Google's newest model costs more per token than the one it shipped six days earlier. It is sold on costing less per job. Both can be true at once, and knowing which one you bought is most of the decision.
Wednesday, November 18, 2026, brought a model priced for constant execution, GPUs placed behind a serverless event, and an agent designed to keep working after launch. The through-line is less about capability than about what happens when running the thing stops being the expensive part.
What this edition claims, and what it does not
No independent test, incident count, customer outcome or stock figure is asserted here. Three sources were read past the announcement summary: Google's Gemini 4 Neon post, Google Labs' Stitch Loop post and AWS's Lambda GPU post. The rest were reviewed at summary level and are described as such. Where a number is vendor-compiled, it is labeled.
Neon: the price is per token, the claim is per task
Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google, introduced Gemini 4 Neon with a framing that is unusually explicit about cost. The model is priced at $3 per million input tokens, $0.15 per million cached input tokens — which Google describes as a 95% discount — and $15 per million output tokens, across a 1 million-token context window.
That is twice the per-token price of Gemini 4 Hydrogen, which Google introduced on November 12 at $1.50 input and $7.50 output. Neon is not the cheaper model per token. It is positioned as the cheaper model per job: at its default high thinking level, Google says it delivers a 32% lower cost per task than Carbon, and roughly 50% to 80% lower cost per task than top competing frontier models at their highest settings.
Those two currencies are not interchangeable, and this part is analysis. A token rate is a price list. A cost per task is a total that depends on how many attempts the work requires, how much context gets re-sent, how much of that context is cached, and how much human review the output needs before it is usable. A model that is cheaper per token and needs three passes can lose to a model that costs more per token and lands in one.
The 95% cached-input discount is where that math actually lives. Repeated context — a repository, a policy document, a running case file — gets cheap on the second and subsequent reads. A workload that never reuses context gets none of it.
The benchmark table, and who assembled it
Google reports Neon at #3 overall on the Artificial Analysis Intelligence Index with a score of 61 at its xhigh thinking level, and publishes a comparison table in which Claude Opus 5.5 scores 59 and GPT-6.2 Sol scores 57. The same table lists Vals Index results at 70.9% for Neon, 70.1% for Opus 5.5 and 66.1% for GPT-6.2 Sol.
Domain claims include #1 on the Legal Research Bench and the Tax Agent Bench, 73.5% on Chartography, 70.5% on Terminal-Bench Science and 84.0% on RiemannBench. Google also reports #2 globally on Arena's new Agent Arena Index with a score of 1336 across what it describes as 4,200-plus real-world agent sessions, and #1 overall on the Artificial Analysis Coding Agent Index at 58.6.
The table is Google's compilation, including competitor scores Google entered. Some underlying benchmarks are third-party — Artificial Analysis, Vals, Arena, Collinear, Gray Swan — and several safety evaluations are Google's own internal testing, including its fabrication-rate and agent-honesty measures. A single assembled table flattens that distinction, and the distinction is the one that matters at procurement time.
Safety claims, and the gate in front of the cyber tier
Google's safety numbers in the post are specific. Neon refuses 1.98% of tasks on Terminal-Bench 4.0 and 0% on Terminal-Bench-Science; it recorded 0.0% cheating on Terminal-Bench 4.0 and took shortcuts on 2.45% of tasks in a coding cheating evaluation. Google reports a 0.09% misaligned action rate on its agentic safety suite and a 99.2% pass rate on its agent honesty evaluation, and says Neon reduces under-service by 43% to 56% and overreach by 15% to 35% across effort levels.
On Gray Swan's indirect prompt injection benchmark, Google states that with real-time safeguards at the highest thinking level, Neon "resisted every attack within this Gray Swan benchmark." That is a strong claim bounded by a specific configuration, and it is presented as such.
The operational detail worth noticing is the distribution gate. A cyber-tuned variant of Neon reaches eligible members of Google's Fairwind Program over the coming days rather than general availability. Google is shipping a cheaper, faster model to everyone and a security tool to vetted defenders, which is a trust boundary expressed as a rollout schedule rather than a documentation page.
Also in the post: Neon can act as what Google calls a fast, high-judgment Orchestrator directing parallel Hydrogen subagents inside Google Antigravity. Two models, one session, different jobs. That is routing as an architectural decision, and it is the reason per-token price alone no longer answers a procurement question.
AWS: GPUs behind an event function, with a stated ceiling
Amazon's Lambda update is short and consequential. Event Functions running on Lambda Managed Instances now support GPU instances in the G-series, deployed as container images, with AWS stating that Lambda handles provisioning, scaling, driver management and event integration. Pricing is standard EC2 pricing plus a management fee.
