Hero title card reading Gemini 4 Neon, live 18 November 2026, with a subtitle reading five request-level changes, and a footer line reading pricing and thinking levels per Google's Gemini API documentation, checked 20 November 2026
Engineering & Research

Gemini 4 Neon Is Live: Five API Changes to Plan For

Author

Gideon Frost

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

G​emini 4 Neon went generally available on 18 November 2026, and the interesting document is not the launch post. It is the migration checklist G​oogle publishes underneath it, which is where you find out that moving production traffic onto G​emini 4 Neon from Gemini 3.8 Flash, or from Gemini 4 Carbon or G​emini 4 Hydrogen, changes more than a model string. Five request-level behaviours change, one of them silently, one of them with a 400, and one of them by a factor you will only notice on the bill. G​oogle's summary of the model itself is modest: frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work and science, sitting between G​emini 4 Hydrogen and Gemini 4 Carbon for high-frequency complex tasks. What follows is the part of the rollout that costs engineering time rather than tokens.

What actually shipped on 18 November

The rate card is flat and does not move: $3.00 per million input tokens, $15.00 per million output tokens, $0.15 per million cached input, paid tier only, no free quota anywhere. Context is one unified 1,048,576-token window shared by input, thinking and output, with a 65,536-token synchronous output cap and a long-decode path back up to a million. There are four thinking levels — low, medium, high and xhigh — and the default is high, not medium. The model is on the paid Gem​ini API across both generateContent and the Interactions API, in AI Studio for AI Pro and Ultra subscribers, and inside the Antigravity agent in Gem​ini Managed Agents.

Two numbers from outside G​oogle are worth carrying into the migration decision. Artificial Analysis scores G​emini 4 Neon at 61.4 on its Intelligence Index at xhigh, which the tracker prints as 61 and ranks sixth of 95 models, and it measures 110 output tokens per second against Gemini 4 Carbon's 51. The same panel also records $6.44 per Intelligence Index task at xhigh and 283 million output tokens across the evaluation set, which is the detail that matters here: this is a model that reaches its scores by emitting a lot of tokens, and every one of them is billed.

A self-built card titled The five changes, as a checklist, with five rows reading sampling parameters accepted then ignored inside range and a 400 outside it, explicit caching removed in favour of implicit caching on prompts of 4096 tokens or more, thinking default high instead of medium with minimal rejected, state carried by previous_interaction_id with thought signatures preserved, and outputs past 65536 tokens requiring background execution with max_output_tokens raised

Sampling parameters: ignored, or a 400

G​oogle's migration notes tell you to remove temperature, top_p and top_k from requests. The exact behaviour is the part to design around: in-range custom values are accepted and then ignored by the server, while out-of-range values return a 400 validation error. So a service that has been quietly passing temperature 1.3 for eighteen months does not fail on the way in — it starts returning the same distribution it would have returned at any other setting, and the only symptom is that your outputs get more deterministic-looking in some places and less in others, with no error to correlate against.

The practical fix is not to clamp the values. It is to delete them and move whatever you were using them for somewhere the model actually honours it: structured outputs for shape, an evaluation suite for variance, and prompt-level changes for style. Anyone relying on temperature as a creativity dial now has no dial.

Explicit caching is gone, implicit is automatic

G​emini 4 Neon does not support explicit context caching. The cachedContents resource and caches.create are not available on it, and requests carrying a cached_content field have to lose it. What replaces that is implicit caching, which runs by default on prompts of at least 4,096 tokens and discounts cached input by 95% — $0.15 per million instead of $3.00 — with no handle to create, expire or pay storage on.

For most teams this is a simplification and a small bill reduction. For a few it is a regression, and the shape of the regression is predictable: if your cache hits depended on a large fixed prefix being registered once and reused across many differently-worded requests, implicit caching only helps when the prefix actually matches, so the winning move is to make the shared prefix byte-identical and put everything variable after it. If your cached prefix is under 4,096 tokens, you no longer have caching at all.

minimal is rejected and high is the default

The thinking default is the quietest change in the list. Gemini 4 Carbon and Gemini 4 Argon default to medium and stop at high. G​emini 4 Neon defaults to high and adds xhigh above it, and it rejects minimal with an error — the level that Gemini 3.6 Flash and the Flash-Lite models still accept. Code that never sets thinking_level therefore gets a deeper, slower, more expensive answer on Neon than it got on Carbon with the same request body.

G​oogle's own guidance is to evaluate at all four levels with a high max_output_tokens so the cap does not truncate the answer, then run at the lowest level that clears the bar. On a model whose published independent score spans 17 index points from its cheapest rung to its deepest, that is not boilerplate. The default you inherit is the most expensive correct setting, not the cheapest one.

