AI & Tech Daily
GPT-6.2 and the cost of getting work finished
OpenAI's Astra and Sol put model choice and the cost of accepted work in focus. Microsoft adds specialist coding models with introductory pricing, Azure brings Cobalt 200 compute to Australia East, and LangSmith links agents to owners and outcomes. Amazon examines credential-hunting campaigns, New York creates an AI enforcement unit, and Isomorphic Labs raises US$5 billion for drug development. Plus, what Australia's free Gemini Tasks rollout makes available, and which features remain US-only.
Full transcript
Read the episode.
I'm Jesse Owen. This is AI and Tech Daily.
Astra, Sol and the price of finished work
An AI model that costs four times as much could still be the cheaper way to finish a difficult job. OpenAI's new Astra and Sol make that a practical question for developers choosing where to spend.
Our main story today is GPT-6.2: how its two model choices change the calculation for producing code, documents and other professional work. OpenAI released Astra and Sol on 17 November in ChatGPT Work, Codex and its API. Astra is aimed at demanding projects; Sol is aimed at everyday coding and knowledge work.
For developers using the API, the standard rates are quite different. Astra costs eight US dollars per million input tokens and forty dollars per million output tokens. Sol costs two dollars for input and ten for output. Tokens are the chunks of text a model processes and produces. At those rates, Astra costs four times as much for the same input and output volume. These are API prices, separate from the question of which paid Work or Codex subscription you use.
That gives a developer a concrete choice. A routine revision and a difficult debugging job can consume similar amounts of text while demanding very different kinds of work. The useful question is how much effort remains after the model responds. Does the developer accept the change, send it back for another pass, or spend an hour untangling it?
OpenAI reports that Astra at its xhigh reasoning setting scored 5.2 points above GPT-6 Astra at max on Agents' Last Exam. For someone already using Astra on demanding tasks, it's a reason to compare the new model on the work that has been difficult to finish.
I'd measure that comparison at the point where somebody accepts the deliverable. With code, that could mean a change that passes the relevant checks and survives review. With a document, it could mean the intended reader can use it without another substantial rewrite. The model bill is one part of that calculation. Review time and repeated attempts are part of it too.
Consider a developer asking for a change that touches several parts of an application. A useful comparison would keep the requested behaviour and acceptance checks the same, then look at the complete path to a satisfactory change. If a more expensive model needs fewer repair rounds, its higher rate may be worthwhile. If both finish the job with similar effort, the lower rate becomes much more attractive.
This also gives organisations a way to think about model access without declaring one model the answer to every task. Everyday drafting and demanding project work have different economics. A developer can value a stronger option for the jobs that get stuck while still caring deeply about the price of routine work repeated across a week.
There's a specific access boundary for security work. Both models launched with tighter restrictions on higher-risk cybersecurity assistance. Access through OpenAI's reduced-refusal Daybreak programme was still planned, so people doing that work need to account for the restrictions in the available release.
My takeaway is to make accepted work the unit of comparison. Astra's premium is worth paying where it removes enough difficult review or rework; Sol becomes compelling where everyday tasks finish well at its lower rate. That is a more useful buying decision than choosing the most expensive model by default.
Microsoft's specialist coding prices
That's the broad model choice. Microsoft is taking a more specialised coding approach, with a date worth putting beside the price.
At Ignite on 17 November, Microsoft introduced MAI Code 2 and Code 2 Flash. Code 2 targets changes across multiple files and debugging. Flash is aimed at routine edits and other high-volume work. GitHub Copilot customers gain another coding-model option, expanding Microsoft's own offering for enterprise development.
A coding assistant can face a wide range of jobs within one repository. Fixing a small, well-defined edit is a different assignment from working out why behaviour breaks across several files. Microsoft's split gives those jobs different model options, rather than treating every request as equally demanding.
Microsoft says these models aren't yet tuned for fundamental science or very long-horizon agentic work. That puts a useful boundary around their intended role: development tasks such as the edits and debugging described at launch. Its performance comparisons cover selected coding benchmarks.
The pricing needs a calendar attached. Code 2's reported rates are fifty US cents per million input tokens and two dollars fifty per million output tokens until 17 May 2027. They reflect a fifty per cent promotional discount. Afterwards, the rates rise to one dollar for input and five dollars for output.
For a development budget extending beyond that date, the same volume of input and output would cost twice as much at the announced later rates. That is easy to lose inside an attractive launch comparison, especially when a trial begins on one budget and an ongoing service lands on another.
I'd treat the introductory period as an opportunity to learn which coding tasks fit the models. Routine edits could be a particularly useful place for a specialist option if a developer can accept them with little repair. Multi-file debugging gives Code 2 a more demanding test: finding the relevant change and keeping the surrounding behaviour intact.
Price the ongoing workload at the post-promotion rate, then judge whether the accepted changes justify it. A favourable trial price can make evaluation cheaper. The lasting value comes from a model that fits the work your developers repeatedly need to finish.
