Monthly Technology Roundup
September 2026 Technology Roundup
What Changed, What Matters, What to Watch
September brought changes across AI, cloud, security, developer platforms, browsers and quantum computing. This roundup covers the developments that changed what you can use, what it costs, how systems are built or how much risk you carry. It is a selection, not a complete list.
The month in 60 seconds
One line per area, each with its tally. The icon or the line takes you to that section.
- AI models2 developments · 2 to exploreCheaper model tiers arrived from both OpenAI and Anthropic. GPT-6 Sol, GPT-6 Luna and Claude Opus 5.5 are priced well below their companies' flagship models, and both companies say the cheaper models come close on many tasks. Those performance statements are vendor claims.
- AI agents2 developments · 2 to exploreProducts for long-running agents arrived at very different stages. OpenAI's Agents API is in public beta, AWS CloudWatch Omni is generally available, Google's AlloyDB for agents is a preview and Microsoft's Autopilot is in private preview. The label matters more than the headline.
- Cybersecurity3 developments · 2 to use now · 1 to exploreTwo NetScaler vulnerabilities were exploited around the time of disclosure. Citrix, CISA and the fixed builds are covered below. This was the most urgent item of the month for anyone who runs the product.
- Security guidance3 developments · 2 to use now · 1 to exploreSecurity teams got two new official resources. NIST published guidance on protecting identity tokens, and Microsoft began publishing machine-readable exploitability statements for its CVEs.
- Software2 developments · 1 to use now · 1 to explorePost-quantum key exchange moved further into default software behaviour. Java 27 now prefers a hybrid post-quantum key exchange in TLS 1.3, and Cloudflare now selects the strongest key exchange each origin server supports.
If you remember only three things from this month
- If you run NetScaler ADC or Gateway, check your build against the fixed versions. Then decide whether an appliance that was exposed before patching needs a compromise check.
- Read the availability label first. Several of this month's agent products are in beta or preview, and plans built on them should say so.
- Treat prices and performance claims differently. Model prices are published and checkable. Most comparisons between cheaper and flagship models come from the vendors, so test on your own workload.
Artificial Intelligence2 to explore
OpenAI and AnthropicAvailableTJ watchExplore
Cheaper model tiers from OpenAI and Anthropic
What happenedOpenAI released GPT-6 Astra on 3 September at $10 per million input tokens and $50 per million output tokens. On 22 September it added GPT-6 Sol at $2 and $10, and GPT-6 Luna at $0.10 and $0.50. GPT-6.1 Sol followed on 29 September at the same $2 and $10 as Sol. Its change is in claimed capability, not price. OpenAI's documentation describes it as near-Astra performance at a lower cost than Astra. Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 1 September at $10 and $50. Anthropic estimates that Fable 5.1 costs about 25% less than Fable 5 on typical workloads, and up to about 45% less on highly agentic work. Mythos 5.1 is available only to vetted cyberdefenders and life scientists, and currently only to US organizations. Claude Opus 5.5 arrived on 22 September at $4 and $20, compared with $5 and $25 for Opus 5. Anthropic says it costs 40% less than Opus 5 at default settings on typical workloads. Claude Sonnet 5.5 followed on 28 September at the same prices as Sonnet 5.
Why it mattersInside one vendor's lineup, input prices now range from $0.10 to $10 per million tokens. For high-volume work such as coding assistance, document processing or agent loops, a cheaper model that is good enough can change what a project costs. How close the cheaper models come to the flagships depends on your tasks, and nobody outside the vendors has published a comparison here.
EvidenceDocumented fact: release dates, list prices, Mythos 5.1 access limits. Vendor claim: near-Astra performance, and Anthropic's 25%, 45% and 40% cost estimates. TechiesJournal interpretation: that cheaper tiers may be good enough for many workloads. Anthropic's figures are estimates of workload cost, not list price cuts. Mythos 5.1 is Limited, so it is not part of the Available status above.
