Released November 18, 2026 · verified November 21, 2026
Gemini 4 Neon: specs, price, and how it stacks up against Opus 5.5 and Fable 5.2
Live since November 18, 2026 at $3 input / $15 output per million tokens. Here is what Google actually published — the spec sheet, the whole rate card, its own benchmark table — and how it lines up against Claude Opus 5.5 and Claude Fable 5.2.
TL;DR
Gemini 4 Neon (gemini-4-neon) is on the paid Gemini API tier at $3.00 per million input tokens and $15.00 per million output tokens, with cached input at $0.15 and a single 1,048,576-token window shared by input, thinking and output. There is no free tier. Google places it between Gemini 4 Hydrogen for everyday work and Gemini 4 Carbon for the hardest long-horizon problems, with four thinking levels (low, medium, high by default, xhigh). Against Claude it is cheaper than both Opus 5.5 ($4/$20) and Fable 5.2 ($8/$40) on the sticker rates, and it takes video and audio input, which the Claude models do not.
Verified November 21, 2026
Every figure on this page was read from Google’s own pages on November 21, 2026: the November 18 launch post on blog.google, the DeepMind model card, and the Gemini API pricing, model and “what’s new” pages. Where Google has not published something — a full safety report, a system-card score — the page says so instead of filling the gap.
Gemini 4 Neon next to Opus 5.5 and Fable 5.2
Model IDs, release dates and list prices come from the site’s model registry, which was checked against Google’s and Anthropic’s pricing documentation on the dates in the note below.
| Spec | Gemini 4 Neon | Claude Opus 5.5 | Claude Fable 5.2 |
|---|---|---|---|
| API model ID | gemini-4-neon | claude-opus-5-5 | claude-fable-5-2 |
| Released | 2026-11-18 | 2026-09-22 | 2026-11-09 |
| Input (per M tokens) | $3.00 | $4.00 | $8.00 |
| Output (per M tokens) | $15.00 | $20.00 | $40.00 |
| Cached input (per M tokens) | $0.15 | $0.20 | $0.20 |
| Batch (per M tokens) | $1.50 / $7.50 | $2.00 / $10.00 | $4.00 / $20.00 |
| Context window | 1M | 1M | 1M |
| Max output | 65K / 1M | 128K | 128K |
| Input modalities | Text, image, video, audio, PDF → text | Text and image → text | Text and image → text |
| Thinking control | low / medium / high (default) / xhigh | Adaptive, always on; depth via effort | Adaptive, always on; depth via effort |
| Data retention | No retention term published; Google lists “used to improve our products: No” | Not stated on the pricing page | 30 days, mandatory |
| Open weights | No | No | No |
Specs re-read 2026-11-21: gemini-4-neon is the stable model ID, the context window is 1,048,576 tokens shared by input, thinking and output, synchronous output defaults to 65,536 tokens and background execution can extend a single response to the full window. Thinking supports low, medium, high (default) and xhigh; minimal is rejected. Image generation, audio generation and the Live API are listed as unsupported. Knowledge cutoff: March 2026. The two Claude columns are Anthropic’s published figures.
The full Gemini 4 Neon rate card
Google publishes four serving modes for the same model. Every figure is USD per million tokens on the paid tier; there is no free tier for Neon.
| Line item | Standard | Batch / Flex | Priority |
|---|---|---|---|
| Input | $3.00 | $1.50 | $5.40 |
| Output (including thinking tokens) | $15.00 | $7.50 | $27.00 |
| Cached input | $0.15 | $0.075 | $0.27 |
| Free tier | Not available | Not available | Not available |
| Grounding with Google Search | 5,000 free queries per month, then $14 per 1,000 | 5,000 free queries per month, then $14 per 1,000 | 5,000 free queries per month, then $14 per 1,000 |
Batch and Flex are both listed at $1.50 input / $7.50 output with cached input at $0.075; Priority is $5.40 / $27.00 with cached input at $0.27. Caching is implicit only — explicit cachedContents are not supported, and a cache hit bills 5% of the input rate. Grounding with Google Search and Google Maps share a 5,000-request monthly allowance across all Gemini 3 and newer models, then $14 per 1,000 requests with a 1,500-per-day cap. Every token is billed, including thinking tokens and internal tool instructions.
