Google released Gemini 4 Neon on November 18, its third new model in eight days, priced at $3 per million input tokens and $15 per million output tokens. On the Artificial Analysis Coding Agent Index, Neon scored 58.6, the highest of any model currently on that leaderboard. The release lands 24 hours after OpenAI shipped GPT-6.2 Astra and Sol, compressing what has historically been a months-long model cadence into a single intense November week that has reshuffled competitive rankings three times in a row.
Three Models in Eight Days: What Neon Slots Into and What It Retires
Google launched Gemini 4 Neon on November 18, following Gemini 4 Carbon on November 10 and Gemini 4 Hydrogen on November 12. The cadence is deliberate. Google is building a three-tier model stack organized by cost and task depth rather than the old Flash-and-Pro binary it used through the Gemini 3.x era. The naming scheme, Hydrogen, Neon, and Carbon, maps roughly to task weight: light everyday execution, balanced depth and efficiency, and the hardest long-horizon reasoning.
Neon sits at $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million, a 95% discount from the standard rate. That price is twice Hydrogen's rate ($1.50 input, $7.50 output) and currently 50% more expensive per token than Carbon's promotional rate, which stands at $2 input and $10 output through November 30. The promotion matters because 9to5Google's breakdown of the pricing schedule confirms that Carbon's price doubles to $4 input and $20 output starting December 1, leaving Neon 25% cheaper than Carbon per token after the promo window closes. For teams running high-volume agentic pipelines, December 1 is a material decision point.
The launch also retires two older models. Argon and Gemini 3.1 Pro are being removed from the Gemini model picker, and Gemini 3.5 Flash-Lite is relabeled "Auto" for paid users. For teams using the Gemini API, this creates two practical migration questions: whether Neon or Hydrogen replaces Argon for their existing workloads, and whether to stay on Carbon before its December price increase makes the case for Neon economically obvious. The announcement was made by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect. Neon is rolling out in Google AI Studio, in Gemini for Business, in the Gemini app for AI Pro and Ultra subscribers, and in Google Antigravity for the same subscriber tiers.
The model carries a 1 million-token context window and four thinking levels: Low, Medium, High, and Extra High (xhigh). Thinking tokens are billed at the output rate. The context caching rate of $0.15 per million is the most aggressive cached-token discount Google has applied to a mid-tier model, and for applications that re-use large shared context blocks across repeated agent calls, such as a fixed system prompt or a standing codebase chunk, that caching rate can drop the effective per-token input cost by up to 95% on cached portions, changing the economics for cache-heavy workloads. Neon also supports Google's Tasks feature for multi-step automated jobs, expanded PDF and spreadsheet handling, and more frequent use of live web sources when answering queries that require current information.
The Coding Index Score That Separates Neon From the Mid-Tier Pack
The benchmark most consequential for developers evaluating Neon is the Artificial Analysis Coding Agent Index, which is not run by Google. Artificial Analysis tracks model performance across coding tasks using a composite methodology and publishes results independently of any lab. Neon scored 58.6 on that index, the highest score currently on the leaderboard, ahead of GPT-6.2 Astra and Claude Fable 5.2. For comparison, Fable 5.2 leads the broader Artificial Analysis Intelligence Index with an overall score of 65, while Astra scores 62 and Neon scores 61 at its xhigh thinking level.
The task-level picture underneath that coding headline is uneven. On Terminal-Bench 4.1, a standard coding evaluation run under shared harness conditions, Neon scores 65.2%, trailing Claude Opus 5.5's 65.8%. On FrontierCode 2.0 at the default High thinking level, Neon scores 46.2% against Opus 5.5's 47.3%. On the shared Mini-SWE harness, which all labs use under equivalent conditions, Neon scores 12.8% against Sol's 11.6%. These are the results that independent replication tends to confirm, and the margins are narrow enough that workload-specific testing matters more than any single number.
Where Neon's coding lead grows more visible is on longer-horizon agentic tasks run inside Google's infrastructure. On SWE-sweep run inside Google Antigravity, Neon reaches 39.0% against Opus 5.5's 10.1% and Sol's 8.3%. This is the number Google foregrounds in its model card, but it reflects a provider-controlled harness environment that optimizes for Neon's specific runtime characteristics. The difference between 39.0% under Google's harness and 12.8% under the shared harness is a measurement environment artifact rather than a pure capability gap. For teams not running inside Google Antigravity, the Mini-SWE results are more predictive of real-world performance.
