Back to the show

AI & Tech Daily

Gemini Neon and the cost of getting AI work finished

18:04

Gemini 4 Neon brings faster agent work, with task costs shaped by effort settings and a temporary Carbon discount. OpenAI reports uneven research-agent success and circuit improvements adopted in AMD designs. AWS adds per-request agent payments, NGINX patches two HTTP/2 flaws, and Anthropic starts public trading. Cloudflare opens its Clairic document-extraction models, while Genie 4 reaches eligible Ultra subscribers, including in Australia.

Full transcript

Read the episode.

I'm Jesse Owen. This is AI and Tech Daily.

Gemini Neon: the cost of finishing

A faster AI model can have a higher price per token and still cost less to finish a job. Google's Gemini 4 Neon puts that calculation in front of developers, with a discount on its rival complicating the answer.

Our main story is Gemini 4 Neon and the question it raises for anyone paying an agent to do lengthy work: how much does an acceptable result cost, and how long are you waiting for it?

Google released Neon on 18 November through its API, and began rolling it out to Pro and Ultra subscribers in Gemini and Antigravity. It's aimed at sustained coding, analytical and scientific work. The context window is one million tokens, meaning it can take a large amount of material into a request. Published API pricing is three US dollars per million input tokens and fifteen dollars per million output tokens. Input is the material supplied to the model; output is what it generates. With output priced at five times the input rate, the balance between those two parts affects the bill. Two tasks with the same total token count can therefore have different costs.

The interesting result comes from Artificial Analysis. In its evaluation, Neon took roughly twenty-one minutes per task at its extra-high effort setting. Carbon, running at high effort, took about thirty-six minutes. That's a fifteen-minute difference. If you're waiting for an agent before you can inspect a result and continue working, that time has a practical value of its own.

The evaluator also measured a lower task cost: six US dollars forty-four for Neon, against seven dollars seventy-six for Carbon. That calculation used Carbon's standard pricing. Carbon's current discount runs until 30 November, with input tokens at two dollars per million and output tokens at ten dollars. Those temporary rates are below Neon's rates.

So the immediate purchasing decision has a date attached. A comparison made during Carbon's discount can produce a different cost ranking from one using its standard tariff. We have a measured speed advantage on these tasks, and a task-cost advantage against the standard price. The discounted bill needs its own calculation.

This is where agent economics gets more interesting than the price printed beside a model name. The token rate is one part of the bill. The amount of work the model does to reach an answer determines how many tokens you're buying. A model with a higher rate can finish with less total spending if it uses sufficiently fewer billable tokens. Effort settings belong in that comparison too; these measurements used different settings for the two models.

Then there's what happens after the agent stops. For a coding job, the useful endpoint is a change that passes the relevant checks and does what the developer asked. If a cheaper attempt leaves another round of work, the cost of that round belongs with the original task. That's the measure I'd use when deciding whether a faster model earns a place in a development workflow.

Elapsed time also affects how you organise the day. A shorter run could make it easier to inspect a proposed change while the problem is still fresh, or fit another iteration before a deadline. An unattended overnight job might place more weight on total spending. The same speed improvement can have quite different value in those two situations.

My read is that Neon gives developers a concrete reason to revisit longer jobs they've found too slow. The useful comparison keeps the task and acceptance checks steady, records the effort setting, and uses the tariff they'll actually pay. That turns the twenty-one-minute result into a question about their own work: can they get a usable answer sooner, at a total cost that makes sense?

Research agents have uneven strengths

That leaves us with a practical buying question. The next finding shifts attention to the work we give these agents.

OpenAI published internal research-agent measurements on 18 November. It estimates that autonomous success on tasks requiring two to four hours of human work rose from twenty-three per cent in March to forty-six per cent in October. Success here means completion without human intervention.

The more revealing numbers sit underneath that overall improvement. For tasks analysed between 1 August and 18 October, estimated success was seventy-nine per cent for monitoring training runs. For writing inference and serving-stack code, it was twenty-five per cent. That's the software involved in running a trained model and delivering its responses.

Both activities belong inside an AI research organisation, yet they present very different opportunities for delegation. An overall success figure can hide that difference. A manager looking at the twenty-five per cent category is considering a workflow where substantial intervention remains common. The monitoring result suggests a much larger share of that particular work can be completed autonomously.

For an engineering manager, I'd read this as a reason to organise agent adoption around named activities. Monitoring a run has a different purpose and finish line from changing the machinery that serves a model. Keeping those categories separate lets an organisation see which work is moving off a researcher's desk and which work is returning for attention.

There's also a useful distinction in the way the outcome is counted. Completion without intervention tells you whether the agent carried the task through on its own. A task that still needs help can contain useful work, but the researcher has to make room for that help. If the goal is to free up uninterrupted time, the frequency of those interruptions becomes part of the value calculation.

Task selection and measurement methods changed across the period, so the rise from March to October isn't a direct measure of researchers' overall productivity. The task-level spread is the stronger guide to where delegation may pay off now.