Who should care Developers · AI teams · architects · technology leaders · Sources OpenAI API changelog and model pages · Anthropic: Claude Fable 5.1 and Mythos 5.1 · Anthropic: Claude Opus 5.5 · Anthropic API release notes
OpenAIPreviewTJ watchExplore
OpenAI's Agents API enters public beta
What happenedOpenAI's Agents API entered public beta on 10 September, open to all developers. It gives applications access to the Codex harness through an OpenAI-managed API. OpenAI handles sessions, orchestration, context compaction and recovery, while the application supplies tools and chooses where the agent runs. Usage is billed at the selected model's API rates, with OpenAI tools at their standard rates and OpenAI-hosted sandboxes at standard container rates. For now the service supports data residency only in the United States and does not support Zero Data Retention. Microsoft also announced Copilot Autopilot on 25 September. It is an agent that lives in a customer's tenant with its own identity, memory, computer and workspace. Microsoft says Autopilot is expanding to private preview at the end of September.
Why it mattersAn agent that works for hours needs a place to keep its state, a way to continue after a failure, and limits on what it can touch. The Agents API takes on part of that work. The residency and retention limits decide whether some organizations can use it yet.
EvidenceDocumented fact: beta status, scope and limits from OpenAI's documentation and announcement, and Autopilot's private preview status from Microsoft's post. Unknown: Autopilot pricing and general availability date. A long-running agent still needs someone to review what it did.
Who should care AI developers · platform teams · architects · security teams · Sources OpenAI Agents API guide · OpenAI announcement · Microsoft: Copilot Autopilot
Cloud Computing1 to use now · 1 to explore · 2 to watch
AWSAvailableTJ watchUse now
Elastic Beanstalk Cluster Mode
What happenedOn 17 September AWS made Elastic Beanstalk Cluster Mode generally available. Several applications can share infrastructure powered by Amazon EKS, and AWS says Elastic Beanstalk deploys, scales, patches, monitors and upgrades the applications. There is no additional charge for Cluster Mode. You pay for the underlying resources, including the EKS control plane fee, EKS Auto Mode compute, Amazon ECR and Amazon CloudWatch.
Why it mattersTeams that wanted something between a simple application platform and running Kubernetes themselves get another option. Sharing infrastructure across several applications may change what each one costs, so the EKS charges are worth modelling before you move anything.
EvidenceDocumented fact: general availability, shared EKS-based infrastructure and charges, per AWS.
Who should care Cloud developers · platform engineers · AWS architects · Sources AWS News Blog
AWS and Google CloudPreviewTJ watchExplore
AWS and Google add options for observing and serving agents
What happenedAWS made Amazon CloudWatch Omni generally available on 22 September. AWS describes it as a purpose-built observability, evaluation and experimentation solution for AI agents. On 24 September Google Cloud announced PostgreSQL for agents in AlloyDB, available in preview. It provisions sandboxed serverless instances in seconds, fully separated from the primary, standby and read replica instances, with up-to-the-second read-only access to production data.
Why it mattersAgents tend to send bursts of queries and tool calls that look different from ordinary application traffic. Omni is about seeing and evaluating what an agent did. AlloyDB's approach is about keeping agent queries away from the production primary. They address different problems, and only one of them is generally available yet.
EvidenceDocumented fact: Omni is generally available and AlloyDB for agents is a preview, per AWS and Google. Vendor claim: Google states the design supports over 3 million queries per second and millions of agents, and does not say how it measured those figures. The card carries the lower of the two statuses.
Who should care Cloud architects · platform teams · AI engineers · SRE teams · Sources AWS: CloudWatch Omni · Google Cloud: PostgreSQL for agents in AlloyDB
CoreWeaveLimitedTJ watchWatch
CoreWeave makes Vera Rubin NVL72 available, with one named customer
What happenedOn 30 September CoreWeave announced the availability of NVIDIA Vera Rubin NVL72 on its cloud. It named Cognition as the first customer anywhere running production workloads on the system. Cognition measured a 4.8 times increase in total token throughput for its SWE-2 inference workloads compared with a GB200 NVL72 baseline, and a 3.8 times increase in output token throughput for reinforcement learning workloads.
Why it mattersNew accelerator generations are usually discussed for months before anyone outside the vendor can run production work on them. A named production customer suggests the hardware has moved beyond demonstrations. The announcement gives no regions or capacity figures, which is why we list the status as Limited. That label is our judgement, not CoreWeave's wording.