AWS also states the ceiling plainly: the capability is "optimized for inference rather than training, and custom models rather than large foundation models already available on services like Amazon Bedrock." That sentence narrows what a reader should expect more than any other line in the post.
The published limits support the reframing this section makes, and this part is analysis. Up to 90 minutes per invocation for asynchronous work, up to 48 concurrent requests per execution environment, and accelerator memory from 3 to 48 GB. The historic constraint on GPU work was idle capacity — you pay for the accelerator whether or not a request arrives. Moving the accelerator behind an event means the planner reasons about concurrency, cold-start behavior and driver versions instead of utilization. Those are different skills, and they are not free.
Google Labs ships an agent that does not clock out
Stitch Loop, from Google Labs, is described by Justin Toupin, Director of Product Management at Google Labs, as "a long-running product agent that connects to your codebase and live metrics to continuously improve what you've shipped." The post is specific enough to be useful and quiet enough to need qualification.
What it states plainly: setup connects GitHub plus product context — logs, user behavior, feedback — through more than 180 integrations, custom MCP servers or local CLI uploads. Three default loops ship out of the box, with dedicated loops for cybersecurity, latency and accessibility. Each loop runs a sequence beginning with Suggestions, then Solutions carrying tradeoffs and scoping, then Build.
What it also states is that access is a waitlist, described as global. That is availability language, not a launch.
The control that matters most sits in the accompanying summary rather than the body of the post reviewed: every pull request requires manual approval. That is the right design for an agent that writes to a production codebase, and it is described rather than demonstrated. No approval-workflow detail, acceptance rate or review burden was reviewed here.
Read alongside the rest of the day, the direction is consistent. Neon is priced to be invoked constantly. Lambda is priced per event. Stitch Loop is described as running 24/7. Three announcements, one assumption: execution is abundant.
Named and not characterized
- Genie 4. Google is rolling out a new Project Genie experience to eligible Google AI Ultra subscribers, including first-time access to the Genie 4 world model, upgrades to Genie 3 and a community gallery. Named only; no capability, benchmark or quality claim is made here.
- Gemini app on lower-memory Android devices. Google said the app now supports lower-memory devices with Gemini Live and custom image creation. Not tested.
- Google for Mexico. Google described AI innovations and partnerships announced at its Mexico event. Country-program detail not reviewed.
- November Demand Gen Drop and Google Marketing Platform Drop. Holiday-season advertising updates, the latter including Spotify audiences, Pinterest inventory and a Loblaw integration. Not characterized here.
- Google Play Las Vegas. A rewards and live-event promotion tied to race week. Not characterized.
- Cybersecurity tips collection. Google published a consumer safety resource. Not evaluated.
- OpenAI, Inside the Codex sandbox. A description of how Codex runs commands while enforcing user-controlled filesystem and network limits across macOS, Linux and Windows. Reviewed at summary level.
- OpenAI, The Silent Consensus. A piece by Dean Ball, Head of Strategic Futures at OpenAI, on how more capable AI could reshape society and how people retain control of institutions. Named only; it is an essay, not a product claim.
JWA interpretation
This section is the publisher's analysis, held separate from the reporting above.
The three substantive items share a structure: each lowers the cost of the next action and raises the cost of the decision to take it. A cheaper-per-task model makes more invocations rational. A GPU behind an event makes more inference requests affordable. An always-on agent makes continuous work normal.
What does not get cheaper is the check. Google's own post is the evidence: the same announcement that publishes a price list also publishes refusal rates, shortcut rates, misalignment rates and an injection-resistance claim, and routes its cyber capability through a vetted program. If verification were free, none of that would need to be in the document.
For a buyer, that argues for one question applied to every one of these releases. Not whether the model is good, and not whether the price fell. What does one completed, reviewed task cost — including the reading, the second source and the person who signs. Neon's 95% cached-input discount is a real number. So is the review hour.
Also on the site today
The most-read article in this publication on November 18 was an analysis of what happens to an assistant's context when it is embedded inside a business application, and why memory scope is an entitlement decision rather than a feature. It carried 19 counted visits at the time of this edition — a pageview count, not a count of unique readers.
What remains unconfirmed
- Independent verification of any Neon benchmark. The comparison table is Google's compilation, and several listed safety benchmarks are Google's internal evaluations.
- Whether the 32% lower cost per task versus Carbon holds for workloads that do not reuse context and therefore capture little of the cached-input discount. No workload-specific figure was reviewed.
- The scope of the Fairwind gating for cyber-tuned Neon, including eligibility criteria and expansion timing.