Google’s own benchmark table
These are the numbers Google printed in the Gemini 4 Neon model card on November 18, 2026, including the two competitor columns Google chose to publish. They are vendor-run evaluations, not independent measurements.
| Benchmark | Gemini 4 Neon | Claude Opus 5.5 | GPT-6.2 Sol |
|---|---|---|---|
| Artificial Analysis Intelligence Index v5.0.1 | 61 | 59 | 57 |
| Vals Index v2.2 | 70.9% | 70.1% | 66.1% |
| Terminal-Bench 4.1 | 65.2% | 65.8% | 61.5% |
| FrontierCode 2.0 | 46.2% | 47.3% | 43.8% |
| SWE-sweep (provider harness) | 39.0% | 10.1% | 8.3% |
| SWE-sweep (mini-SWE harness) | 12.8% | 3.6% | 11.6% |
| Zapier AutomationBench | 56.3% | 42.5% | 41.3% |
| HLE-Diamond | 71.3% | 67.9% | 64.0% |
| Chartography | 73.5% | 66.3% | 75.1% |
| GraphWalks, 256K→1M (BFS F1) | 95.2% | 66.8% | 80.1% |
| CWE-Bench v2 | 70.7% | 67.1% | 66.0% |
| WebExploitBench v1.2 | 51.3% | 37.1% | — |
Read from the model card on 2026-11-21. A dash is a cell Google left empty. On Google’s own table Neon trails Opus 5.5 on Terminal-Bench 4.1 (65.2% vs 65.8%), FrontierCode 2.0 (46.2% vs 47.3%) and Chartography (73.5% vs 75.1%), and leads it on the AA index, the Vals Index, SWE-sweep, AutomationBench, HLE-Diamond and GraphWalks. The launch post also claims first place on the Artificial Analysis Coding Agent Index at 58.6 and an AA Intelligence Index of 61 at xhigh thinking — vendor claims, quoted as claims. The Neon Safety & Security Report is not published yet.
Neon vs Opus 5.5 vs Fable 5.2: what actually differs
Four things decide most of these calls — price, what goes in, how long the context runs, and what the vendor promises about your data.
Price per task, not per token
At $3/$15 Neon sits roughly 25% below Opus 5.5 ($4/$20) and at about a third of Fable 5.2’s $8/$40 on the sticker rates. Google’s launch post claims a 32% lower cost per task than Gemini 4 Carbon at the default high thinking level, and roughly 50%–80% lower cost per task than competing frontier models on the Vals and Agent Arena indices — Google’s framing on its own cost-per-task basis, not a list-price comparison.
Video and audio go in directly
Neon accepts text, images, video, audio and PDFs and returns text. Opus 5.5 and Fable 5.2 accept text and images and return text. If the workload is “read this recording” or “watch this screen capture”, Neon is the only one of the three that takes the file as-is; the Claude models need a transcript or extracted frames.
One shared 1M window
All three advertise a 1M-token context, but Neon’s is unified: input, thinking and output draw on the same 1,048,576 tokens, with 65,536 synchronous output and long decode up to the full window through background execution. Fable 5.2 and Opus 5.5 keep a 128K output ceiling. Neon’s long-context price is flat — no surcharge above a threshold, unlike OpenAI’s 272K tier.
Data handling is where they diverge most
Fable 5.2 carries a mandatory 30-day retention window and is not ZDR-eligible without express authorisation. Opus 5.5’s retention is not stated on Anthropic’s pricing page. Google lists Neon as not used to improve its products and publishes no zero-retention commitment on the model page — read Google’s terms for your own compliance case rather than inferring one from this page.
Where you can buy it
Neon runs on the paid Gemini API (generateContent and the Interactions API), Google AI Studio for AI Pro and Ultra subscribers, Gemini for business, Google Antigravity and the Gemini app. Anthropic sells Opus 5.5 and Fable 5.2 on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS — so a team already standardised on Bedrock or Foundry has a shorter path to the Claude models.