EvidenceDocumented fact: the announcement, its date and the named customer. Observed result: the throughput figures were measured by Cognition on its own workloads, and are not an independent benchmark. TechiesJournal interpretation: the Limited status.
Who should care AI infrastructure teams · cloud architects · organizations running large AI workloads · Sources CoreWeave announcement
Government of CanadaAvailableTJ watchWatch
Canada publishes data-centre principles, and welcomes a large planned project
What happenedOn 3 September the Government of Canada launched its Responsible Data Centre Development Principles. The five principles say data centres should create lasting local benefits, not shift electricity costs to Canadians, minimize water use and environmental impacts, be transparent about local impacts, and bring strategic value to Canada. Twenty-three organizations signed at launch, including AWS, Google, Microsoft, Meta and OpenAI. Nineteen more joined on 22 September, bringing the total to 42. Separately, on 14 September the government welcomed a planned Bell investment in Saskatchewan. Bell plans up to 900 megawatts of new capacity, advancing toward a 1.2 gigawatt AI infrastructure hub, with total capital investment of up to $52.5 billion.
Why it mattersData centres need electricity, cooling water, land and grid connections, and the costs can reach well beyond the companies that use the compute. The principles are a set of commitments that organizations sign. They show electricity and water are becoming part of the conversation about AI infrastructure. The Saskatchewan item is a different kind of story: a private company's plan that the government welcomed.
EvidenceDocumented fact: the principles, their dates and signatory counts, and the Bell plan as the government describes it. The Bell project is a plan. It is not built capacity. The Available status refers to the principles being in effect.
Who should care Technology leaders · infrastructure architects · policy teams · organizations planning large AI deployments · Sources Principles launch, 3 September · 19 additional signatories, 22 September · Saskatchewan investment, 14 September
Cybersecurity2 to use now · 1 to explore
Citrix and CISAAvailableTJ watchUse now
NetScaler ADC and Gateway: two exploited vulnerabilities
What happenedOn 27 September Citrix published bulletin CTX697096, covering eight vulnerabilities in NetScaler ADC and NetScaler Gateway, CVE-2026-88771 to CVE-2026-88778. Two of them have a CVSS v4 score of 9.5. CVE-2026-88771 allows an unauthenticated attacker to execute arbitrary commands because of improper input validation. CVE-2026-88772 is a memory overflow that can lead to remote code execution or denial of service. Citrix says exploits of both have been observed on unmitigated deployments. CISA added both to its Known Exploited Vulnerabilities catalog on 27 September, with a remediation due date of 30 September. Fixed builds include 14.1-73.37 and later, 13.1-64.23 and later, and the matching FIPS and NDcPP builds listed in the bulletin. Earlier in the month, on 9 September, CISA added a separate NetScaler authentication bypass, CVE-2026-19490, to the same catalog.
Why it mattersNetScaler appliances often sit at the network edge, handling remote access and application delivery. A flaw that is exploited there does not wait for the normal patch cycle. Because exploitation was observed before the fixes were available, installing the fixed build closes the known flaws but does not show whether an appliance that was exposed earlier was affected. Teams in that position should decide whether to check for signs of compromise.
EvidenceDocumented fact: the CVEs, scores, exploitation statement, fixed builds and KEV dates, per Citrix and CISA. Unknown: who is exploiting these flaws and how widely. The sources reviewed do not say. The Available status refers to the fixed builds.
Who should care Infrastructure and network teams · security operations · anyone who runs NetScaler ADC or Gateway · Sources Citrix bulletin CTX697096 · CISA Known Exploited Vulnerabilities catalog
NISTAvailableTJ watchUse now
NIST publishes guidance on protecting identity tokens
What happenedOn 15 September NIST published IR 8587, "Protecting Tokens and Assertions from Forgery, Theft, and Misuse: Implementation Recommendations for Agencies and Cloud Service Providers." One of the four authors is from CISA. It covers identity providers, authorization servers, key management, token lifecycle, single sign-on, federation and API access.
Why it mattersA stolen or forged token can give an attacker access without needing a password. The report gives teams a structured set of recommendations to compare against how they issue, sign, store and accept tokens. It is written for US federal agencies and cloud service providers, and it is guidance, not regulation. Teams outside that audience can still use it as a checklist.
EvidenceDocumented fact: title, date, authorship and scope, per the NIST publication page.
Who should care Identity architects · cloud security teams · API and platform teams · Sources NIST IR 8587
MicrosoftAvailableTJ watchExplore
Microsoft publishes VEX statements for its CVEs
What happenedOn 8 September Microsoft began publishing Vulnerability Exploitability eXchange (VEX) statements for all Microsoft-assigned CVEs. A VEX statement is a machine-readable record of whether a vulnerability affects a particular product. Microsoft says the change does not increase the number of security updates customers need to deploy.
Why it mattersA CVE number tells you a flaw exists. It does not tell you whether the product and version you run are affected. VEX data lets vulnerability-management tools answer that question automatically instead of someone reading each advisory. It covers Microsoft-assigned CVEs only, so it will not replace that work for third-party software.
EvidenceDocumented fact: the scope and start date, per Microsoft's Security Response Center. How well your tools can use the data depends on the tools.
Who should care Vulnerability-management teams · security operations · tool owners · Sources Microsoft Security Response Center
DevOps1 to watch
DockerPreviewTJ watchWatch
Docker publishes the Sandbox Kit specification
What happenedOn 24 September Docker published the Sandbox Kit specification, open source under Apache 2.0. A Kit is an ordinary OCI image. Its manifest carries the declarations in one annotation, vnd.docker.sandbox.kit.descriptor, and the image layers carry the content. The declarations list the agent, its tools and everything it asks to reach, such as hosts, credentials and volumes. Docker also said it is bringing the specification to the CNCF under neutral governance.
Why it mattersAn agent needs permissions as well as a model. One agent may need a repository but not production credentials. Another may need a browser limited to approved domains. Writing those permissions down in a standard package could make agent environments easier to reproduce and review across different runtimes.
EvidenceDocumented fact: the specification, its packaging and Docker's statement about the CNCF. The CNCF step is Docker's announced intent. No acceptance by the CNCF has been announced, and a specification becomes more significant when runtimes beyond Docker adopt it.
Who should care Developers · DevOps teams · platform engineers · security architects · Sources Docker: specification and CNCF · Docker: Sandbox Kit specification
Software & Systems1 to use now · 1 to explore
Google and MozillaAvailableTJ watchUse now
Chrome and Firefox move to two-week release cycles
What happenedFirefox 155 on 1 September was the first Firefox release on Mozilla's new two-week rhythm. Chrome 153 on 8 September was the first Chrome stable release on a two-week cycle. Both browsers had used four-week cycles. Chrome's Extended Stable channel keeps major updates to every eight weeks for customers who cannot take a two-week cycle, with security fixes backported weekly. Mozilla says it will monitor the transition closely and adjust as it learns.
Why it mattersFixes and finished features reach users sooner. Teams that test web applications or manage browsers on company devices get more frequent changes to absorb, and Extended Stable is the option built for those who cannot.
EvidenceDocumented fact: dates and cadences, per Google and Mozilla. Mozilla notes that the new cadence does not mean Firefox will ship twice as many features.
Who should care Web developers · QA teams · enterprise IT · application owners · Sources Chrome: two-week release cycle · Mozilla: Firefox release cadence
OpenJDK and CloudflareAvailableTJ watchExplore
Post-quantum key exchange becomes a default in Java 27 and Cloudflare origin connections
What happenedJava 27 reached general availability on 15 September. JEP 527 adds hybrid post-quantum key exchange for TLS 1.3, and by default the JDK places X25519MLKEM768 at the front of the list of key exchange groups, making it the most preferred. The default list can be changed with the jdk.tls.namedGroups system property. Cloudflare made two changes. On 8 September it described Automatic Key Exchange for origin connections, which picks the strongest key exchange each origin server supports and prefers X25519MLKEM768. Post-quantum support for origin connections has existed since 2023, so what is new is the automatic selection. On 10 September its 1.1.1.1 resolver began validating DNSSEC signatures made with ML-DSA-44. Signing in Cloudflare's authoritative DNS and the matching record support in its registrar are planned, not live.
Why it mattersCryptography takes years to change across runtimes, servers, DNS and certificates. These are working pieces arriving now, so teams can start testing them while the migration is still early. A hybrid exchange only happens when both ends support it, so what your servers and proxies support matters as much as the runtime.
EvidenceDocumented fact: release dates, JEP 527 behaviour and the Cloudflare capabilities described, per OpenJDK and Cloudflare. Preparing for post-quantum cryptography does not mean today's encryption has been broken.
Who should care Security teams · Java developers · architects · infrastructure teams · Sources OpenJDK: JEP 527 · Cloudflare: Automatic Key Exchange · Cloudflare: post-quantum DNSSEC on 1.1.1.1
Quantum Computing1 to watch
Sandia, Quantinuum, NVIDIA, IBMResearchTJ watchWatch
A new benchmark and two IBM preprints on measuring quantum progress
What happenedA preprint submitted to arXiv on 10 September proposes QUOPS, a benchmark for the size of the largest circuits a quantum computer can run successfully and the speed at which it runs them. Its authors come from Sandia National Laboratories, Quantinuum, NVIDIA and the University of New Mexico. They applied it to processors from Quantinuum, Google and IBM. The paper says current systems fall short of what large problems such as factoring RSA-2048 need by roughly five orders of magnitude in circuit size, while today's physical-qubit processors have or nearly have the required operations rate. IBM discussed two separate preprints in a 15 September blog post. One, submitted on 11 September, combines error detection with probabilistic error cancellation. It reports sampling overhead reduced by up to a factor of 63 compared with cancellation without post-selection, in an experiment with six Trotter steps on ibm_aachen. The other, first posted in July, reports roughly a 10 times improvement in effective gate error rates after syndrome post-selection, on a circuit using 76 physical qubits and 314 T gates, with a fidelity lower bound of 0.349 at 95% confidence.
Why it mattersQuantum computers are often compared by physical qubit count, which says little about how much reliable computation a machine can do. QUOPS is one attempt to measure the work that actually runs. The IBM papers address a different part of the problem: how much useful computation error handling can recover before full fault tolerance exists.
EvidenceObserved result: the figures come from specific experiments in preprints, with no peer-review status shown. The 63 times figure is an upper bound on an inferred sampling overhead against one baseline. The 10 times figure compares gate error after post-selection with the physical gate error, and it does not mean IBM machines became ten times more accurate overall. QUOPS is a proposal, not an accepted standard.
Who should care Quantum researchers · technology strategists · organizations assessing quantum readiness · Sources QUOPS preprint (arXiv 2609.12146) · IBM spacetime mitigation (arXiv 2609.13108) · IBM sampling hard circuits (arXiv 2607.25941) · IBM Quantum blog, 15 September
Nothing in this month matches that combination. Show everything.
2 developments worth watchingNot the biggest stories, but the ones whose next move will tell us something.
- Kubernetes workload-aware schedulingKubernetes 1.37 was released on 26 August, so it falls outside this roundup's window, but its September feature write-ups are why it belongs here. The Workload and PodGroup APIs for gang scheduling, workload-aware preemption and shared DRA ResourceClaims for PodGroups graduated to Beta. DRA Extended Resource support graduated to general availability. Distributed AI and high-performance computing jobs often need all their pods to start together, which is the problem these APIs address. The open question is how many production clusters adopt the Beta features, so check what is enabled in your own version before relying on them.
- IonQ Superion 256IonQ's 8 September release says Superion 256 is available to order now, with customer deliveries in 2027. The first system was pre-sold in the first quarter of 2026. IonQ's claims about fidelity and future cost are the company's own. Worth following, but not available today.
Don't overread this monthWhere the headlines ran ahead of the evidence.
- HeardAgents can run for longer, so they can now work unsupervised.
- Measured viewAgents that run longer are not agents that run unsupervised. Managed sessions and observability are real progress. They also show how much surrounding engineering an agent needs before it can be trusted, and most of this month's agent products are still in beta or preview.
- HeardCheaper models mean we now know which model is best.
- Measured viewCheaper models do not establish a single best model. The prices are published. Most of the comparisons with flagship models come from the companies selling them. Testing on your own workload tells you more than any vendor table.
- HeardFaster browser releases mean twice the features.