Cobalt 200 reaches Australia East
There's also a useful development underneath the models, in the computers running the rest of the application.
Microsoft made its Cobalt 200 virtual machines generally available on 17 November across fourteen Azure regions, including Australia East. These are Arm-based CPUs for work such as application services, databases, compiling code and executing an agent's tools. Australian Azure customers have an available region now, ahead of a promised expansion to twenty-three regions.
An AI application often involves several kinds of computing. The model generates a response or chooses a tool. A service then runs that tool, retrieves information or works with a database. A coding workflow may also compile the changed software. Those surrounding jobs can consume time even when the model itself responds quickly.
Cobalt 200 addresses that ordinary application compute. Its relevance to an agent is the work carried out around the model: the service handling a request, the database answering it, or the machine running the requested operation. GPU-based model training is a separate job.
Microsoft reports up to fifty per cent better performance than Cobalt 100, with the result depending on the workload. Existing Arm workloads can move across generations. For an Azure operator already running compatible software, that makes the new machines a concrete infrastructure option to compare. Software architecture compatibility and availability of the required virtual machine in the chosen region still govern deployment.
The distinction between faster compute and a faster complete application is useful here. If a tool spends much of its time doing CPU work, a faster machine could shorten that step. If the delay sits elsewhere, changing the processor may leave most of the wait untouched. That helps an operator identify which part of the application is worth measuring.
For Australian organisations, the practical opportunity is to examine eligible workloads in Australia East without waiting for the wider regional expansion. I'd look first at the services doing meaningful work between model responses. Improving those steps could make an agent feel more responsive and reduce infrastructure pressure while leaving the model choice alone.
An owner and an outcome for each agent
Once several agents are doing useful work, it helps to know who owns each one and where its results end up.
LangChain introduced its rebuilt LangSmith Agent Platform on 13 November and detailed the available product on 16 November. It brings an agent's tracing, evaluation, deployment and governance records together around a persistent identity. Each agent has a named owner, an immutable ID and a type. Development, staging and production environments sit beneath that identity.
In plain terms, tracing records what happened during a run, while evaluation helps assess the result. Connecting those records to the same agent gives the person responsible a way to follow its behaviour through development and into operation. The immutable ID keeps that identity stable as the surrounding work changes.
General-purpose and coding-agent types are available. The coding views include merged pull requests and token cost. A merged pull request is a code change accepted into the repository, so it gives a developer or manager a useful downstream event to put beside the cost of producing it.
That could change a conversation about an agent which looks very busy. A high run count tells you it has been active. A record of merged work and associated spend helps show what that activity produced. If costs rise while accepted changes don't, the owner has somewhere concrete to begin investigating. Traces can then help explain which steps were repeated or where a run failed.
Compatible external runtimes can connect through OpenTelemetry, a standard way to export operational data. That means adopting these views doesn't necessarily require moving an agent to LangChain's runtime. The useful coverage depends on the integration and the telemetry the existing runtime supplies.
LangChain says the platform is available globally through its cloud service, bring-your-own-cloud deployments and self-hosting. Organisation-wide dashboards and identity-provider registration remain on the roadmap.
For developers responsible for several agents, I like the shift towards a stable owner and a visible result. It gives a failed run somewhere to go and makes spending easier to discuss in terms of delivered work. The strongest use would be to connect that evidence to the person who can improve the agent, so the record leads to a better next run.
Credential hunting that funds more hunting
There's a darker use for that same ability to keep working: hunting for credentials that pay for more attempts.
Amazon published evidence on 13 November from MadPot, its network of decoy sensors. Requests captured between 3 October and 5 November contained offensive agents' operating histories. Those records described stealing AI-provider credentials and using them to fund further credential hunting.
An AI-provider credential gives software access to model services billed to an account. Someone who obtains it may be able to run their own requests at the account holder's expense. In the campaigns described by these records, access to one account could supply the model usage for looking for more credentials.
The targets included exposed AI gateways and application-management interfaces. The reported techniques relied on exposed secrets, misconfiguration and known vulnerabilities. For an organisation running an internet-facing gateway, that places attention on how the service is exposed, who can reach its management functions, and where credentials might be visible.
Amazon directly captured the requests and their embedded histories; outcomes described against other organisations remain unconfirmed. It assessed activity referring to potentially affected AWS customers and notified them as appropriate.
An AI gateway is particularly relevant because it can sit between an application and the model providers it uses. The organisation may be concentrating access and spending behind that service. A weakness there can therefore create both a security problem and an unexpected bill. Looking at unusual usage belongs alongside examining access controls and internet exposure.
For an operator, the spending record can help make an incident visible. Unexpected model consumption is worth investigating alongside unexpected access, because the service being abused is itself metered. A credential used outside its intended application can have a financial consequence even when the application continues to appear healthy.