State, signatures, and a new interaction ID

Multi-turn and tool-calling sessions move from "resend the transcript" to two specific mechanics. In the Interactions API you carry state with previous_interaction_id rather than rebuilding history. In client-managed generateContent histories you return prior model turns and their thought signatures unchanged on every turn, and you must not copy a signature across parallel functionCall parts — the signature is bound to the call it came from.

There is also a failure mode worth knowing before it appears in production: prompting the model through system instructions to emit raw-text JSON tool calls, with zero tools declared natively, can return MALFORMED_FUNCTION_CALL. Declare tools with function calling, or constrain the response with structured outputs. This one is easy to miss because it is not a schema change — it is a prompt pattern that used to work.

Long outputs need background execution

Synchronous responses stop at 65,536 tokens, streaming or not. To go past that, you set max_output_tokens above 65,536 and set background to true, and the server continues in 64K hops until the window is exhausted. Set background without raising max_output_tokens and the job still stops at 65,536, which is the exact mistake the docs call out in a note.

The second half of this change is in your HTTP client, not your prompt. G​oogle tells you to raise timeouts to at least fifteen minutes, because hard prompts on this model can run for several minutes before the first answer token. The independent measurement agrees with the direction: Artificial Analysis records 56.5 seconds to first answer token and 21.5 minutes per Intelligence Index task for Neon, against 45.96 seconds and 28.4 minutes for Gemini 4 Argon. Faster overall, slower to start.

Screenshot of Google's What's new in Gemini 4 Neon documentation page, showing the model table with the gemini-4-neon model ID, thinking levels low, medium, high and xhigh, and the paid tier price of 3 dollars per million input tokens and 15 dollars per million output tokens

What the migration costs before it saves anything

Billing follows every token the model processes, including thinking tokens and, when you use grounding, the internal system instructions and function declarations. That is standard for the family but it interacts badly with two Neon specifics: the high default burns more thinking tokens than the medium default your Carbon code was written against, and the model's own token profile is heavy. The tracker counted 283 million output tokens for Neon across the Intelligence Index against 205 million for Carbon — 38% more, at a time when the per-token price is also higher than Carbon's introductory rate.

None of that argues against moving. It argues against moving by changing one string and watching the dashboards. The honest sequence is: pin the old model, duplicate a slice of traffic onto gemini-4-neon at medium and at high, and compare cost per completed task rather than cost per token. Cost per token on this model is the least informative number in the sheet.

Screenshot of the Gemini API pricing page for Gemini 4 Neon showing the Standard tab with an input price of 3 dollars and an output price of 15 dollars per million tokens, context caching at 0.15 dollars with a note that only implicit caching is supported, a free tier reading not available, and a callout stating that every token including thinking tokens is billed

Where a router earns its keep in this migration

This is a textbook canary rollout, and it is the reason a routing layer is useful even for a single-vendor migration. One API key covers 200+ models, so the old path and the new one sit behind the same endpoint while you move a percentage of traffic; automatic failover means a 400 from an out-of-range parameter or a timeout on a long-running task falls back to the model that was working yesterday instead of paging someone. Prices are passed through at provider list price with no markup, so Neon's $3.00 and $15.00 are the same numbers on your invoice as on G​oogle's page, and any change G​oogle makes to them lands on our side the same day.

One thing to be explicit about: G​emini 4 Neon is not in our catalogue yet. Gemini 4 Carbon and Gemini 4 Argon are, and so is gemini-3.1-pro-preview, and we are not going to describe a rate card as an endpoint. What the router gives you today is a stable place to keep the fallback chain while you evaluate Neon on G​oogle's own API, and the same request shape when it lands.

Who should move now

Move now if your workload is repository-level debugging, multi-file refactoring or a long agent session with many tool calls — the things G​oogle built Neon for, and the things the independent wall-clock numbers favour. Wait if you depend on explicit caching, if you cannot raise client timeouts, or if your pipeline is extraction-shaped: G​oogle's own documentation warns that xhigh "can cause overthinking or exhaust max_output_tokens on simpler prompts such as text or JSON extraction," and at high-by-default you are already one rung above where a Carbon-shaped pipeline used to sit.

The checkable thing to watch is not another launch post. It is whether the 61 holds up at high — where the tracker scores it 59.6, behind Carbon's 60.0 at Carbon's own ceiling — and whether G​oogle adds explicit caching to the model. Both are documented behaviours, both are visible from outside, and both change the answer to "is this the default for our agent fleet" more than any demo does.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily