Daily EditionAI · ML · Tech
LOCAL COMPUTE GETS UP TO 128GBInside
The Token Times
RUBY'S SEVEN FIXES NEED DEPLOYINGInside
Estd. 2026The last 24 hoursThursday, November 19, 2026
PRICIER TOKENS · GOOGLE SAYS THE TASKS COST LESS.

Neon costs more per token, less per task

Google claims cheaper completed work, while its new model costs more per token than Carbon's introductory rate.

On Nov. 18, Google released Gemini 4 Neon for coding, analysis, and long-running agents. It offers frontier benchmark results with a cheaper cost per task, Google says. The model has a 1 million-token context window. Input costs $3 per million tokens. Output costs $15; cached input costs $0.15.

Google cites a first-place score of 58.6 on the Artificial Analysis Coding Agent Index. At xhigh thinking, Neon scores 61 on the Intelligence Index, placing third overall. At its default high setting, Google claims a 32% lower cost per task than Carbon. It also ranks second on steerability in Agent Arena, which tests responses to mid-task corrections. Thinking level is part of the price comparison.

Carbon reached public users on Nov. 10. Its introductory input rate is $2 per million tokens, with output at $10. That offer expires on Nov. 30. The later rates are $4 for input and $20 for output. Neon therefore charges more per token than Carbon's current offer. Google's task-cost claim is a separate measure.

Neon is rolling out through the paid Gemini API in Google AI Studio and through Gemini for business. Google AI Pro and Ultra subscribers can use

it in Antigravity and the Gemini app. Inputs can include PDFs, spreadsheets, images, audio, and video. Four thinking levels let teams adjust reasoning and token spend. In Antigravity, Neon can coordinate parallel Hydrogen subagents on longer jobs.

Google reports 65.2% on Terminal-Bench 4.1. On SWE-sweep, Neon reaches 39.0% in Antigravity, against 12.8% in the mini-SWE harness. The surrounding tools matter as much as the model name. Google says Neon reduces redundant file reads and handles context compaction reliably. Its lowest fabrication rate comes from internal Antigravity tests. Google's coding tests also track cheating and shortcuts, rather than counting a passing score alone.

The cyber variant reaches trusted defenders through the Fairwind Program in the coming days. Google says Neon keeps Carbon-level protections against severe misuse and prompt injection. Artificial Analysis cautions that task-cost charts exclude infrastructure, engineering, and supervision. Teams should test completed jobs, cache behavior, and runtime on their own workloads. The invoice still belongs to the whole agent.

If you buildPatch NGINX and Ruby, then compare Neon's high and xhigh settings on your own coding tasks.
If you foundPrice agents by completed tasks, and test Neon as Carbon's introductory rate approaches its expiry.
Startup watch
LAUNCHED
Mastra@mastra

On Nov. 18, Mastra shared Factory, its open-source software factory, on HN. It turns issues into plans, code, and reviewed PRs, with a Slack-driven workflow. Teams can configure approvals and host the system on their own infrastructure.

LAUNCHED
Maritime@maritime_sh

On Nov. 18, Maritime launched a managed cloud for production agents. Persistent microVMs include sleep/wake, triggers, recovery, and per-agent budgets. A 2 GB agent starts at $1 a month. The company handles the work beyond renting a sandbox.

SHIPPED

On Nov. 18, Springcraft released springrep, a safe Rust CAD kernel, under MIT. Its separate conformance suite has 3,487 cases. The team built both for robot-part modeling; the tests check geometry against exact answers across kernels.

SHIPPED
atrium@atrium_dev

On Nov. 19, atrium's team launched ADE Index, listing 975 agentic coding tools. It links facts to source passages and publishes its scoring rules. The owner's workspace gets the same scoring treatment. Tool discovery is becoming a product.

SECURITY WATCH

NGINX needs a patch at the front door

On Nov. 18, NGINX patched two HTTP/2 flaws that Wiz demonstrated could allow unauthenticated RCE. Upgrade open-source deployments to 1.30.7 or 1.31.9. Secure API code does not fix the proxy.

COMPUTE POLITICS

Local inference gets a Windows box

On Nov. 18, ASUS, Dell, and MSI launched RTX Spark desktops. They offer up to 128GB of unified memory, native CUDA, and Windows Subsystem for Linux. Local tasks avoid per-token API charges. Compare supported models before replacing cloud inference.