My reading is that model credentials belong with other production credentials in the organisation's security and spending controls. The specific issue here is unauthorised access to a paid capability that can support further activity. Keeping ownership, access and usage visible gives the operator a clearer chance of spotting that abuse.
New York builds AI enforcement capacity
On the governance side, New York has named the team that will turn AI obligations into investigative work.
On 17 November, the state's Attorney General announced the AI Safety, Standards, Enforcement, and Transparency unit. Its remit includes enforcing AI-related laws and supplying technical expertise for investigations. The office was still building the team.
The unit will lead enforcement of the RAISE Act, which the announcement says takes effect on 1 January 2027 and covers large frontier-model developers. For businesses subject to those requirements, there's now an identified enforcement function as that commencement date approaches. Australian law is unchanged.
The operational detail is the combination of legal enforcement and technical expertise. Investigating an AI-related incident requires people who can understand the technology as well as the obligations attached to it. Establishing a dedicated unit gives that work an organisational home.
For covered firms, I'd read this as a reason to give incident documentation and safety disclosures practical attention. Records that explain what happened, and the basis for a disclosure, become more useful when a technically equipped investigator may need to examine them. The next visible development will be how the office staffs and uses that capacity.
Five billion dollars for drug development
A very different kind of investment is going into the distance between a computational drug design and clinical testing.
Isomorphic Labs announced a five-billion-US-dollar Series C funding round on 17 November, co-led by Greenoaks and Sequoia Capital. The company says the money will support its IsoDDE drug-design engine, therapeutic programmes, and the infrastructure for translational and clinical development.
That spending plan reaches beyond improving the design software. Translational work concerns carrying discoveries towards their use in medicine, while clinical-development infrastructure supports the work needed to take programmes into human testing. Those are consequential destinations for capital in a business built around AI-led drug design.
The progress reported in the company release is preclinical. Human-trial efficacy results and an approved treatment are later milestones, with trial-entry timing still unestablished.
For organisations following this field, the interesting question is how the new funding changes the programmes that reach those milestones. My view is that the capital could help connect computational discoveries to the development work needed to test them. The useful follow-up is a programme entering clinical trials, then measured outcomes that show what the resulting treatment can do for patients.
What Changes for You
If you've been waiting to hand a few everyday jobs to an assistant, Australia's access is changing.
Google began rolling out Tasks in Gemini for free on 17 November in Australia and five other markets. It widens access to the agent previously offered through the paid Spark subscription. Tasks uses Gemini 4 Hydrogen, and supported jobs include meal planning, calendar management and reviewing emails for recurring expenses.
The change is access to delegated work without that extra subscription. Reviewing recurring expenses, for example, can involve connected email data, rather than only the text you type into a conversation. Calendar management similarly reaches into a personal service. That makes the account permissions part of the decision about which jobs to hand over.
Availability is rolling out, so an individual Australian account may not have it immediately. There is also a clear regional boundary around the newly announced additions. Business calls, Google Pay purchases, Google Password Manager sign-ins, and assignments sent by text or email start in the United States. Australia's free rollout covers the supported Tasks offering, while those additions remain US-first.
Google says tasks run in a dedicated, hardware-isolated virtual machine, with confirmation or pre-approval for sensitive actions. You can interrupt a task or disconnect connected apps. Google also acknowledges failures and occasions when the user has to finish the job.
For someone deciding whether this is useful, I'd focus on the job and the access it needs. A meal plan and a review of your email expenses involve different amounts of personal information. The appealing part is being able to try supported delegated work without another subscription; the practical trade-off is allowing the assistant into the services that make that work possible.
The useful result is a finished task you can check. For a calendar change, that means the intended details are in place. For recurring expenses, it means you can see which payments deserve your attention and make the decision yourself.
You'll find the sources and full transcript at owenonthenet.com. Thanks for listening.
Sources
Reporting behind this episode.
- community.openai.com/t/gpt-6-2-astra-and-sol-are-now-available/1417531/1
- finance.yahoo.com/technology/ai/articles/microsoft-aims-challenge-openai-anthropic-181007858.html
- infoworld.com/article/4250152/microsoft-launches-mai-code-2-models-to-challenge-openai-and-anthropic-in-ai-coding.html
- azure.microsoft.com/en-us/blog/azure-cobalt-200-vms-now-available-proven-on-real-workloads-and-built-for-the-agentic-era
- langchain.com/blog/series-c
- langchain.com/blog/we-rebuilt-langsmith-around-the-agent
- aws.amazon.com/blogs/security/amazon-threat-intelligence-observes-threat-actors-running-cybercrime-campaigns-with-artex-ai-agents
- ag.ny.gov/press-release/2026/attorney-general-james-announces-creation-ai-safety-standards-enforcement-and
- prnewswire.com/in/news-releases/isomorphic-labs-announces-5-billion-series-c-funding-302942093.html
- blog.google/innovation-and-ai/products/gemini-app/tasks-in-gemini-free-new-features
- theverge.com/tech/1035756/google-spark-ai-tasks-free