The organisational opportunity is quite specific: direct autonomous work towards activities with a demonstrated record of completion, and budget review time for the more difficult categories. That could let a research group get useful capacity from agents before every part of its engineering work is equally suited to them. Choosing the work is a management decision with as much practical weight as choosing the model.

AI circuit changes reach AMD designs

There's a more tangible example of AI-assisted engineering inside the chips themselves.

OpenAI says AMD has incorporated some model-proposed circuit changes into a design sent for manufacturing, with others included in future production design files. The results, published on 18 November, put AI-generated optimisation inside an established chip-design process.

The models worked on SystemVerilog, the language used here to describe circuit behaviour. They rewrote those descriptions to explore alternative implementations. Synopsys tools then checked that the changes preserved the blocks' behaviour and assessed the engineering trade-offs.

That sequence is important to understanding the advance. A circuit can be described in different ways while carrying out the same function. The design challenge is finding an implementation that makes better use of power, performance or area. Area is the amount of chip space the circuit occupies. Improving one of those dimensions can affect the others, so designers need a way to compare candidate changes against the job the circuit already does.

OpenAI reports improvements of up to fourteen per cent in that optimisation against heavily optimised AMD flagship-GPU design baselines. The figure applies to the circuit-design results being reported. Measurements of finished GPUs aren't available in this report.

What interests me is the starting point. These were already heavily optimised designs. Finding further improvements there suggests a potentially useful role for models in searching through alternatives that an engineering group might otherwise have to explore manually. The model proposes a rewrite; the engineering tools establish whether the behaviour survives and what the trade-offs look like.

There are two different tests in that process. Preserving behaviour asks whether the revised block still performs its intended function. Assessing power, performance and area asks whether it's a better implementation for the designer's goals. A change needs to make sense on both counts before it becomes useful engineering work.

AMD's reported adoption gives that process a concrete endpoint: selected changes have reached manufacturing design material. For organisations designing chips, the opportunity is to shorten the search for useful improvements while retaining the checks already built into the workflow. Engineers can spend their attention comparing viable options and deciding which trade-offs suit the product. The valuable contribution is an additional source of candidate designs that can be evaluated within that process.

Agents can pay through AWS

A different obstacle appears when an agent needs a service it can't already access.

AWS Marketplace introduced agentic payments through Amazon Bedrock AgentCore on 19 November. Supported APIs, data and content can now be bought per request using the x402 protocol, with the charge appearing in US dollars on an existing AWS account.

For the AWS billing route, there's no cryptocurrency wallet to set up and no separate billing registration with each participating seller. An authorised agent can purchase a supported service while carrying out a task. Search is one example: the agent may need information from a paid service to continue its work.

The appeal for developers is the removal of a small but awkward break in the workflow. Without a suitable purchasing arrangement, reaching a paid resource can leave a task waiting for someone to organise access. Bringing that transaction into an existing cloud account gives the software a supported way to continue, provided the service and the payment route are available.

AWS includes spending limits and approved-seller controls. Private Marketplace policies still apply. Transactions appear in CloudTrail and Cost Explorer, giving organisations records of the activity and its cost through their AWS tools.

Those controls let a developer express purchasing authority separately from the task itself. Asking an agent to answer a question can involve paid services along the way. The organisation can define which sellers are available and how much the agent is allowed to spend, without having a person approve each individual request inside those boundaries.

I think that's the most useful design consequence here. Permission to spend becomes a normal part of setting up the agent, alongside access to the tools it needs. A spending limit can define the budget for that work, while seller approval narrows the services available to it. The transaction records then help the organisation understand what the task consumed.

Availability is limited to participating products and AWS Regions that support AgentCore payments. Sellers also have to integrate the Marketplace payment mechanism. A developer's first practical question is therefore whether the particular service they need is supported in the region they're using.

For organisations already buying through AWS, this could reduce the administrative work around occasional paid API requests. Its usefulness grows with the services available: each relevant seller gives an agent another task dependency it can resolve through the same billing relationship.

NGINX fixes exposed HTTP/2 flaws

If you operate web infrastructure, this one has a more immediate maintenance consequence.

F5 released NGINX fixes on 18 November during coordinated disclosure of two HTTP/2 memory-safety flaws. Wiz demonstrated unauthenticated remote code execution, meaning an attacker could get code running through a reachable service without first logging in.

Both flaws require attacker-reachable HTTP/2. One was demonstrated against ordinary static-file serving. The other involves reverse-proxy processing under connection-slot pressure, where the server is handling requests on behalf of another service and available connection slots become constrained.

That static-file example makes the exposure worth checking even on a server with a relatively simple job. The relevant questions are whether HTTP/2 is reachable by an attacker, whether the installed package is affected, and whether its configuration meets the conditions for either flaw.

For NGINX Open Source, the fixed releases are version 1.30.7 on the stable branch and 1.31.9 on mainline. Those numbers are worth keeping exact because they identify the upstream fixes. NGINX Plus has separate fixed releases, and Linux distribution packages need their own patch check.