Science and mathematical benchmarks show Neon's most unambiguous leads. It scores 70.5% on Terminal-Bench Science, which Google describes as best in class across all currently available models. It scores 84.0% on RiemannBench, a mathematical reasoning evaluation, and 71.3% on HLE-Diamond. On Chartography, a multimodal reasoning test, Neon scores 73.5%, which trails GPT-6.2 Sol's 75.1%. On Agents' Last Exam v1.1, Neon scores 75.8%, which Google describes as on par with leading models. For research, analytical, and scientific workloads, the benchmark evidence is stronger and less harness-dependent than the coding results.
What Google Claims Against Opus 5.5 and Sol, and What the Table Leaves Out
Google's model card for Neon includes a direct comparison against Claude Opus 5.5 and GPT-6.2 Sol across 19 benchmarks, with Neon leading on 17. The largest margin comes on GraphWalks, a long-context retrieval test spanning context windows from 256K to 1 million tokens: Neon scores 95.2%, against Opus 5.5's 66.8% and Sol's 80.1%. On AutomationBench, run by Zapier to test real-world task automation, Neon scores 56.3% against 42.5% for Opus 5.5 and 41.3% for Sol. On WebExploitBench v1.2, a cybersecurity evaluation, Neon scores 51.3% at roughly half the cost of competing large models, according to Digital Trends. On the Legal Research Bench and the Tax Agent Bench, Neon ranks first.
Critics argue, however, that the 17-of-19 record reflects a deliberately selected peer set rather than a comprehensive competitive comparison. The two models chosen for comparison, Opus 5.5 and GPT-6.2 Sol, are not the strongest models currently in the market. Claude Fable 5.2 leads the Artificial Analysis Intelligence Index with a score of 65. GPT-6.2 Astra scores 62. Neon's overall score of 61 at xhigh puts it third overall, behind both Fable 5.2 and Astra. Google's comparison table includes neither Fable 5.2 nor Astra. That peer-set choice positions the 17-of-19 record as a claim about the mid-tier competitive bracket rather than the full frontier, and enterprise buyers evaluating frontier-class capability should read the benchmark table with that scope in mind.
The bear case for Neon's benchmark claims goes further. The two tasks where Neon trails, Terminal-Bench 4.1 and FrontierCode 2.0, are both in coding, the category Google is most aggressively marketing as a Neon strength. The risk is that teams optimize for Neon based on the 17-of-19 record and discover in production that the specific coding evaluations most relevant to their workflows are precisely the ones where Opus 5.5 or Astra still leads. For coding-first teams, the AA Coding Agent Index score of 58.6 is a more useful single-number summary than the head-to-head table, because the index covers a wider range of tasks under independent conditions.
On enterprise benchmarks, the evidence is more directly applicable. Neon ranks first on Surge AI evaluations for finance, accounting, spreadsheet work, and constraint-following tasks. It ranks first on Andon Labs' Vending-Bench 2 with a value score of $16,843, which measures economic output delivered in multi-step commercial automation tasks. Cybersecurity defenders with access to Google's Fairwind Program can run Neon at its highest thinking levels on specialized security evaluations, where Neon scores 90.5% on SEC-bench Pro and 70.7% on CWE-bench v2, a 3.4 percentage-point uplift over Carbon. For corporate teams running structured analytical workflows, the enterprise benchmark set is a more actionable signal than the intelligence index scores alone.
The December 1 Pricing Change and What Carbon Teams Should Model Now
The most consequential information embedded in Neon's launch may not be Neon's own benchmarks but the Carbon pricing change that arrives simultaneously. Carbon currently runs at a promotional rate of $2 per million input tokens and $10 per million output tokens, a rate Google set when Carbon launched on November 10 and that expires November 30. Starting December 1, Carbon moves to $4 input and $20 output per million tokens, doubling both rates, according to the full pricing schedule reported by DataCamp's Gemini 4 Neon analysis.
After December 1, Neon at $3 input and $15 output is 25% cheaper than Carbon per token. At the High thinking level, Google says Neon already costs 32% less per task than Carbon, and at xhigh, 36% or more less than competing large-scale frontier models on a per-task basis. If those efficiency ratios hold post-December, the economic argument for staying on Carbon becomes a question of measurable output quality. For workloads where Carbon's additional reasoning depth produces results Neon cannot match, the premium is justified. For workloads where the two models produce equivalent output, Neon at $3 versus Carbon at $4 is a 25% cost reduction that compounds at scale in high-volume agentic pipelines. Teams currently running Carbon should model both their average thinking level and the task-type distribution now, before December 1 turns that into a live billing decision rather than a theoretical one.
