THE SIGNAL IN ONE SENTENCE
Google has released Gemini 4 Neon, a new frontier model positioned between the speed-first Hydrogen model and the heavier Carbon model. Google says Neon is built for complex coding, analytical reasoning, knowledge work and scientific tasks that need serious capability without the full cost of its largest model. It accepts text, images, audio and video, offers a combined context window of up to one million tokens and has four thinking levels. The model began rolling out on November 18 through the paid Gemini API tier in Google AI Studio, Gemini for business, Google Antigravity and the Gemini app for eligible Pro and Ultra subscribers. Google lists API pricing at $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. Google also published an impressive table. Neon scored 61 on the Artificial Analysis Intelligence Index, 70.9 percent on the Vals Index, 65.2 percent on Terminal-Bench 4.1 and 46.2 percent on FrontierCode 2.0 at the settings named in Google's model card. The launch post says Neon costs 32 percent less per task than Carbon at its default high thinking level and roughly 50 to 80 percent less per task than top competing frontier models on selected third-party indices. Useful numbers. Not your receipt. The plain signal is simple: make Neon earn the switch on accepted work. Accepted work is the output that survives your real finish line. The code passes the repository tests and review. The spreadsheet preserves formulas and matches the source data. The research memo cites the right evidence and marks uncertainty. The analysis can be reproduced. The customer-facing draft follows policy and does not invent a convenient fact. Token price is only one input to that finish line. A cheaper run can become expensive when it needs three retries, a larger thinking setting, a longer prompt, a human rewrite or a second model to verify the first one. A more expensive run can be cheaper when it succeeds once and leaves a clean evidence trail. So do not begin with the leaderboard. Begin with a small pile of work your organization already understands. Choose twenty to fifty representative tasks from the last month. Keep the awkward ones. Include the bug that crosses three packages, the spreadsheet with hidden dependencies, the slide deck with a picky visual system, the policy question with an exception and the scientific task where the correct answer is that the evidence is incomplete. Remove private material that the evaluation does not need. Preserve the task instructions, source files, tools, permissions, expected output and acceptance criteria. Then freeze the environment. Record the exact model name, thinking level, prompt, system instructions, tools, context files, temperature or other sampling controls, cache behavior, region, date and harness version. If Neon runs inside Antigravity, an API workflow and the Gemini app, treat those as different systems. The surrounding harness can change planning, tool use, retries and cost even when the model name is identical. Google's own launch material makes that point accidentally and helpfully. On SWE-sweep, Neon scores 12.8 percent in the mini-SWE harness and 39.0 percent in Google Antigravity. That is not a small wrapper effect. It is a reminder that a model result belongs to the model, harness, tools and settings together. Compare systems, not labels. For each task, run the current production system and the Neon candidate under the same acceptance test. Count a task as accepted only when it passes the test that matters in operation. Record partial credit separately so a polished near miss does not become a quiet success. Now price the whole attempt. Capture input tokens, cached input, output tokens, thinking level, tool calls, retries, wall-clock time, human review minutes, correction minutes and any verification model used afterward. Add the cost of failed attempts. If a worker waits twenty minutes for a five-dollar saving, the saving may be fictional. The right denominator is accepted tasks, not tokens. Total the model and tool cost for the evaluation, add the value of review and repair time, then divide by the number of accepted outputs. Report the distribution too. An average can hide a small set of spectacularly expensive failures. Use at least four buckets. First-pass accepted means the output passed without repair. Accepted after repair means a person or agent corrected it. Rejected means it did not reach the finish line. Unsafe or out-of-bounds means it violated a permission, policy or data rule even if the artifact looked good. That last bucket cannot be averaged away. Neon is advertised for complex software engineering and includes strong cybersecurity results. The model card also describes security evaluations and acknowledges that more capable models require stronger protections. A coding score does not grant production credentials. Run the candidate with the least privilege needed for the task. Use disposable environments for consequential code changes. Keep secrets