The flaws span long histories of upstream releases, but fixed stable releases are exceptions to those broad ranges. Distribution maintainers may also backport fixes, applying a security change while retaining an older-looking version number. An operator using a distribution package needs the maintainer's patch information to determine its status.

If patching can't happen immediately, temporarily disabling HTTP/2 mitigates these two vulnerabilities. That gives operators a specific way to remove the required exposure while arranging the appropriate fixed package. It addresses this pair of flaws; other outstanding security fixes still need their usual attention.

For an organisation with several kinds of NGINX deployment, the useful prioritisation starts with exposed HTTP/2 services and their package evidence. An internet-facing server using an affected build presents a different situation from an installation where the required protocol isn't reachable. That lets the operator direct urgent work towards deployments that meet the conditions.

Anthropic enters public markets

The financing behind AI development also changed this week, with a major supplier entering public markets.

Anthropic began trading on Nasdaq under the ticker ANT on 19 November, after pricing its initial public offering the day before. The offering comprised nine hundred and eighty million shares at eighty-three US dollars each.

Most of those shares are being sold by Anthropic itself, with the remainder sold by existing holders. The company's disclosed share allocation works out to approximately seventy-nine point four billion US dollars in gross company proceeds before fees. That's the portion relevant to the money raised by the business; proceeds from existing holders' sales go to those holders.

The offering was scheduled to close on 20 November, subject to the usual conditions. Completed settlement and final net proceeds weren't established in the available reporting.

For organisations relying on AI suppliers, the consequential development is the scale of the announced funding capacity. Capital can support future model development and infrastructure, although the offering figures don't specify how that spending will translate into products. I'd watch what Anthropic builds and makes available with that capacity. Those decisions have a more direct connection to customers' plans than the movement of its shares on the first trading day.

What Changes for You

There are also two new ways to put these tools to work, depending on whether you're building software or exploring ideas.

Cloudflare released updated Clairic document-extraction models on 18 November, available through Workers AI or as downloadable weights. Clairic has twenty-seven billion parameters; the smaller Clairic-flash has nine billion. They accept text, images and PDFs, with up to four images or rendered PDF pages combined per request.

For a developer extracting invoice totals, receipt items or support-ticket fields, the useful output is structured data: typed fields, confidence information, and source-text spans where evidence can be matched. That provides a way to connect a returned value with the document it came from. The published Clairic weights use the Apache two-point-zero licence.

Hosted input currently costs ten cents per million tokens for either model. On 1 December, the larger model rises to twenty-four cents; Clairic-flash stays at ten. Output tokens aren't charged, though repeated processing passes can bill the same input again. Existing Clairic users receive the larger replacement in place, so that price change deserves attention.

This requires software integration, and self-hosting requires suitable resources. Cloudflare's launch calls Clairic graduated from preview, while its documentation still labels it preview. For builders, specialised extraction looks worth evaluating where a general reasoning model is doing routine field collection.

And if you'd rather explore a generated world, Google began rolling out Genie 4 in Project Genie on 18 November. Eligible AI Ultra subscribers aged eighteen or older can use the browser-based service in supported countries, including Australia. Pro subscribers can join a waitlist.

You can change an environment mid-session using text prompts, and Genie 4 adds real-time native audio. That gives creative exploration another dimension: change the scene while you're inside it, then experience how it responds.

It's an experimental prototype, with camera latency, physical interactions, audio precision and persistent scenes still presenting research challenges. I'd see its immediate value in interactive creative prototyping, where trying a different environment can help develop an idea. For an eligible subscriber, the new opportunity is to explore and revise a scene during the same session.

You'll find the sources and full transcript at owenonthenet.com. Thanks for listening.

Sources

Reporting behind this episode.

  1. blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon
  2. artificialanalysis.ai/articles/gemini-4-neon-intelligence-gain
  3. openai.com/index/research-acceleration-jagged-frontier
  4. openai.com/index/advancing-chip-design-with-our-models
  5. aws.amazon.com/about-aws/whats-new/2026/11/marketplace-agentic-payments-agentcore
  6. docs.aws.amazon.com/bedrock-agentcore/latest/devguide/payments-process-payment.html
  7. nginx.org/en/security_advisories.html
  8. wiz.io/blog/nginx-http2-rce-vulnerabilities
  9. bleepingcomputer.com/news/security/critical-nginx-nginx2root-bugs-let-attackers-hijack-servers
  10. anthropic.com/news/anthropic-prices-initial-public-offering
  11. theverge.com/ai-artificial-intelligence/1035384/anthropic-ipo-trading-nasdaq
  12. blog.cloudflare.com/open-clairic
  13. developers.cloudflare.com/workers-ai/models/clairic
  14. huggingface.co/Cloudflare/clairic
  15. blog.google/innovation-and-ai/models-and-research/google-deepmind/genie-4-now-available-project-genie
  16. support.google.com/labs/answer/16875695
  17. support.google.com/googleone/answer/16476748