Neon also runs at Batch and Flex pricing of $1.50 input and $7.50 output, the same as Hydrogen's standard rate. Teams doing asynchronous, non-latency-sensitive processing can route through Batch or Flex and run Neon at Hydrogen prices. The Priority tier costs $5.40 input and $27.00 output. That is 35% more than Carbon's post-December standard rate. For time-critical production use cases requiring guaranteed low latency, Neon at Priority is more expensive than Carbon at standard, so the cost equation does not uniformly favor Neon. The choice between tiers depends on latency requirements, and teams with hard latency budgets may find Carbon at standard pricing still the better post-December option for their most time-sensitive calls.
What to Watch Before December 1 and in the Weeks After
December 1 is the first hard signal. If teams running Carbon-intensive agentic pipelines begin migrating to Neon after the price change, that migration rate will appear in developer forum activity, API billing patterns, and the leaderboard rankings on Vals and Artificial Analysis over the subsequent two to three weeks. If Neon's cost advantage proves durable in production, it accelerates Google's pricing power in the mid-tier API market, where OpenAI's GPT-6.2 Sol at $2/$10 and Anthropic's Opus 5.5 currently bracket the competitive range. If migration is slower than Google's pricing logic implies, it suggests that Carbon's capability ceiling is worth the premium to more teams than the benchmark comparison predicts.
Independent replication of Google's benchmark claims is the second signal to track. Artificial Analysis and Epoch AI both published independent analyses of GPT-6.2 Astra and Sol within days of their November 17 release. Similar treatment for Neon should arrive within 7 to 14 days. If those results confirm Neon's leads across a wider range of conditions, the pricing narrative strengthens. If independent results show Neon closer to Opus 5.5 and Sol in a broader measurement environment, the story narrows back to cost efficiency rather than capability leadership. The specific tasks to watch are Terminal-Bench 4.1 and FrontierCode 2.0, the two rows where Neon currently trails under shared harness conditions.
Cursor and Gemini CLI integration is the near-term technical milestone most critical to the coding benchmark story. Neon is not yet listed in Cursor's model documentation or the Gemini CLI docs as of the November 18 launch. For the coding index score to translate into real developer adoption, those integrations need to ship. The specific timing has not been announced. If Cursor support arrives before December 1, it gives developers a reason to switch during the same window the pricing change makes Carbon more expensive. If the integration delays past December, the two catalysts decouple and the adoption curve is likely slower than the benchmark story alone implies.
Longer term, Google's release pace over the next 60 days will indicate whether the three-model burst of November 10 to 18 is the start of a sustained new cadence or a concentrated launch event. Three Gemini 4 models in eight days is unprecedented velocity for Google's AI team. Whether the team ships Gemini 4 updates, new variants optimized for specific verticals, or additional family members before year end will tell more about the structural change in Google's AI operation than any single benchmark row. For enterprise buyers negotiating API contracts for 2027, the pace of model releases is as relevant as the benchmark scores: a faster cadence means faster deprecation cycles and more frequent migration decisions.
Google is betting that most agentic workloads don't need Carbon's ceiling. They need Neon's floor, priced 25% below Carbon once the promotional rate expires on December 1.
Key Takeaways
- Coding Agent Index #1 at 58.6 on the Artificial Analysis leaderboard, the first third-party benchmark where Neon leads all frontier models currently on record.
- $3/$15 per million tokens, with cached input at $0.15 per million and a 1 million-token context window; Batch and Flex tiers run at $1.50/$7.50, the same as Hydrogen's standard rate.
- Carbon doubles to $4/$20 on December 1, making Neon 25% cheaper than Carbon per token after the promotional pricing window closes November 30.
- 17 of 19 benchmarks vs. Claude Opus 5.5 and GPT-6.2 Sol per Google's model card, but independent scores place Neon at 61 on the AA Intelligence Index, third behind Fable 5.2 (65) and Astra (62).
- Not yet integrated into Cursor or Gemini CLI as of the November 18 launch, limiting access to the developer toolchains most central to the coding benchmark claims.
Questions Worth Asking
- If Neon's SWE-sweep advantage shrinks from 39.0% to 12.8% when measured under a shared harness, should coding teams use provider-harness or shared-harness results when choosing between models?
- Google benchmarked Neon against Opus 5.5 and GPT-6.2 Sol but not against Fable 5.2 or GPT-6.2 Astra. What does that peer-set selection reveal about where Google believes Neon actually competes?
- If Carbon's December pricing makes Neon the default mid-tier choice, does that represent a deliberate strategy to migrate users away from Carbon, or is the promo expiry simply a corrective return to cost-reflective rates?