out of prompts unless the task truly requires them. Log file writes, commands, network requests, package installations and external changes. Require human approval before deployment, payment, account changes, data deletion or contact with a customer. The thinking setting needs its own test. Neon offers minimal, low, high and xhigh levels. That is useful because a routine extraction task should not automatically pay for deep reasoning. It is also a new tuning surface where teams can accidentally spend more than they planned. Run the same acceptance set at the lowest plausible setting, then promote only the failures that might benefit from more reasoning. Record whether the higher setting actually changed acceptance, latency and total cost. Do not route every hard-looking task to xhigh because the name sounds reassuring. A routing rule should be evidence-based. For example, a low setting might handle formatting and bounded transformations. High might take multi-file refactors or technical analysis. xhigh might be reserved for rare tasks where a longer reasoning path produces a measurable gain and the extra delay is acceptable. The categories are yours to prove. Test corrections in the middle of work too. Google says Neon ranks highly on steerability, meaning it can adapt to user corrections during a task. Build a correction test rather than trusting the label. Change a requirement halfway through. Revoke a tool. Correct a false assumption. Ask the model to preserve finished work while replacing one section. Check whether it updates the plan, abandons invalid work and leaves the artifact internally consistent. This matters in agentic systems because the expensive mistake is often not the first wrong step. It is continuing confidently after the human has supplied the correction. Long context deserves similar suspicion. One million tokens is capacity, not comprehension. Test the document lengths you actually use. Plant answerable facts, contradictory versions, stale instructions and irrelevant attachments at different positions. Measure retrieval, citation accuracy, instruction priority and whether performance degrades as the context grows. Do not stuff the whole company into the prompt because the window is large. Use retrieval, permissions and data minimization. A large context window can increase exposure, latency and the number of plausible but wrong connections. The smallest sufficient context is usually easier to audit. Multimodal tasks need their own acceptance set. Neon accepts images, audio and video as well as text. Check charts with small labels, screenshots with similar controls, noisy audio, long recordings, scanned tables and clips where the important event is brief. Require the output to point back to the source frame, timestamp, region or cell when the task depends on visual or audio evidence. Then test the ugly edge. Malformed files. Missing permissions. A tool that returns an error. A repository with unrelated changes. A spreadsheet with circular references. A source that contradicts the user's premise. A prompt injection hidden in a document. The production question is not whether the model can complete a beautiful demo. It is whether it notices when the work should stop. Set explicit stop conditions before the pilot. A candidate should pause when it lacks a required source, cannot verify a consequential claim, reaches a spending limit, encounters a protected system, sees conflicting instructions or fails the same step twice. Record whether it stops cleanly and preserves enough evidence for a person to continue. Run the evaluation more than once. Frontier models are probabilistic, tools change and vendor services evolve. Repeat the high-value and high-risk cases across several runs. Track first-pass acceptance, variance and recurring failure patterns. A task that succeeds once and fails twice is not a stable automation. Keep a permanent regression set after the decision. Google can update a hosted model, harness, safety layer or tool integration. Your own prompts and connected systems will change too. Re-run the protected set on a schedule and after any material update. Version the results and keep a rollback route to the previous system. If Neon wins, say where. It may become the default for repository refactors but not spreadsheet work. It may beat Carbon at high while Hydrogen remains better for routine volume. It may work well in Antigravity and poorly in a thin internal harness. It may save money only when the cache is warm. That is not a messy conclusion. It is an operating map. Google has released a capable and attractively priced model. The launch chart is a useful invitation to test it. The switch belongs to your acceptance set.
01
WHAT ACTUALLY CHANGED
Google released Gemini 4 Neon on November 18, 2026.
Neon is positioned between the speed-focused Gemini 4 Hydrogen and the larger Gemini 4 Carbon.
The model accepts text, images, audio and video.
Its combined input and output context window is up to one million tokens.
