THE APEX TIMES
Alphabet’s Google DeepMind introduces Gemini 4 Neon, a new model positioned for long-running, tool-heavy tasks, with published benchmarks and token-based pricing for input and output.
Google is rolling out a new large AI model aimed at speeding up complex work while cutting the cost of running it. In a release posted Wednesday, the company described Gemini 4 Neon as a “frontier-level” model designed for fast end-to-end task completion across software engineering, knowledge work, and scientific workflows.
Gemini 4 Neon is marketed as strong at multi-step instruction following and for working with messy, real-world business inputs. Google says the model can synthesize across spreadsheets, PDFs, charts, audio, and video within a single 1 million-token context window, then convert those materials into structured deliverables. The company also frames Neon as built for agent-style workflows that keep running over long stretches, reducing redundant file reads and handling “context compaction reliably” during extended sessions.
Google provided a series of benchmark claims to support the performance profile. It said Neon ranks ahead of comparable models, and at its “xhigh thinking level” it scored 61 on the Artificial Analysis Intelligence Index, placing it #3 overall behind larger, slower models. On the Vals Index, Google said Neon is #2 and sits on the Pareto frontier, a concept used in AI evaluation to describe points where improvements come with relatively favorable tradeoffs rather than just higher cost or larger models.
The release also emphasizes controllability and lower “fabrication” risk. Google said Neon ranks #2 on steerability in a new “Agent Arena Index,” meaning it can adapt cleanly to mid-task user corrections. The company additionally claimed it reached its lowest fabrication rate in internal “Google Antigravity testing,” though it did not provide external verification details or specify the absolute methodology behind that internal test name.
Pricing was a prominent part of the announcement. Google set Gemini 4 Neon at $3 per million input tokens, with a $0.15 per million cached input tokens rate that it says represents a 95% discount for reused context. Output tokens are priced at $15 per million, and the model supports a 1 million-token context window. In addition, Google said Neon can be cheaper per task than at least one named peer, stating that at the default “high” thinking level it pairs a higher index score with a 32% lower cost per task than “Carbon,” and that it is roughly 50% to 80% lower cost per task than top competing frontier models at high settings. The company also said xhigh thinking can deliver deeper multi-step reasoning for the hardest engineering and analytical workloads while still costing 36% or more less per task than competing models at their highest setting.
Google positioned Neon as particularly relevant to developer and production coding work, including debugging across repositories. It said Neon ranks #1 on the Artificial Analysis Coding Agent Index, scores 65.2% on Terminal-Bench 4.1, and reports 46.2% on FrontierCode 2.0 at a default high setting. For “production-ready code,” Google claims Neon performs multi-file debugging and refactoring with targeted, surgical edits, and that it shows strong results on a “mini-SWE” harness and a separate “Google Antigravity” evaluation, where it said performance reached 39.0% after improvements.
The company also introduced model variants tuned for security and cybersecurity use cases. Gemini 4 Neon for cyber, according to Google, produces more secure code than Carbon across AutoBaxBench, including a state-of-the-art result on its hard tier (76%). Google also said that on WebExploitBench v1.2 it is state-of-the-art (51.26%) at half the cost of competing large models, and it described a code hardening lifecycle workflow aimed at finding vulnerabilities, validating them, and patching. It said this cyber variant is rolling out to “trusted defenders” in the coming days as part of a program called the Fairwind Program.
The release situates Neon within a broader Gemini lineup and what Google says is an approach to balancing intelligence and efficiency. Google described a set of “thinking levels” within Neon, saying teams can dial intelligence and token spend to match their workloads, with four levels scaling steadily. It also described “Carbon” as tackling hardest, long-horizon projects and “Hydrogen” as handling fast everyday execution, with Neon positioned as the middle ground for high-frequency and complex workflows.
What is not fully disclosed is how readily these benchmark results translate into customer deployments. While Google provides detailed rankings and percentage outcomes, the company does not in the release explain how those benchmarks are normalized for cost, latency, security constraints, or the kinds of tool environments agents will face in production. It also does not specify the timeline for general availability beyond the cyber variant’s “coming days” rollout, or whether there are limits on which developers and enterprises can access the model first. For now, the practical question for Alphabet will be whether Gemini 4 Neon’s claimed efficiency and reliability hold up in real agent workloads where tools, permissions, and failure recovery are the hard parts.
Investors moved quickly after early outlines connected United Launch Alliance’s work to the BE-4 engine family, which Blue Origin’s New Glenn is expected to use for a planned 2027 flight carrying AST SpaceMobile’s BlueBird satellites.
A report tied to Anthropic’s debut highlights a cash position large enough to dwarf the combined figures of Amazon and NVIDIA, while also pointing to a new exchange-traded fund built around companies described as “ANT-linked.”
A Yahoo Finance report highlights how Berkshire Hathaway’s roughly $10 billion investment in Alphabet appears to be losing value, with Bill Ackman commenting that the timing and pricing were unexpected even as he noted the stake was bought at a perceived discount.
The company said it has reshaped leadership around national security and export-control oversight after Huang’s exit, emphasizing governance changes ahead of its next quarter and the continuing geopolitical scrutiny on advanced chips.
Anthropic priced its IPO at $83 per share ahead of its debut, while Nvidia heads toward a seventh straight weekly decline amid fresh US pressure on AI chip exports.
A report citing Cathie Wood’s ARK funds says they sold about $84.8 million of Nvidia shares before the company’s earnings period and used the cash to buy holdings tied to crypto markets.
Reports tied to Nvidia’s former chief executive, including allegations of multiple affairs involving a suspected Chinese intelligence-linked figure, are now being read alongside a broader, politically sensitive backdrop for advanced chip sales across the United States and China.
A celebrity investor publicly criticized NVIDIA’s former CEO Jensen Huang, then appeared to reverse course several days later, echoing a message that Nvidia employees “know better than the board,” according to a Yahoo Finance report.
A reporting update says the major banks and asset managers that earlier this year promised as much as $500 billion in financing support for Nvidia customers have not yet moved capital into that program.
The sale, disclosed in a regulatory filing, trims Lisa Su’s direct stake in Advanced Micro Devices while the company’s shares have climbed sharply this year.
Comments
Community guidelines. Comments may be removed for violations.
Loading comments...