with GPT-6.2 Sol and Claude Sonnet 5.5 to price the same input. The Gemini 4 Hydrogen guide gives a starting point for teams with faster, routine coding tasks. Teams commissioning an evaluation through our AI transformation service can use the same starting issue and acceptance criteria.

Choose a default your reviewer can defend

Use Neon at high when a job crosses files, tools or media and someone can verify the result. Keep a lower effort for bounded work with a quick acceptance check. Give extra high a repeatable failure to solve before making anyone wait for it on every request.

── more in #large-language-models 4 stories · sorted by recency
── more on @google 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/gemini-4-neon-price-…] indexed:0 read:10min 2026-11-19 · —