Google offers minimal, low, high and xhigh thinking levels.
Neon began rolling out through the paid Gemini API tier, Google AI Studio, Gemini for business, Google Antigravity and the Gemini app for eligible subscribers.
Google lists API prices of $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens.
Google reports frontier-level results across coding, knowledge work, science, multimodal and cybersecurity evaluations.
Google says Neon costs less per task than Carbon and selected competing frontier models on named benchmark settings.
The reported SWE-sweep result changes substantially between the mini-SWE harness and Google Antigravity, illustrating that the harness is part of the measured system.
02
WHY THIS MATTERS
A model priced below the largest frontier tier can change which complex tasks are economical to automate.
Token price does not include retries, tool use, verification, review or repair.
Benchmark scores can move sharply with the harness, tools and reasoning settings.
Four thinking levels create useful control and another place for costs to drift.
A million-token context window can support large jobs while increasing audit and data-exposure challenges.
Multimodal access expands the range of work and the number of failure modes that need local testing.
Strong coding and cyber results make permission boundaries and action logs more important, not less.
A permanent regression set can detect later changes in the hosted model or surrounding system.
03
WHERE IT COULD HELP
- Build a twenty-to-fifty-task acceptance set from recent real work.
- Include hard cases, edge cases and tasks whose correct result is uncertainty or refusal.
- Freeze the model name, thinking level, prompt, tools, files, permissions and harness version for every run.
- Treat the API, Gemini app and Antigravity as different systems until local evidence shows otherwise.
- Measure first-pass acceptance, acceptance after repair, rejection and out-of-bounds behavior separately.
- Calculate cost per accepted task using tokens, tools, retries, review time and correction time.
- Test each task at the lowest plausible thinking level and promote only when higher effort improves acceptance.
- Run mid-task correction tests to see whether the agent updates its plan and finished artifact.
- Test long context with conflicting versions, stale instructions and irrelevant material.
- Require multimodal answers to identify the supporting frame, timestamp, region or cell.
- Log commands, file writes, network access, package installations and external changes.
- Keep consequential actions behind explicit human approval.
- Define stop conditions for missing evidence, repeated failure, conflicting instructions and budget limits.
- Repeat high-value and high-risk tasks to measure variance.
- Keep a versioned regression set and a tested rollback route after deployment.
KEEP A HAND ON THE WHEEL
Most capability and cost-per-task claims in the launch material are reported by Google, even when they use third-party benchmarks. Local results can differ because of task selection, harnesses, tools, prompts, thinking levels and acceptance criteria. The model card compares Neon with named competing systems under stated settings, but a benchmark table does not measure your complete workflow or labor cost. The one-million-token window describes capacity. It does not prove reliable retrieval, instruction following or citation across the full window. Google reports substantial differences between Neon in a mini-SWE harness and in Antigravity on SWE-sweep. Do not attribute a system result to the model alone. Paid access, subscription eligibility and product availability vary by channel and can change during rollout. Security and cybersecurity capability require least privilege, disposable environments, action logs, approval gates and rapid shutdown paths. Higher thinking levels can improve difficult work while adding latency and cost. Route with measured acceptance gains rather than task appearance. Hosted models and surrounding tools can change. Keep a dated regression set, versioned results and a rollback path.
04
TERMS WORTH KEEPING
OPEN GLOSSARY CARD
Least privilege
Giving a person or program only the access needed for its current job and no more.
OPEN GLOSSARY CARD
Benchmark
A fixed test used to compare how systems perform on the same tasks.
OPEN GLOSSARY CARD
Regression test
A repeatable check that reveals whether a change broke behavior that previously worked.
SOURCES AND VERIFICATION STATUS
This article was written from the materials below. Product claims and dates were checked against those sources on November 19, 2026.
THE PUBLICATION ENGINE
WANT A SIGNAL OF YOUR OWN?
We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.
WORK WITH US