- Measured viewFaster browser releases do not mean twice the features. Mozilla says so directly. A two-week cycle mostly shortens how long finished work waits for a release.
- HeardBetter quantum error handling means broad quantum advantage has arrived.
- Measured viewBetter quantum error handling does not mean broad quantum advantage. The IBM results are preprints about specific experiments. The QUOPS paper finds that the circuits today's machines run reliably are far smaller than the large problems people hope to solve. IonQ's new system is not due for delivery until 2027.
Worth going deeperTwo subjects from this month deserve deeper treatment beyond this roundup.
- From CVE to "Does This Affect Me?": What VEX Changes in Vulnerability TriageComing later
Planned Article - Beyond Qubit Count: How Should We Measure Whether a Quantum Computer Is Actually Improving?Coming later
Planned Article
References
Artificial Intelligence
- OpenAI API changelog and model pages for GPT-6 Astra, Sol, Luna and 6.1 Sol · documentation, 29 September 2026
- OpenAI Agents API guide · documentation, 10 September 2026
- OpenAI, Introducing the Agents API and hosted sandboxes · announcement, 10 September 2026
- Anthropic, Claude Fable 5.1 and Mythos 5.1 · announcement, 1 September 2026
- Anthropic, Claude Opus 5.5 · announcement, 22 September 2026
- Anthropic API release notes and Sonnet 5.5 pricing · documentation, 28 September 2026
- Microsoft, Introducing the new Copilot with Home, Code and Autopilot · announcement, 25 September 2026
Cloud Computing
- AWS, Introducing Amazon CloudWatch Omni · announcement, 22 September 2026
- Google Cloud, Announcing PostgreSQL for agents in AlloyDB · announcement, 24 September 2026
- AWS, Elastic Beanstalk Cluster Mode · announcement, 17 September 2026
- CoreWeave, Vera Rubin NVL72 availability · announcement, 30 September 2026
- Government of Canada, Responsible Data Centre Development Principles · government release, 3 September 2026
- Government of Canada, 19 additional signatories · government release, 22 September 2026
- Government of Canada, Saskatchewan sovereign AI infrastructure · government release, 14 September 2026
Cybersecurity
- Citrix, NetScaler security bulletin CTX697096 · security bulletin, 27 September 2026
- CISA, Known Exploited Vulnerabilities catalog · catalogue, 27 September 2026
- NIST IR 8587, Protecting Tokens and Assertions from Forgery, Theft, and Misuse · guidance, 15 September 2026
- Microsoft Security Response Center, expanded VEX · announcement, 8 September 2026
DevOps
- Docker, Sandbox Kit specification and CNCF announcement · announcement, 24 September 2026
- Docker, Sandbox Kit specification · specification, 24 September 2026
- Kubernetes v1.37 release · release notes, 26 August 2026
- Kubernetes, workload-aware scheduling in v1.37 · project blog, 8 September 2026
- Kubernetes, DRA updates in v1.37 · project blog, 3 September 2026
Software & Systems
- OpenJDK, JDK 27 · project page, 15 September 2026
- OpenJDK, JEP 527 Post-Quantum Hybrid Key Exchange for TLS 1.3 · specification, 15 September 2026
- Cloudflare, Automatic Key Exchange · announcement, 8 September 2026
- Cloudflare, post-quantum DNSSEC on 1.1.1.1 · announcement, 10 September 2026
- Cloudflare, post-quantum to origins · announcement, 29 September 2023
- Google Chrome, two-week release cycle · announcement, 8 September 2026
- Mozilla, Firefox release cadence · announcement, 19 August 2026
Quantum Computing
- QUOPS, Benchmarking the computational power of quantum computers (arXiv 2609.12146) · preprint, 10 September 2026
- IBM, Spacetime mitigation of logical errors (arXiv 2609.13108) · preprint, 11 September 2026
- IBM, Sampling hard circuits with verifiably high fidelity (arXiv 2607.25941) · preprint, 28 July 2026
- IBM Quantum blog, From error mitigation to fault-tolerant quantum computing · blog post, 15 September 2026
- IonQ, Superion product line · press release, 8 September 2026
Monthly Technology Roundup · Coverage 1 to 30 September 2026 · Published 6 October 2026 · Corrections welcome.