The positioning sentences are Google’s and Anthropic’s own wording from their model and pricing pages, read on 2026-11-21; the prices are the registry’s. Nothing here is an independent performance measurement.
Practical notes before you switch
API-level differences that cost time if you discover them after the fact.
Model ID and availability
The ID is gemini-4-neon and it is stable, not a preview. It is paid-tier only: a free-tier project gets an error, not a degraded answer. It works through both generateContent and the Interactions API.
Thinking levels replace temperature
Neon supports low, medium, high (default) and xhigh, and rejects minimal. Sampling parameters are ignored: temperature, top_p and top_k are dropped if in range and return a 400 if out of range. Set thinking_level instead of tuning temperature.
Outputs past 64K need background mode
Synchronous responses stop at 65,536 tokens. To go further, set max_output_tokens above that and background to true so the server continues in 64K hops, or use client-side continuation in the current SDKs. Setting background without max_output_tokens still stops at 65,536.
Implicit caching only
Neon has no explicit cache objects — cachedContents and caches.create are unsupported. Implicit caching runs automatically on prompts of at least 4,096 tokens and discounts cache hits by 95%, to $0.15 per million input tokens.
It is also built to orchestrate
Google’s launch post describes Neon directing parallel Hydrogen subagents in Antigravity, and the API docs list gemini-4-neon as a supported model for the Antigravity agent in Gemini Managed Agents, standalone or paired with Hydrogen. That is a supported configuration, not a benchmark result.
No free tier
Every free-tier cell on the pricing page reads “Not available”. Budget for paid usage from the first request; Google’s own quickstart examples assume a paid project.
Read from the Gemini API “what’s new” and model pages on 2026-11-21. Migrating from Gemini 4 Carbon, Argon or Hydrogen needs no request-schema change — only the model ID, and the default thinking level if you depend on it.
Where Neon sits in the Gemini 4 family
Google shipped four Gemini 4 models. Neon sits between Hydrogen, for everyday work, and Carbon, for the hardest long-horizon problems.
Gemini 4 Carbon
The hardest, longest-horizon work: architecture, deep debugging, science and security problems. Default thinking level medium, capped at high.
Gemini 4 Neon
Complex, high-frequency work — coding, analytical reasoning, enterprise workflows — where frontier quality matters but cost per task does too. Default high, xhigh available.
Gemini 4 Hydrogen
Everyday coding and agentic execution at the lowest latency. Same 1M window and four thinking levels, lower price.
The descriptions are Google’s own positioning from the launch post and the model card. Gemini 4 Argon, the fourth model, is a separate reasoning model with its own model card.
Sources
Google’s pages, all read November 21, 2026, plus Anthropic’s pricing documentation for the two Claude columns.
- Google: Introducing Gemini 4 Neon
The November 18, 2026 launch post: positioning, the $3/$15 price, the 1M context, the thinking levels, the benchmark claims and the channel list.
- DeepMind: Gemini 4 Neon model card
Published November 18, 2026: inputs and outputs, the 1,048,576-token window, 64K/1M output, the March 2026 knowledge cutoff, the evaluation table and the frontier-safety assessment.
- Gemini API pricing
Standard, Batch, Flex and Priority rows for gemini-4-neon, cached input, and the grounding allowances.
- Gemini API: Gemini 4 Neon model page
Model code, token limits, supported capabilities, consumption options and the November 2026 update date.
- Gemini API: What’s new in Gemini 4 Neon
Thinking levels, long decode, the migration checklist from Gemini 3 and Gemini 4 models, and the flat-pricing note.
- Anthropic: Claude pricing
Opus 5.5 at $4/$20 and Fable 5.2 at $8/$40, both with a 1M context and 128K output, used for the two Claude columns.
What is not published yet
Google lists the Neon Safety & Security Report as “coming soon”, and the model card carries frontier-safety determinations forward from Gemini 4 Argon rather than fresh Neon results. There is no system card and no independent third-party benchmark on this page. If a number is not on Google’s pages, it is not on this page.
This site is independent of Google and Anthropic.
Price it on your own workload
Sticker prices are only a starting point. Put your request volume, prompt size and cache-hit rate into the calculator and compare Neon against the Claude models.