- Cold-start behavior, practical cost and real-traffic concurrency for GPU-backed Lambda event functions. No benchmark or customer data reviewed.
- How Stitch Loop's approval workflow, suggestion scoring and review burden perform in practice. The post describes the mechanism; no result or customer evidence was reviewed.
- Whether waitlist access to Stitch Loop is broadly available or capacity-limited. The post states a global waitlist, not a date.
- Genie 4's measured capabilities beyond the existence of the rollout. Not characterized.
- Any measured effect of these releases on enterprise procurement or deployment behavior. None is asserted.
- Market data. No stock, index or valuation figures were reviewed for this edition, and no intraday or closing prices are reported.
Evidence notes and limits
Evidence for this edition was reviewed through 23:40 UTC on Wednesday, November 18, 2026.
Google's Gemini 4 Neon announcement, dated November 18, 2026 and attributed to Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect, was retrieved in full. From it this article draws: the pricing of $3 per million input tokens, $0.15 per million cached input tokens described as a 95% discount, $15 per million output tokens, and the 1 million-token context window; the #3 overall placement on the Artificial Analysis Intelligence Index at 61 at xhigh thinking level; the comparison table listing 59 for Claude Opus 5.5 and 57 for GPT-6.2 Sol, and Vals Index figures of 70.9%, 70.1% and 66.1% for the same three models; the #1 placements on the Legal Research Bench and Tax Agent Bench; the Chartography, Terminal-Bench Science, RiemannBench, Agent Arena Index, AutomationBench and Coding Agent Index results; the stated 32% lower cost per task than Carbon and the roughly 50% to 80% range against competing frontier models; the four thinking levels; the refusal, cheating, shortcut, misalignment, agent-honesty and under-service/overreach figures; the quoted Gray Swan statement; the Orchestrator description directing parallel Hydrogen subagents; and the Fairwind Program rollout. All of these are reproduced as Google-reported. The comparison table is identified as Google's compilation, including competitor figures Google entered, and no independent reproduction was available.
The comparison to Gemini 4 Hydrogen's $1.50 and $7.50 per million token pricing is drawn from this publication's November 13, 2026 coverage of that launch, which cited Google's announcement. It is used here only to establish that Neon's per-token price is higher, not to characterize either model's quality.
AWS's Lambda GPU announcement, dated November 18, 2026, was retrieved with substantial detail. It is the sole source for G-series GPU support on Event Functions in Lambda Managed Instances; deployment as container images; the described handling of provisioning, scaling, driver management and event integration; standard EC2 pricing plus a management fee; the stated optimization for inference rather than training and for custom models rather than large foundation models already available on services like Amazon Bedrock; the 90-minute extended timeout for asynchronous and event source mapping invocations; up to 48 concurrent requests per execution environment; and accelerator memory from 3 to 48 GB. No benchmark, cold-start figure, real-traffic measurement or customer outcome was reviewed, and none is characterized.
Google Labs' Stitch Loop post, dated November 18, 2026 and attributed to Justin Toupin, Director of Product Management, Google Labs, was retrieved. It is the source for the quoted description of Stitch Loop as "a long-running product agent that connects to your codebase and live metrics to continuously improve what you've shipped"; for the connection of GitHub and product context through more than 180 integrations, custom MCP servers or local CLI uploads; for three default loops plus loops for cybersecurity, latency and accessibility; for the sequence beginning with Suggestions, then Solutions and Build; and for the global waitlist. The statement that every pull request requires manual approval appears in the AI-generated summary accompanying the post, not in the body text retrieved, and is presented here as a described control rather than a verified one. No pricing, acceptance rate, review-burden data or customer outcome was reviewed.
Google's Genie 4 post, the lower-memory Gemini app post, the Google for Mexico post, the Demand Gen Drop, the Google Marketing Platform Drop, the Google Play Las Vegas post and the cybersecurity tips collection were each reviewed at announcement-summary level and appear only in the named-and-not-characterized list.
OpenAI's Codex sandbox post and The Silent Consensus were each reviewed at announcement-summary level. No enforcement mechanism, efficacy result, argument or conclusion from either is characterized.
The reading of token price versus task cost as separate currencies, the observation that the cached-input discount determines which workloads benefit, the point about routing as an architectural decision, the reframing of GPU-behind-an-event as a concurrency problem rather than a utilization problem, and the argument that verification is the cost that did not fall are this publication's analysis. Nothing here establishes that any organization adopted Neon, reduced spend, joined the Stitch Loop waitlist, or changed a procurement decision.
The advertisement above is separate from this analysis and is not a source for any claim in it.